linux-stable/kernel/sched/fair.c

11286 lines
295 KiB
C
Raw Normal View History

License cleanup: add SPDX GPL-2.0 license identifier to files with no license Many source files in the tree are missing licensing information, which makes it harder for compliance tools to determine the correct license. By default all files without license information are under the default license of the kernel, which is GPL version 2. Update the files which contain no license information with the 'GPL-2.0' SPDX license identifier. The SPDX identifier is a legally binding shorthand, which can be used instead of the full boiler plate text. This patch is based on work done by Thomas Gleixner and Kate Stewart and Philippe Ombredanne. How this work was done: Patches were generated and checked against linux-4.14-rc6 for a subset of the use cases: - file had no licensing information it it. - file was a */uapi/* one with no licensing information in it, - file was a */uapi/* one with existing licensing information, Further patches will be generated in subsequent months to fix up cases where non-standard license headers were used, and references to license had to be inferred by heuristics based on keywords. The analysis to determine which SPDX License Identifier to be applied to a file was done in a spreadsheet of side by side results from of the output of two independent scanners (ScanCode & Windriver) producing SPDX tag:value files created by Philippe Ombredanne. Philippe prepared the base worksheet, and did an initial spot review of a few 1000 files. The 4.13 kernel was the starting point of the analysis with 60,537 files assessed. Kate Stewart did a file by file comparison of the scanner results in the spreadsheet to determine which SPDX license identifier(s) to be applied to the file. She confirmed any determination that was not immediately clear with lawyers working with the Linux Foundation. Criteria used to select files for SPDX license identifier tagging was: - Files considered eligible had to be source code files. - Make and config files were included as candidates if they contained >5 lines of source - File already had some variant of a license header in it (even if <5 lines). All documentation files were explicitly excluded. The following heuristics were used to determine which SPDX license identifiers to apply. - when both scanners couldn't find any license traces, file was considered to have no license information in it, and the top level COPYING file license applied. For non */uapi/* files that summary was: SPDX license identifier # files ---------------------------------------------------|------- GPL-2.0 11139 and resulted in the first patch in this series. If that file was a */uapi/* path one, it was "GPL-2.0 WITH Linux-syscall-note" otherwise it was "GPL-2.0". Results of that was: SPDX license identifier # files ---------------------------------------------------|------- GPL-2.0 WITH Linux-syscall-note 930 and resulted in the second patch in this series. - if a file had some form of licensing information in it, and was one of the */uapi/* ones, it was denoted with the Linux-syscall-note if any GPL family license was found in the file or had no licensing in it (per prior point). Results summary: SPDX license identifier # files ---------------------------------------------------|------ GPL-2.0 WITH Linux-syscall-note 270 GPL-2.0+ WITH Linux-syscall-note 169 ((GPL-2.0 WITH Linux-syscall-note) OR BSD-2-Clause) 21 ((GPL-2.0 WITH Linux-syscall-note) OR BSD-3-Clause) 17 LGPL-2.1+ WITH Linux-syscall-note 15 GPL-1.0+ WITH Linux-syscall-note 14 ((GPL-2.0+ WITH Linux-syscall-note) OR BSD-3-Clause) 5 LGPL-2.0+ WITH Linux-syscall-note 4 LGPL-2.1 WITH Linux-syscall-note 3 ((GPL-2.0 WITH Linux-syscall-note) OR MIT) 3 ((GPL-2.0 WITH Linux-syscall-note) AND MIT) 1 and that resulted in the third patch in this series. - when the two scanners agreed on the detected license(s), that became the concluded license(s). - when there was disagreement between the two scanners (one detected a license but the other didn't, or they both detected different licenses) a manual inspection of the file occurred. - In most cases a manual inspection of the information in the file resulted in a clear resolution of the license that should apply (and which scanner probably needed to revisit its heuristics). - When it was not immediately clear, the license identifier was confirmed with lawyers working with the Linux Foundation. - If there was any question as to the appropriate license identifier, the file was flagged for further research and to be revisited later in time. In total, over 70 hours of logged manual review was done on the spreadsheet to determine the SPDX license identifiers to apply to the source files by Kate, Philippe, Thomas and, in some cases, confirmation by lawyers working with the Linux Foundation. Kate also obtained a third independent scan of the 4.13 code base from FOSSology, and compared selected files where the other two scanners disagreed against that SPDX file, to see if there was new insights. The Windriver scanner is based on an older version of FOSSology in part, so they are related. Thomas did random spot checks in about 500 files from the spreadsheets for the uapi headers and agreed with SPDX license identifier in the files he inspected. For the non-uapi files Thomas did random spot checks in about 15000 files. In initial set of patches against 4.14-rc6, 3 files were found to have copy/paste license identifier errors, and have been fixed to reflect the correct identifier. Additionally Philippe spent 10 hours this week doing a detailed manual inspection and review of the 12,461 patched files from the initial patch version early this week with: - a full scancode scan run, collecting the matched texts, detected license ids and scores - reviewing anything where there was a license detected (about 500+ files) to ensure that the applied SPDX license was correct - reviewing anything where there was no detection but the patch license was not GPL-2.0 WITH Linux-syscall-note to ensure that the applied SPDX license was correct This produced a worksheet with 20 files needing minor correction. This worksheet was then exported into 3 different .csv files for the different types of files to be modified. These .csv files were then reviewed by Greg. Thomas wrote a script to parse the csv files and add the proper SPDX tag to the file, in the format that the file expected. This script was further refined by Greg based on the output to detect more types of files automatically and to distinguish between header and source .c files (which need different comment types.) Finally Greg ran the script using the .csv files to generate the patches. Reviewed-by: Kate Stewart <kstewart@linuxfoundation.org> Reviewed-by: Philippe Ombredanne <pombredanne@nexb.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2017-11-01 14:07:57 +00:00
// SPDX-License-Identifier: GPL-2.0
/*
* Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
*
* Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
*
* Interactivity improvements by Mike Galbraith
* (C) 2007 Mike Galbraith <efault@gmx.de>
*
* Various enhancements by Dmitry Adamushko.
* (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
*
* Group scheduling enhancements by Srivatsa Vaddagiri
* Copyright IBM Corporation, 2007
* Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
*
* Scaled math optimizations by Thomas Gleixner
* Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
*
* Adaptive scheduling granularity, math enhancements by Peter Zijlstra
* Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
*/
#include "sched.h"
#include <trace/events/sched.h>
/*
* Targeted preemption latency for CPU-bound tasks:
*
* NOTE: this latency value is not the same as the concept of
* 'timeslice length' - timeslices in CFS are of variable length
* and have no persistent notion like in traditional, time-slice
* based scheduling concepts.
*
* (to see the precise effective timeslice length of your workload,
* run vmstat and monitor the context-switches (cs) field)
*
* (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
*/
unsigned int sysctl_sched_latency = 6000000ULL;
static unsigned int normalized_sysctl_sched_latency = 6000000ULL;
/*
* The initial- and re-scaling of tunables is configurable
*
* Options are:
*
* SCHED_TUNABLESCALING_NONE - unscaled, always *1
* SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
* SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
*
* (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
*/
enum sched_tunable_scaling sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG;
/*
* Minimal preemption granularity for CPU-bound tasks:
*
* (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
*/
unsigned int sysctl_sched_min_granularity = 750000ULL;
static unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
/*
* This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity
*/
static unsigned int sched_nr_latency = 8;
/*
* After fork, child runs first. If set to 0 (default) then
* parent will (try to) run first.
*/
unsigned int sysctl_sched_child_runs_first __read_mostly;
/*
* SCHED_OTHER wake-up granularity.
*
* This option delays the preemption effects of decoupled workloads
* and reduces their over-scheduling. Synchronous workloads will still
* have immediate wakeup/sleep latencies.
*
* (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
*/
unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
static unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
int sched_thermal_decay_shift;
static int __init setup_sched_thermal_decay_shift(char *str)
{
int _shift = 0;
if (kstrtoint(str, 0, &_shift))
pr_warn("Unable to set scheduler thermal pressure decay shift parameter\n");
sched_thermal_decay_shift = clamp(_shift, 0, 10);
return 1;
}
__setup("sched_thermal_decay_shift=", setup_sched_thermal_decay_shift);
#ifdef CONFIG_SMP
/*
* For asym packing, by default the lower numbered CPU has higher priority.
*/
int __weak arch_asym_cpu_priority(int cpu)
{
return -cpu;
}
/*
* The margin used when comparing utilization with CPU capacity.
*
* (default: ~20%)
*/
#define fits_capacity(cap, max) ((cap) * 1280 < (max) * 1024)
#endif
#ifdef CONFIG_CFS_BANDWIDTH
/*
* Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
* each time a cfs_rq requests quota.
*
* Note: in the case that the slice exceeds the runtime remaining (either due
* to consumption or the quota being specified to be smaller than the slice)
* we will always only issue the remaining available time.
*
* (default: 5 msec, units: microseconds)
*/
unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
#endif
static inline void update_load_add(struct load_weight *lw, unsigned long inc)
{
lw->weight += inc;
lw->inv_weight = 0;
}
static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
{
lw->weight -= dec;
lw->inv_weight = 0;
}
static inline void update_load_set(struct load_weight *lw, unsigned long w)
{
lw->weight = w;
lw->inv_weight = 0;
}
/*
* Increase the granularity value when there are more CPUs,
* because with more CPUs the 'effective latency' as visible
* to users decreases. But the relationship is not linear,
* so pick a second-best guess by going with the log2 of the
* number of CPUs.
*
* This idea comes from the SD scheduler of Con Kolivas:
*/
static unsigned int get_update_sysctl_factor(void)
{
unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
unsigned int factor;
switch (sysctl_sched_tunable_scaling) {
case SCHED_TUNABLESCALING_NONE:
factor = 1;
break;
case SCHED_TUNABLESCALING_LINEAR:
factor = cpus;
break;
case SCHED_TUNABLESCALING_LOG:
default:
factor = 1 + ilog2(cpus);
break;
}
return factor;
}
static void update_sysctl(void)
{
unsigned int factor = get_update_sysctl_factor();
#define SET_SYSCTL(name) \
(sysctl_##name = (factor) * normalized_sysctl_##name)
SET_SYSCTL(sched_min_granularity);
SET_SYSCTL(sched_latency);
SET_SYSCTL(sched_wakeup_granularity);
#undef SET_SYSCTL
}
void sched_init_granularity(void)
{
update_sysctl();
}
sched/fair: Rework sched_fair time accounting Christian suffers from a bad BIOS that wrecks his i5's TSC sync. This results in him occasionally seeing time going backwards - which crashes the scheduler ... Most of our time accounting can actually handle that except the most common one; the tick time update of sched_fair. There is a further problem with that code; previously we assumed that because we get a tick every TICK_NSEC our time delta could never exceed 32bits and math was simpler. However, ever since Frederic managed to get NO_HZ_FULL merged; this is no longer the case since now a task can run for a long time indeed without getting a tick. It only takes about ~4.2 seconds to overflow our u32 in nanoseconds. This means we not only need to better deal with time going backwards; but also means we need to be able to deal with large deltas. This patch reworks the entire code and uses mul_u64_u32_shr() as proposed by Andy a long while ago. We express our virtual time scale factor in a u32 multiplier and shift right and the 32bit mul_u64_u32_shr() implementation reduces to a single 32x32->64 multiply if the time delta is still short (common case). For 64bit a 64x64->128 multiply can be used if ARCH_SUPPORTS_INT128. Reported-and-Tested-by: Christian Engelmayer <cengelma@gmx.at> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Cc: fweisbec@gmail.com Cc: Paul Turner <pjt@google.com> Cc: Stanislaw Gruszka <sgruszka@redhat.com> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Link: http://lkml.kernel.org/r/20131118172706.GI3866@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-11-18 17:27:06 +00:00
#define WMULT_CONST (~0U)
#define WMULT_SHIFT 32
sched/fair: Rework sched_fair time accounting Christian suffers from a bad BIOS that wrecks his i5's TSC sync. This results in him occasionally seeing time going backwards - which crashes the scheduler ... Most of our time accounting can actually handle that except the most common one; the tick time update of sched_fair. There is a further problem with that code; previously we assumed that because we get a tick every TICK_NSEC our time delta could never exceed 32bits and math was simpler. However, ever since Frederic managed to get NO_HZ_FULL merged; this is no longer the case since now a task can run for a long time indeed without getting a tick. It only takes about ~4.2 seconds to overflow our u32 in nanoseconds. This means we not only need to better deal with time going backwards; but also means we need to be able to deal with large deltas. This patch reworks the entire code and uses mul_u64_u32_shr() as proposed by Andy a long while ago. We express our virtual time scale factor in a u32 multiplier and shift right and the 32bit mul_u64_u32_shr() implementation reduces to a single 32x32->64 multiply if the time delta is still short (common case). For 64bit a 64x64->128 multiply can be used if ARCH_SUPPORTS_INT128. Reported-and-Tested-by: Christian Engelmayer <cengelma@gmx.at> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Cc: fweisbec@gmail.com Cc: Paul Turner <pjt@google.com> Cc: Stanislaw Gruszka <sgruszka@redhat.com> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Link: http://lkml.kernel.org/r/20131118172706.GI3866@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-11-18 17:27:06 +00:00
static void __update_inv_weight(struct load_weight *lw)
{
unsigned long w;
if (likely(lw->inv_weight))
return;
w = scale_load_down(lw->weight);
if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
lw->inv_weight = 1;
else if (unlikely(!w))
lw->inv_weight = WMULT_CONST;
else
lw->inv_weight = WMULT_CONST / w;
}
/*
sched/fair: Rework sched_fair time accounting Christian suffers from a bad BIOS that wrecks his i5's TSC sync. This results in him occasionally seeing time going backwards - which crashes the scheduler ... Most of our time accounting can actually handle that except the most common one; the tick time update of sched_fair. There is a further problem with that code; previously we assumed that because we get a tick every TICK_NSEC our time delta could never exceed 32bits and math was simpler. However, ever since Frederic managed to get NO_HZ_FULL merged; this is no longer the case since now a task can run for a long time indeed without getting a tick. It only takes about ~4.2 seconds to overflow our u32 in nanoseconds. This means we not only need to better deal with time going backwards; but also means we need to be able to deal with large deltas. This patch reworks the entire code and uses mul_u64_u32_shr() as proposed by Andy a long while ago. We express our virtual time scale factor in a u32 multiplier and shift right and the 32bit mul_u64_u32_shr() implementation reduces to a single 32x32->64 multiply if the time delta is still short (common case). For 64bit a 64x64->128 multiply can be used if ARCH_SUPPORTS_INT128. Reported-and-Tested-by: Christian Engelmayer <cengelma@gmx.at> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Cc: fweisbec@gmail.com Cc: Paul Turner <pjt@google.com> Cc: Stanislaw Gruszka <sgruszka@redhat.com> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Link: http://lkml.kernel.org/r/20131118172706.GI3866@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-11-18 17:27:06 +00:00
* delta_exec * weight / lw.weight
* OR
* (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
*
* Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
sched/fair: Rework sched_fair time accounting Christian suffers from a bad BIOS that wrecks his i5's TSC sync. This results in him occasionally seeing time going backwards - which crashes the scheduler ... Most of our time accounting can actually handle that except the most common one; the tick time update of sched_fair. There is a further problem with that code; previously we assumed that because we get a tick every TICK_NSEC our time delta could never exceed 32bits and math was simpler. However, ever since Frederic managed to get NO_HZ_FULL merged; this is no longer the case since now a task can run for a long time indeed without getting a tick. It only takes about ~4.2 seconds to overflow our u32 in nanoseconds. This means we not only need to better deal with time going backwards; but also means we need to be able to deal with large deltas. This patch reworks the entire code and uses mul_u64_u32_shr() as proposed by Andy a long while ago. We express our virtual time scale factor in a u32 multiplier and shift right and the 32bit mul_u64_u32_shr() implementation reduces to a single 32x32->64 multiply if the time delta is still short (common case). For 64bit a 64x64->128 multiply can be used if ARCH_SUPPORTS_INT128. Reported-and-Tested-by: Christian Engelmayer <cengelma@gmx.at> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Cc: fweisbec@gmail.com Cc: Paul Turner <pjt@google.com> Cc: Stanislaw Gruszka <sgruszka@redhat.com> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Link: http://lkml.kernel.org/r/20131118172706.GI3866@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-11-18 17:27:06 +00:00
* we're guaranteed shift stays positive because inv_weight is guaranteed to
* fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
*
* Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
* weight/lw.weight <= 1, and therefore our shift will also be positive.
*/
sched/fair: Rework sched_fair time accounting Christian suffers from a bad BIOS that wrecks his i5's TSC sync. This results in him occasionally seeing time going backwards - which crashes the scheduler ... Most of our time accounting can actually handle that except the most common one; the tick time update of sched_fair. There is a further problem with that code; previously we assumed that because we get a tick every TICK_NSEC our time delta could never exceed 32bits and math was simpler. However, ever since Frederic managed to get NO_HZ_FULL merged; this is no longer the case since now a task can run for a long time indeed without getting a tick. It only takes about ~4.2 seconds to overflow our u32 in nanoseconds. This means we not only need to better deal with time going backwards; but also means we need to be able to deal with large deltas. This patch reworks the entire code and uses mul_u64_u32_shr() as proposed by Andy a long while ago. We express our virtual time scale factor in a u32 multiplier and shift right and the 32bit mul_u64_u32_shr() implementation reduces to a single 32x32->64 multiply if the time delta is still short (common case). For 64bit a 64x64->128 multiply can be used if ARCH_SUPPORTS_INT128. Reported-and-Tested-by: Christian Engelmayer <cengelma@gmx.at> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Cc: fweisbec@gmail.com Cc: Paul Turner <pjt@google.com> Cc: Stanislaw Gruszka <sgruszka@redhat.com> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Link: http://lkml.kernel.org/r/20131118172706.GI3866@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-11-18 17:27:06 +00:00
static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
{
sched/fair: Rework sched_fair time accounting Christian suffers from a bad BIOS that wrecks his i5's TSC sync. This results in him occasionally seeing time going backwards - which crashes the scheduler ... Most of our time accounting can actually handle that except the most common one; the tick time update of sched_fair. There is a further problem with that code; previously we assumed that because we get a tick every TICK_NSEC our time delta could never exceed 32bits and math was simpler. However, ever since Frederic managed to get NO_HZ_FULL merged; this is no longer the case since now a task can run for a long time indeed without getting a tick. It only takes about ~4.2 seconds to overflow our u32 in nanoseconds. This means we not only need to better deal with time going backwards; but also means we need to be able to deal with large deltas. This patch reworks the entire code and uses mul_u64_u32_shr() as proposed by Andy a long while ago. We express our virtual time scale factor in a u32 multiplier and shift right and the 32bit mul_u64_u32_shr() implementation reduces to a single 32x32->64 multiply if the time delta is still short (common case). For 64bit a 64x64->128 multiply can be used if ARCH_SUPPORTS_INT128. Reported-and-Tested-by: Christian Engelmayer <cengelma@gmx.at> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Cc: fweisbec@gmail.com Cc: Paul Turner <pjt@google.com> Cc: Stanislaw Gruszka <sgruszka@redhat.com> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Link: http://lkml.kernel.org/r/20131118172706.GI3866@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-11-18 17:27:06 +00:00
u64 fact = scale_load_down(weight);
int shift = WMULT_SHIFT;
sched/fair: Rework sched_fair time accounting Christian suffers from a bad BIOS that wrecks his i5's TSC sync. This results in him occasionally seeing time going backwards - which crashes the scheduler ... Most of our time accounting can actually handle that except the most common one; the tick time update of sched_fair. There is a further problem with that code; previously we assumed that because we get a tick every TICK_NSEC our time delta could never exceed 32bits and math was simpler. However, ever since Frederic managed to get NO_HZ_FULL merged; this is no longer the case since now a task can run for a long time indeed without getting a tick. It only takes about ~4.2 seconds to overflow our u32 in nanoseconds. This means we not only need to better deal with time going backwards; but also means we need to be able to deal with large deltas. This patch reworks the entire code and uses mul_u64_u32_shr() as proposed by Andy a long while ago. We express our virtual time scale factor in a u32 multiplier and shift right and the 32bit mul_u64_u32_shr() implementation reduces to a single 32x32->64 multiply if the time delta is still short (common case). For 64bit a 64x64->128 multiply can be used if ARCH_SUPPORTS_INT128. Reported-and-Tested-by: Christian Engelmayer <cengelma@gmx.at> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Cc: fweisbec@gmail.com Cc: Paul Turner <pjt@google.com> Cc: Stanislaw Gruszka <sgruszka@redhat.com> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Link: http://lkml.kernel.org/r/20131118172706.GI3866@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-11-18 17:27:06 +00:00
__update_inv_weight(lw);
sched/fair: Rework sched_fair time accounting Christian suffers from a bad BIOS that wrecks his i5's TSC sync. This results in him occasionally seeing time going backwards - which crashes the scheduler ... Most of our time accounting can actually handle that except the most common one; the tick time update of sched_fair. There is a further problem with that code; previously we assumed that because we get a tick every TICK_NSEC our time delta could never exceed 32bits and math was simpler. However, ever since Frederic managed to get NO_HZ_FULL merged; this is no longer the case since now a task can run for a long time indeed without getting a tick. It only takes about ~4.2 seconds to overflow our u32 in nanoseconds. This means we not only need to better deal with time going backwards; but also means we need to be able to deal with large deltas. This patch reworks the entire code and uses mul_u64_u32_shr() as proposed by Andy a long while ago. We express our virtual time scale factor in a u32 multiplier and shift right and the 32bit mul_u64_u32_shr() implementation reduces to a single 32x32->64 multiply if the time delta is still short (common case). For 64bit a 64x64->128 multiply can be used if ARCH_SUPPORTS_INT128. Reported-and-Tested-by: Christian Engelmayer <cengelma@gmx.at> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Cc: fweisbec@gmail.com Cc: Paul Turner <pjt@google.com> Cc: Stanislaw Gruszka <sgruszka@redhat.com> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Link: http://lkml.kernel.org/r/20131118172706.GI3866@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-11-18 17:27:06 +00:00
if (unlikely(fact >> 32)) {
while (fact >> 32) {
fact >>= 1;
shift--;
}
}
fact = mul_u32_u32(fact, lw->inv_weight);
sched/fair: Rework sched_fair time accounting Christian suffers from a bad BIOS that wrecks his i5's TSC sync. This results in him occasionally seeing time going backwards - which crashes the scheduler ... Most of our time accounting can actually handle that except the most common one; the tick time update of sched_fair. There is a further problem with that code; previously we assumed that because we get a tick every TICK_NSEC our time delta could never exceed 32bits and math was simpler. However, ever since Frederic managed to get NO_HZ_FULL merged; this is no longer the case since now a task can run for a long time indeed without getting a tick. It only takes about ~4.2 seconds to overflow our u32 in nanoseconds. This means we not only need to better deal with time going backwards; but also means we need to be able to deal with large deltas. This patch reworks the entire code and uses mul_u64_u32_shr() as proposed by Andy a long while ago. We express our virtual time scale factor in a u32 multiplier and shift right and the 32bit mul_u64_u32_shr() implementation reduces to a single 32x32->64 multiply if the time delta is still short (common case). For 64bit a 64x64->128 multiply can be used if ARCH_SUPPORTS_INT128. Reported-and-Tested-by: Christian Engelmayer <cengelma@gmx.at> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Cc: fweisbec@gmail.com Cc: Paul Turner <pjt@google.com> Cc: Stanislaw Gruszka <sgruszka@redhat.com> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Link: http://lkml.kernel.org/r/20131118172706.GI3866@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-11-18 17:27:06 +00:00
while (fact >> 32) {
fact >>= 1;
shift--;
}
sched/fair: Rework sched_fair time accounting Christian suffers from a bad BIOS that wrecks his i5's TSC sync. This results in him occasionally seeing time going backwards - which crashes the scheduler ... Most of our time accounting can actually handle that except the most common one; the tick time update of sched_fair. There is a further problem with that code; previously we assumed that because we get a tick every TICK_NSEC our time delta could never exceed 32bits and math was simpler. However, ever since Frederic managed to get NO_HZ_FULL merged; this is no longer the case since now a task can run for a long time indeed without getting a tick. It only takes about ~4.2 seconds to overflow our u32 in nanoseconds. This means we not only need to better deal with time going backwards; but also means we need to be able to deal with large deltas. This patch reworks the entire code and uses mul_u64_u32_shr() as proposed by Andy a long while ago. We express our virtual time scale factor in a u32 multiplier and shift right and the 32bit mul_u64_u32_shr() implementation reduces to a single 32x32->64 multiply if the time delta is still short (common case). For 64bit a 64x64->128 multiply can be used if ARCH_SUPPORTS_INT128. Reported-and-Tested-by: Christian Engelmayer <cengelma@gmx.at> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Cc: fweisbec@gmail.com Cc: Paul Turner <pjt@google.com> Cc: Stanislaw Gruszka <sgruszka@redhat.com> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Link: http://lkml.kernel.org/r/20131118172706.GI3866@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-11-18 17:27:06 +00:00
return mul_u64_u32_shr(delta_exec, fact, shift);
}
const struct sched_class fair_sched_class;
/**************************************************************
* CFS operations on generic schedulable entities:
*/
#ifdef CONFIG_FAIR_GROUP_SCHED
static inline struct task_struct *task_of(struct sched_entity *se)
{
SCHED_WARN_ON(!entity_is_task(se));
return container_of(se, struct task_struct, se);
}
/* Walk up scheduling entities hierarchy */
#define for_each_sched_entity(se) \
for (; se; se = se->parent)
static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
{
return p->se.cfs_rq;
}
/* runqueue on which this entity is (to be) queued */
static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
{
return se->cfs_rq;
}
/* runqueue "owned" by this group */
static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
{
return grp->my_q;
}
static inline void cfs_rq_tg_path(struct cfs_rq *cfs_rq, char *path, int len)
{
if (!path)
return;
if (cfs_rq && task_group_is_autogroup(cfs_rq->tg))
autogroup_path(cfs_rq->tg, path, len);
else if (cfs_rq && cfs_rq->tg->css.cgroup)
cgroup_path(cfs_rq->tg->css.cgroup, path, len);
else
strlcpy(path, "(null)", len);
}
sched/fair: Fix insertion in rq->leaf_cfs_rq_list Sargun reported a crash: "I picked up c40f7d74c741a907cfaeb73a7697081881c497d0 sched/fair: Fix infinite loop in update_blocked_averages() by reverting a9e7f6544b9c and put it on top of 4.19.13. In addition to this, I uninlined list_add_leaf_cfs_rq for debugging. This revealed a new bug that we didn't get to because we kept getting crashes from the previous issue. When we are running with cgroups that are rapidly changing, with CFS bandwidth control, and in addition using the cpusets cgroup, we see this crash. Specifically, it seems to occur with cgroups that are throttled and we change the allowed cpuset." The algorithm used to order cfs_rq in rq->leaf_cfs_rq_list assumes that it will walk down to root the 1st time a cfs_rq is used and we will finish to add either a cfs_rq without parent or a cfs_rq with a parent that is already on the list. But this is not always true in presence of throttling. Because a cfs_rq can be throttled even if it has never been used but other CPUs of the cgroup have already used all the bandwdith, we are not sure to go down to the root and add all cfs_rq in the list. Ensure that all cfs_rq will be added in the list even if they are throttled. [ mingo: Fix !CGROUPS build. ] Reported-by: Sargun Dhillon <sargun@sargun.me> Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: tj@kernel.org Fixes: 9c2791f936ef ("Fix hierarchical order in rq->leaf_cfs_rq_list") Link: https://lkml.kernel.org/r/1548825767-10799-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-01-30 05:22:47 +00:00
static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
{
struct rq *rq = rq_of(cfs_rq);
int cpu = cpu_of(rq);
if (cfs_rq->on_list)
sched/fair: Fix insertion in rq->leaf_cfs_rq_list Sargun reported a crash: "I picked up c40f7d74c741a907cfaeb73a7697081881c497d0 sched/fair: Fix infinite loop in update_blocked_averages() by reverting a9e7f6544b9c and put it on top of 4.19.13. In addition to this, I uninlined list_add_leaf_cfs_rq for debugging. This revealed a new bug that we didn't get to because we kept getting crashes from the previous issue. When we are running with cgroups that are rapidly changing, with CFS bandwidth control, and in addition using the cpusets cgroup, we see this crash. Specifically, it seems to occur with cgroups that are throttled and we change the allowed cpuset." The algorithm used to order cfs_rq in rq->leaf_cfs_rq_list assumes that it will walk down to root the 1st time a cfs_rq is used and we will finish to add either a cfs_rq without parent or a cfs_rq with a parent that is already on the list. But this is not always true in presence of throttling. Because a cfs_rq can be throttled even if it has never been used but other CPUs of the cgroup have already used all the bandwdith, we are not sure to go down to the root and add all cfs_rq in the list. Ensure that all cfs_rq will be added in the list even if they are throttled. [ mingo: Fix !CGROUPS build. ] Reported-by: Sargun Dhillon <sargun@sargun.me> Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: tj@kernel.org Fixes: 9c2791f936ef ("Fix hierarchical order in rq->leaf_cfs_rq_list") Link: https://lkml.kernel.org/r/1548825767-10799-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-01-30 05:22:47 +00:00
return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list;
cfs_rq->on_list = 1;
/*
* Ensure we either appear before our parent (if already
* enqueued) or force our parent to appear after us when it is
* enqueued. The fact that we always enqueue bottom-up
* reduces this to two cases and a special case for the root
* cfs_rq. Furthermore, it also means that we will always reset
* tmp_alone_branch either when the branch is connected
* to a tree or when we reach the top of the tree
*/
if (cfs_rq->tg->parent &&
cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
/*
* If parent is already on the list, we add the child
* just before. Thanks to circular linked property of
* the list, this means to put the child at the tail
* of the list that starts by parent.
*/
list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
&(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
/*
* The branch is now connected to its tree so we can
* reset tmp_alone_branch to the beginning of the
* list.
*/
rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
sched/fair: Fix insertion in rq->leaf_cfs_rq_list Sargun reported a crash: "I picked up c40f7d74c741a907cfaeb73a7697081881c497d0 sched/fair: Fix infinite loop in update_blocked_averages() by reverting a9e7f6544b9c and put it on top of 4.19.13. In addition to this, I uninlined list_add_leaf_cfs_rq for debugging. This revealed a new bug that we didn't get to because we kept getting crashes from the previous issue. When we are running with cgroups that are rapidly changing, with CFS bandwidth control, and in addition using the cpusets cgroup, we see this crash. Specifically, it seems to occur with cgroups that are throttled and we change the allowed cpuset." The algorithm used to order cfs_rq in rq->leaf_cfs_rq_list assumes that it will walk down to root the 1st time a cfs_rq is used and we will finish to add either a cfs_rq without parent or a cfs_rq with a parent that is already on the list. But this is not always true in presence of throttling. Because a cfs_rq can be throttled even if it has never been used but other CPUs of the cgroup have already used all the bandwdith, we are not sure to go down to the root and add all cfs_rq in the list. Ensure that all cfs_rq will be added in the list even if they are throttled. [ mingo: Fix !CGROUPS build. ] Reported-by: Sargun Dhillon <sargun@sargun.me> Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: tj@kernel.org Fixes: 9c2791f936ef ("Fix hierarchical order in rq->leaf_cfs_rq_list") Link: https://lkml.kernel.org/r/1548825767-10799-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-01-30 05:22:47 +00:00
return true;
}
if (!cfs_rq->tg->parent) {
/*
* cfs rq without parent should be put
* at the tail of the list.
*/
list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
&rq->leaf_cfs_rq_list);
/*
* We have reach the top of a tree so we can reset
* tmp_alone_branch to the beginning of the list.
*/
rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
sched/fair: Fix insertion in rq->leaf_cfs_rq_list Sargun reported a crash: "I picked up c40f7d74c741a907cfaeb73a7697081881c497d0 sched/fair: Fix infinite loop in update_blocked_averages() by reverting a9e7f6544b9c and put it on top of 4.19.13. In addition to this, I uninlined list_add_leaf_cfs_rq for debugging. This revealed a new bug that we didn't get to because we kept getting crashes from the previous issue. When we are running with cgroups that are rapidly changing, with CFS bandwidth control, and in addition using the cpusets cgroup, we see this crash. Specifically, it seems to occur with cgroups that are throttled and we change the allowed cpuset." The algorithm used to order cfs_rq in rq->leaf_cfs_rq_list assumes that it will walk down to root the 1st time a cfs_rq is used and we will finish to add either a cfs_rq without parent or a cfs_rq with a parent that is already on the list. But this is not always true in presence of throttling. Because a cfs_rq can be throttled even if it has never been used but other CPUs of the cgroup have already used all the bandwdith, we are not sure to go down to the root and add all cfs_rq in the list. Ensure that all cfs_rq will be added in the list even if they are throttled. [ mingo: Fix !CGROUPS build. ] Reported-by: Sargun Dhillon <sargun@sargun.me> Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: tj@kernel.org Fixes: 9c2791f936ef ("Fix hierarchical order in rq->leaf_cfs_rq_list") Link: https://lkml.kernel.org/r/1548825767-10799-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-01-30 05:22:47 +00:00
return true;
}
/*
* The parent has not already been added so we want to
* make sure that it will be put after us.
* tmp_alone_branch points to the begin of the branch
* where we will add parent.
*/
list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch);
/*
* update tmp_alone_branch to points to the new begin
* of the branch
*/
rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
sched/fair: Fix insertion in rq->leaf_cfs_rq_list Sargun reported a crash: "I picked up c40f7d74c741a907cfaeb73a7697081881c497d0 sched/fair: Fix infinite loop in update_blocked_averages() by reverting a9e7f6544b9c and put it on top of 4.19.13. In addition to this, I uninlined list_add_leaf_cfs_rq for debugging. This revealed a new bug that we didn't get to because we kept getting crashes from the previous issue. When we are running with cgroups that are rapidly changing, with CFS bandwidth control, and in addition using the cpusets cgroup, we see this crash. Specifically, it seems to occur with cgroups that are throttled and we change the allowed cpuset." The algorithm used to order cfs_rq in rq->leaf_cfs_rq_list assumes that it will walk down to root the 1st time a cfs_rq is used and we will finish to add either a cfs_rq without parent or a cfs_rq with a parent that is already on the list. But this is not always true in presence of throttling. Because a cfs_rq can be throttled even if it has never been used but other CPUs of the cgroup have already used all the bandwdith, we are not sure to go down to the root and add all cfs_rq in the list. Ensure that all cfs_rq will be added in the list even if they are throttled. [ mingo: Fix !CGROUPS build. ] Reported-by: Sargun Dhillon <sargun@sargun.me> Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: tj@kernel.org Fixes: 9c2791f936ef ("Fix hierarchical order in rq->leaf_cfs_rq_list") Link: https://lkml.kernel.org/r/1548825767-10799-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-01-30 05:22:47 +00:00
return false;
}
static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
{
if (cfs_rq->on_list) {
sched/fair: Optimize update_blocked_averages() Removing a cfs_rq from rq->leaf_cfs_rq_list can break the parent/child ordering of the list when it will be added back. In order to remove an empty and fully decayed cfs_rq, we must remove its children too, so they will be added back in the right order next time. With a normal decay of PELT, a parent will be empty and fully decayed if all children are empty and fully decayed too. In such a case, we just have to ensure that the whole branch will be added when a new task is enqueued. This is default behavior since : commit f6783319737f ("sched/fair: Fix insertion in rq->leaf_cfs_rq_list") In case of throttling, the PELT of throttled cfs_rq will not be updated whereas the parent will. This breaks the assumption made above unless we remove the children of a cfs_rq that is throttled. Then, they will be added back when unthrottled and a sched_entity will be enqueued. As throttled cfs_rq are now removed from the list, we can remove the associated test in update_blocked_averages(). Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: sargun@sargun.me Cc: tj@kernel.org Cc: xiexiuqi@huawei.com Cc: xiezhipeng1@huawei.com Link: https://lkml.kernel.org/r/1549469662-13614-2-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-02-06 16:14:21 +00:00
struct rq *rq = rq_of(cfs_rq);
/*
* With cfs_rq being unthrottled/throttled during an enqueue,
* it can happen the tmp_alone_branch points the a leaf that
* we finally want to del. In this case, tmp_alone_branch moves
* to the prev element but it will point to rq->leaf_cfs_rq_list
* at the end of the enqueue.
*/
if (rq->tmp_alone_branch == &cfs_rq->leaf_cfs_rq_list)
rq->tmp_alone_branch = cfs_rq->leaf_cfs_rq_list.prev;
list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
cfs_rq->on_list = 0;
}
}
static inline void assert_list_leaf_cfs_rq(struct rq *rq)
{
SCHED_WARN_ON(rq->tmp_alone_branch != &rq->leaf_cfs_rq_list);
}
/* Iterate thr' all leaf cfs_rq's on a runqueue */
#define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \
leaf_cfs_rq_list)
/* Do the two (enqueued) entities belong to the same group ? */
static inline struct cfs_rq *
is_same_group(struct sched_entity *se, struct sched_entity *pse)
{
if (se->cfs_rq == pse->cfs_rq)
return se->cfs_rq;
return NULL;
}
static inline struct sched_entity *parent_entity(struct sched_entity *se)
{
return se->parent;
}
static void
find_matching_se(struct sched_entity **se, struct sched_entity **pse)
{
int se_depth, pse_depth;
/*
* preemption test can be made between sibling entities who are in the
* same cfs_rq i.e who have a common parent. Walk up the hierarchy of
* both tasks until we find their ancestors who are siblings of common
* parent.
*/
/* First walk up until both entities are at same depth */
se_depth = (*se)->depth;
pse_depth = (*pse)->depth;
while (se_depth > pse_depth) {
se_depth--;
*se = parent_entity(*se);
}
while (pse_depth > se_depth) {
pse_depth--;
*pse = parent_entity(*pse);
}
while (!is_same_group(*se, *pse)) {
*se = parent_entity(*se);
*pse = parent_entity(*pse);
}
}
#else /* !CONFIG_FAIR_GROUP_SCHED */
static inline struct task_struct *task_of(struct sched_entity *se)
{
return container_of(se, struct task_struct, se);
}
#define for_each_sched_entity(se) \
for (; se; se = NULL)
static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
{
return &task_rq(p)->cfs;
}
static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
{
struct task_struct *p = task_of(se);
struct rq *rq = task_rq(p);
return &rq->cfs;
}
/* runqueue "owned" by this group */
static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
{
return NULL;
}
static inline void cfs_rq_tg_path(struct cfs_rq *cfs_rq, char *path, int len)
{
if (path)
strlcpy(path, "(null)", len);
}
sched/fair: Fix insertion in rq->leaf_cfs_rq_list Sargun reported a crash: "I picked up c40f7d74c741a907cfaeb73a7697081881c497d0 sched/fair: Fix infinite loop in update_blocked_averages() by reverting a9e7f6544b9c and put it on top of 4.19.13. In addition to this, I uninlined list_add_leaf_cfs_rq for debugging. This revealed a new bug that we didn't get to because we kept getting crashes from the previous issue. When we are running with cgroups that are rapidly changing, with CFS bandwidth control, and in addition using the cpusets cgroup, we see this crash. Specifically, it seems to occur with cgroups that are throttled and we change the allowed cpuset." The algorithm used to order cfs_rq in rq->leaf_cfs_rq_list assumes that it will walk down to root the 1st time a cfs_rq is used and we will finish to add either a cfs_rq without parent or a cfs_rq with a parent that is already on the list. But this is not always true in presence of throttling. Because a cfs_rq can be throttled even if it has never been used but other CPUs of the cgroup have already used all the bandwdith, we are not sure to go down to the root and add all cfs_rq in the list. Ensure that all cfs_rq will be added in the list even if they are throttled. [ mingo: Fix !CGROUPS build. ] Reported-by: Sargun Dhillon <sargun@sargun.me> Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: tj@kernel.org Fixes: 9c2791f936ef ("Fix hierarchical order in rq->leaf_cfs_rq_list") Link: https://lkml.kernel.org/r/1548825767-10799-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-01-30 05:22:47 +00:00
static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
{
sched/fair: Fix insertion in rq->leaf_cfs_rq_list Sargun reported a crash: "I picked up c40f7d74c741a907cfaeb73a7697081881c497d0 sched/fair: Fix infinite loop in update_blocked_averages() by reverting a9e7f6544b9c and put it on top of 4.19.13. In addition to this, I uninlined list_add_leaf_cfs_rq for debugging. This revealed a new bug that we didn't get to because we kept getting crashes from the previous issue. When we are running with cgroups that are rapidly changing, with CFS bandwidth control, and in addition using the cpusets cgroup, we see this crash. Specifically, it seems to occur with cgroups that are throttled and we change the allowed cpuset." The algorithm used to order cfs_rq in rq->leaf_cfs_rq_list assumes that it will walk down to root the 1st time a cfs_rq is used and we will finish to add either a cfs_rq without parent or a cfs_rq with a parent that is already on the list. But this is not always true in presence of throttling. Because a cfs_rq can be throttled even if it has never been used but other CPUs of the cgroup have already used all the bandwdith, we are not sure to go down to the root and add all cfs_rq in the list. Ensure that all cfs_rq will be added in the list even if they are throttled. [ mingo: Fix !CGROUPS build. ] Reported-by: Sargun Dhillon <sargun@sargun.me> Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: tj@kernel.org Fixes: 9c2791f936ef ("Fix hierarchical order in rq->leaf_cfs_rq_list") Link: https://lkml.kernel.org/r/1548825767-10799-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-01-30 05:22:47 +00:00
return true;
}
static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
{
}
static inline void assert_list_leaf_cfs_rq(struct rq *rq)
{
}
#define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
static inline struct sched_entity *parent_entity(struct sched_entity *se)
{
return NULL;
}
static inline void
find_matching_se(struct sched_entity **se, struct sched_entity **pse)
{
}
#endif /* CONFIG_FAIR_GROUP_SCHED */
static __always_inline
sched/fair: Rework sched_fair time accounting Christian suffers from a bad BIOS that wrecks his i5's TSC sync. This results in him occasionally seeing time going backwards - which crashes the scheduler ... Most of our time accounting can actually handle that except the most common one; the tick time update of sched_fair. There is a further problem with that code; previously we assumed that because we get a tick every TICK_NSEC our time delta could never exceed 32bits and math was simpler. However, ever since Frederic managed to get NO_HZ_FULL merged; this is no longer the case since now a task can run for a long time indeed without getting a tick. It only takes about ~4.2 seconds to overflow our u32 in nanoseconds. This means we not only need to better deal with time going backwards; but also means we need to be able to deal with large deltas. This patch reworks the entire code and uses mul_u64_u32_shr() as proposed by Andy a long while ago. We express our virtual time scale factor in a u32 multiplier and shift right and the 32bit mul_u64_u32_shr() implementation reduces to a single 32x32->64 multiply if the time delta is still short (common case). For 64bit a 64x64->128 multiply can be used if ARCH_SUPPORTS_INT128. Reported-and-Tested-by: Christian Engelmayer <cengelma@gmx.at> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Cc: fweisbec@gmail.com Cc: Paul Turner <pjt@google.com> Cc: Stanislaw Gruszka <sgruszka@redhat.com> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Link: http://lkml.kernel.org/r/20131118172706.GI3866@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-11-18 17:27:06 +00:00
void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
/**************************************************************
* Scheduling class tree data structure manipulation methods:
*/
static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
{
s64 delta = (s64)(vruntime - max_vruntime);
if (delta > 0)
max_vruntime = vruntime;
return max_vruntime;
}
static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
{
s64 delta = (s64)(vruntime - min_vruntime);
if (delta < 0)
min_vruntime = vruntime;
return min_vruntime;
}
static inline int entity_before(struct sched_entity *a,
struct sched_entity *b)
{
return (s64)(a->vruntime - b->vruntime) < 0;
}
static void update_min_vruntime(struct cfs_rq *cfs_rq)
{
struct sched_entity *curr = cfs_rq->curr;
struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline);
u64 vruntime = cfs_rq->min_vruntime;
if (curr) {
if (curr->on_rq)
vruntime = curr->vruntime;
else
curr = NULL;
}
if (leftmost) { /* non-empty tree */
struct sched_entity *se;
se = rb_entry(leftmost, struct sched_entity, run_node);
if (!curr)
vruntime = se->vruntime;
else
vruntime = min_vruntime(vruntime, se->vruntime);
}
/* ensure we never gain time by being placed backwards. */
cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
#ifndef CONFIG_64BIT
smp_wmb();
cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
#endif
}
/*
* Enqueue an entity into the rb-tree:
*/
static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
{
struct rb_node **link = &cfs_rq->tasks_timeline.rb_root.rb_node;
struct rb_node *parent = NULL;
struct sched_entity *entry;
bool leftmost = true;
/*
* Find the right place in the rbtree:
*/
while (*link) {
parent = *link;
entry = rb_entry(parent, struct sched_entity, run_node);
/*
* We dont care about collisions. Nodes with
* the same key stay together.
*/
if (entity_before(se, entry)) {
link = &parent->rb_left;
} else {
link = &parent->rb_right;
leftmost = false;
}
}
rb_link_node(&se->run_node, parent, link);
rb_insert_color_cached(&se->run_node,
&cfs_rq->tasks_timeline, leftmost);
}
static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
{
rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline);
}
struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
{
struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
if (!left)
return NULL;
return rb_entry(left, struct sched_entity, run_node);
}
static struct sched_entity *__pick_next_entity(struct sched_entity *se)
{
struct rb_node *next = rb_next(&se->run_node);
if (!next)
return NULL;
return rb_entry(next, struct sched_entity, run_node);
}
#ifdef CONFIG_SCHED_DEBUG
struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
{
struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
if (!last)
return NULL;
return rb_entry(last, struct sched_entity, run_node);
}
/**************************************************************
* Scheduling class statistics methods:
*/
int sched_proc_update_handler(struct ctl_table *table, int write,
void __user *buffer, size_t *lenp,
loff_t *ppos)
{
int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
unsigned int factor = get_update_sysctl_factor();
if (ret || !write)
return ret;
sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
sysctl_sched_min_granularity);
#define WRT_SYSCTL(name) \
(normalized_sysctl_##name = sysctl_##name / (factor))
WRT_SYSCTL(sched_min_granularity);
WRT_SYSCTL(sched_latency);
WRT_SYSCTL(sched_wakeup_granularity);
#undef WRT_SYSCTL
return 0;
}
#endif
/*
* delta /= w
*/
sched/fair: Rework sched_fair time accounting Christian suffers from a bad BIOS that wrecks his i5's TSC sync. This results in him occasionally seeing time going backwards - which crashes the scheduler ... Most of our time accounting can actually handle that except the most common one; the tick time update of sched_fair. There is a further problem with that code; previously we assumed that because we get a tick every TICK_NSEC our time delta could never exceed 32bits and math was simpler. However, ever since Frederic managed to get NO_HZ_FULL merged; this is no longer the case since now a task can run for a long time indeed without getting a tick. It only takes about ~4.2 seconds to overflow our u32 in nanoseconds. This means we not only need to better deal with time going backwards; but also means we need to be able to deal with large deltas. This patch reworks the entire code and uses mul_u64_u32_shr() as proposed by Andy a long while ago. We express our virtual time scale factor in a u32 multiplier and shift right and the 32bit mul_u64_u32_shr() implementation reduces to a single 32x32->64 multiply if the time delta is still short (common case). For 64bit a 64x64->128 multiply can be used if ARCH_SUPPORTS_INT128. Reported-and-Tested-by: Christian Engelmayer <cengelma@gmx.at> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Cc: fweisbec@gmail.com Cc: Paul Turner <pjt@google.com> Cc: Stanislaw Gruszka <sgruszka@redhat.com> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Link: http://lkml.kernel.org/r/20131118172706.GI3866@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-11-18 17:27:06 +00:00
static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
{
if (unlikely(se->load.weight != NICE_0_LOAD))
sched/fair: Rework sched_fair time accounting Christian suffers from a bad BIOS that wrecks his i5's TSC sync. This results in him occasionally seeing time going backwards - which crashes the scheduler ... Most of our time accounting can actually handle that except the most common one; the tick time update of sched_fair. There is a further problem with that code; previously we assumed that because we get a tick every TICK_NSEC our time delta could never exceed 32bits and math was simpler. However, ever since Frederic managed to get NO_HZ_FULL merged; this is no longer the case since now a task can run for a long time indeed without getting a tick. It only takes about ~4.2 seconds to overflow our u32 in nanoseconds. This means we not only need to better deal with time going backwards; but also means we need to be able to deal with large deltas. This patch reworks the entire code and uses mul_u64_u32_shr() as proposed by Andy a long while ago. We express our virtual time scale factor in a u32 multiplier and shift right and the 32bit mul_u64_u32_shr() implementation reduces to a single 32x32->64 multiply if the time delta is still short (common case). For 64bit a 64x64->128 multiply can be used if ARCH_SUPPORTS_INT128. Reported-and-Tested-by: Christian Engelmayer <cengelma@gmx.at> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Cc: fweisbec@gmail.com Cc: Paul Turner <pjt@google.com> Cc: Stanislaw Gruszka <sgruszka@redhat.com> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Link: http://lkml.kernel.org/r/20131118172706.GI3866@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-11-18 17:27:06 +00:00
delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
return delta;
}
/*
* The idea is to set a period in which each task runs once.
*
* When there are too many tasks (sched_nr_latency) we have to stretch
* this period because otherwise the slices get too small.
*
* p = (nr <= nl) ? l : l*nr/nl
*/
static u64 __sched_period(unsigned long nr_running)
{
if (unlikely(nr_running > sched_nr_latency))
return nr_running * sysctl_sched_min_granularity;
else
return sysctl_sched_latency;
}
/*
* We calculate the wall-time slice from the period by taking a part
* proportional to the weight.
*
* s = p*P[w/rw]
*/
static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
{
u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
for_each_sched_entity(se) {
struct load_weight *load;
struct load_weight lw;
cfs_rq = cfs_rq_of(se);
load = &cfs_rq->load;
if (unlikely(!se->on_rq)) {
lw = cfs_rq->load;
update_load_add(&lw, se->load.weight);
load = &lw;
}
sched/fair: Rework sched_fair time accounting Christian suffers from a bad BIOS that wrecks his i5's TSC sync. This results in him occasionally seeing time going backwards - which crashes the scheduler ... Most of our time accounting can actually handle that except the most common one; the tick time update of sched_fair. There is a further problem with that code; previously we assumed that because we get a tick every TICK_NSEC our time delta could never exceed 32bits and math was simpler. However, ever since Frederic managed to get NO_HZ_FULL merged; this is no longer the case since now a task can run for a long time indeed without getting a tick. It only takes about ~4.2 seconds to overflow our u32 in nanoseconds. This means we not only need to better deal with time going backwards; but also means we need to be able to deal with large deltas. This patch reworks the entire code and uses mul_u64_u32_shr() as proposed by Andy a long while ago. We express our virtual time scale factor in a u32 multiplier and shift right and the 32bit mul_u64_u32_shr() implementation reduces to a single 32x32->64 multiply if the time delta is still short (common case). For 64bit a 64x64->128 multiply can be used if ARCH_SUPPORTS_INT128. Reported-and-Tested-by: Christian Engelmayer <cengelma@gmx.at> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Cc: fweisbec@gmail.com Cc: Paul Turner <pjt@google.com> Cc: Stanislaw Gruszka <sgruszka@redhat.com> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Link: http://lkml.kernel.org/r/20131118172706.GI3866@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-11-18 17:27:06 +00:00
slice = __calc_delta(slice, se->load.weight, load);
}
return slice;
}
/*
* We calculate the vruntime slice of a to-be-inserted task.
*
* vs = s/w
*/
static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
{
return calc_delta_fair(sched_slice(cfs_rq, se), se);
}
#include "pelt.h"
sched/fair: Update scale invariance of PELT The current implementation of load tracking invariance scales the contribution with current frequency and uarch performance (only for utilization) of the CPU. One main result of this formula is that the figures are capped by current capacity of CPU. Another one is that the load_avg is not invariant because not scaled with uarch. The util_avg of a periodic task that runs r time slots every p time slots varies in the range : U * (1-y^r)/(1-y^p) * y^i < Utilization < U * (1-y^r)/(1-y^p) with U is the max util_avg value = SCHED_CAPACITY_SCALE At a lower capacity, the range becomes: U * C * (1-y^r')/(1-y^p) * y^i' < Utilization < U * C * (1-y^r')/(1-y^p) with C reflecting the compute capacity ratio between current capacity and max capacity. so C tries to compensate changes in (1-y^r') but it can't be accurate. Instead of scaling the contribution value of PELT algo, we should scale the running time. The PELT signal aims to track the amount of computation of tasks and/or rq so it seems more correct to scale the running time to reflect the effective amount of computation done since the last update. In order to be fully invariant, we need to apply the same amount of running time and idle time whatever the current capacity. Because running at lower capacity implies that the task will run longer, we have to ensure that the same amount of idle time will be applied when system becomes idle and no idle time has been "stolen". But reaching the maximum utilization value (SCHED_CAPACITY_SCALE) means that the task is seen as an always-running task whatever the capacity of the CPU (even at max compute capacity). In this case, we can discard this "stolen" idle times which becomes meaningless. In order to achieve this time scaling, a new clock_pelt is created per rq. The increase of this clock scales with current capacity when something is running on rq and synchronizes with clock_task when rq is idle. With this mechanism, we ensure the same running and idle time whatever the current capacity. This also enables to simplify the pelt algorithm by removing all references of uarch and frequency and applying the same contribution to utilization and loads. Furthermore, the scaling is done only once per update of clock (update_rq_clock_task()) instead of during each update of sched_entities and cfs/rt/dl_rq of the rq like the current implementation. This is interesting when cgroup are involved as shown in the results below: On a hikey (octo Arm64 platform). Performance cpufreq governor and only shallowest c-state to remove variance generated by those power features so we only track the impact of pelt algo. each test runs 16 times: ./perf bench sched pipe (higher is better) kernel tip/sched/core + patch ops/seconds ops/seconds diff cgroup root 59652(+/- 0.18%) 59876(+/- 0.24%) +0.38% level1 55608(+/- 0.27%) 55923(+/- 0.24%) +0.57% level2 52115(+/- 0.29%) 52564(+/- 0.22%) +0.86% hackbench -l 1000 (lower is better) kernel tip/sched/core + patch duration(sec) duration(sec) diff cgroup root 4.453(+/- 2.37%) 4.383(+/- 2.88%) -1.57% level1 4.859(+/- 8.50%) 4.830(+/- 7.07%) -0.60% level2 5.063(+/- 9.83%) 4.928(+/- 9.66%) -2.66% Then, the responsiveness of PELT is improved when CPU is not running at max capacity with this new algorithm. I have put below some examples of duration to reach some typical load values according to the capacity of the CPU with current implementation and with this patch. These values has been computed based on the geometric series and the half period value: Util (%) max capacity half capacity(mainline) half capacity(w/ patch) 972 (95%) 138ms not reachable 276ms 486 (47.5%) 30ms 138ms 60ms 256 (25%) 13ms 32ms 26ms On my hikey (octo Arm64 platform) with schedutil governor, the time to reach max OPP when starting from a null utilization, decreases from 223ms with current scale invariance down to 121ms with the new algorithm. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: patrick.bellasi@arm.com Cc: pjt@google.com Cc: pkondeti@codeaurora.org Cc: quentin.perret@arm.com Cc: rjw@rjwysocki.net Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Link: https://lkml.kernel.org/r/1548257214-13745-3-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-01-23 15:26:53 +00:00
#ifdef CONFIG_SMP
static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
static unsigned long task_h_load(struct task_struct *p);
sched/fair: Add 'group_misfit_task' load-balance type To maximize throughput in systems with asymmetric CPU capacities (e.g. ARM big.LITTLE) load-balancing has to consider task and CPU utilization as well as per-CPU compute capacity when load-balancing in addition to the current average load based load-balancing policy. Tasks with high utilization that are scheduled on a lower capacity CPU need to be identified and migrated to a higher capacity CPU if possible to maximize throughput. To implement this additional policy an additional group_type (load-balance scenario) is added: 'group_misfit_task'. This represents scenarios where a sched_group has one or more tasks that are not suitable for its per-CPU capacity. 'group_misfit_task' is only considered if the system is not overloaded or imbalanced ('group_imbalanced' or 'group_overloaded'). Identifying misfit tasks requires the rq lock to be held. To avoid taking remote rq locks to examine source sched_groups for misfit tasks, each CPU is responsible for tracking misfit tasks themselves and update the rq->misfit_task flag. This means checking task utilization when tasks are scheduled and on sched_tick. Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: gaku.inami.xh@renesas.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1530699470-29808-3-git-send-email-morten.rasmussen@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-07-04 10:17:40 +00:00
static unsigned long capacity_of(int cpu);
/* Give new sched_entity start runnable values to heavy its load in infant time */
void init_entity_runnable_average(struct sched_entity *se)
sched: Set an initial value of runnable avg for new forked task We need to initialize the se.avg.{decay_count, load_avg_contrib} for a new forked task. Otherwise random values of above variables cause a mess when a new task is enqueued: enqueue_task_fair enqueue_entity enqueue_entity_load_avg and make fork balancing imbalance due to incorrect load_avg_contrib. Further more, Morten Rasmussen notice some tasks were not launched at once after created. So Paul and Peter suggest giving a start value for new task runnable avg time same as sched_slice(). PeterZ said: > So the 'problem' is that our running avg is a 'floating' average; ie. it > decays with time. Now we have to guess about the future of our newly > spawned task -- something that is nigh impossible seeing these CPU > vendors keep refusing to implement the crystal ball instruction. > > So there's two asymptotic cases we want to deal well with; 1) the case > where the newly spawned program will be 'nearly' idle for its lifetime; > and 2) the case where its cpu-bound. > > Since we have to guess, we'll go for worst case and assume its > cpu-bound; now we don't want to make the avg so heavy adjusting to the > near-idle case takes forever. We want to be able to quickly adjust and > lower our running avg. > > Now we also don't want to make our avg too light, such that it gets > decremented just for the new task not having had a chance to run yet -- > even if when it would run, it would be more cpu-bound than not. > > So what we do is we make the initial avg of the same duration as that we > guess it takes to run each task on the system at least once -- aka > sched_slice(). > > Of course we can defeat this with wakeup/fork bombs, but in the 'normal' > case it should be good enough. Paul also contributed most of the code comments in this commit. Signed-off-by: Alex Shi <alex.shi@intel.com> Reviewed-by: Gu Zheng <guz.fnst@cn.fujitsu.com> Reviewed-by: Paul Turner <pjt@google.com> [peterz; added explanation of sched_slice() usage] Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1371694737-29336-4-git-send-email-alex.shi@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-06-20 02:18:47 +00:00
{
struct sched_avg *sa = &se->avg;
sched: Set an initial value of runnable avg for new forked task We need to initialize the se.avg.{decay_count, load_avg_contrib} for a new forked task. Otherwise random values of above variables cause a mess when a new task is enqueued: enqueue_task_fair enqueue_entity enqueue_entity_load_avg and make fork balancing imbalance due to incorrect load_avg_contrib. Further more, Morten Rasmussen notice some tasks were not launched at once after created. So Paul and Peter suggest giving a start value for new task runnable avg time same as sched_slice(). PeterZ said: > So the 'problem' is that our running avg is a 'floating' average; ie. it > decays with time. Now we have to guess about the future of our newly > spawned task -- something that is nigh impossible seeing these CPU > vendors keep refusing to implement the crystal ball instruction. > > So there's two asymptotic cases we want to deal well with; 1) the case > where the newly spawned program will be 'nearly' idle for its lifetime; > and 2) the case where its cpu-bound. > > Since we have to guess, we'll go for worst case and assume its > cpu-bound; now we don't want to make the avg so heavy adjusting to the > near-idle case takes forever. We want to be able to quickly adjust and > lower our running avg. > > Now we also don't want to make our avg too light, such that it gets > decremented just for the new task not having had a chance to run yet -- > even if when it would run, it would be more cpu-bound than not. > > So what we do is we make the initial avg of the same duration as that we > guess it takes to run each task on the system at least once -- aka > sched_slice(). > > Of course we can defeat this with wakeup/fork bombs, but in the 'normal' > case it should be good enough. Paul also contributed most of the code comments in this commit. Signed-off-by: Alex Shi <alex.shi@intel.com> Reviewed-by: Gu Zheng <guz.fnst@cn.fujitsu.com> Reviewed-by: Paul Turner <pjt@google.com> [peterz; added explanation of sched_slice() usage] Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1371694737-29336-4-git-send-email-alex.shi@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-06-20 02:18:47 +00:00
memset(sa, 0, sizeof(*sa));
sched/fair: Fix incorrect task group ->load_avg A scheduler performance regression has been reported by Joseph Salisbury, which he bisected back to: 3d30544f0212 ("sched/fair: Apply more PELT fixes) The regression triggers when several levels of task groups are involved (read: SystemD) and cpu_possible_mask != cpu_present_mask. The root cause is that group entity's load (tg_child->se[i]->avg.load_avg) is initialized to scale_load_down(se->load.weight). During the creation of a child task group, its group entities on possible CPUs are attached to parent's cfs_rq (tg_parent) and their loads are added to the parent's load (tg_parent->load_avg) with update_tg_load_avg(). But only the load on online CPUs will then be updated to reflect real load, whereas load on other CPUs will stay at the initial value. The result is a tg_parent->load_avg that is higher than the real load, the weight of group entities (tg_parent->se[i]->load.weight) on online CPUs is smaller than it should be, and the task group gets a less running time than what it could expect. ( This situation can be detected with /proc/sched_debug. The ".tg_load_avg" of the task group will be much higher than sum of ".tg_load_avg_contrib" of online cfs_rqs of the task group. ) The load of group entities don't have to be intialized to something else than 0 because their load will increase when an entity is attached. Reported-by: Joseph Salisbury <joseph.salisbury@canonical.com> Tested-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: <stable@vger.kernel.org> # 4.8.x Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: joonwoop@codeaurora.org Fixes: 3d30544f0212 ("sched/fair: Apply more PELT fixes) Link: http://lkml.kernel.org/r/1476881123-10159-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-10-19 12:45:23 +00:00
/*
* Tasks are initialized with full load to be seen as heavy tasks until
sched/fair: Fix incorrect task group ->load_avg A scheduler performance regression has been reported by Joseph Salisbury, which he bisected back to: 3d30544f0212 ("sched/fair: Apply more PELT fixes) The regression triggers when several levels of task groups are involved (read: SystemD) and cpu_possible_mask != cpu_present_mask. The root cause is that group entity's load (tg_child->se[i]->avg.load_avg) is initialized to scale_load_down(se->load.weight). During the creation of a child task group, its group entities on possible CPUs are attached to parent's cfs_rq (tg_parent) and their loads are added to the parent's load (tg_parent->load_avg) with update_tg_load_avg(). But only the load on online CPUs will then be updated to reflect real load, whereas load on other CPUs will stay at the initial value. The result is a tg_parent->load_avg that is higher than the real load, the weight of group entities (tg_parent->se[i]->load.weight) on online CPUs is smaller than it should be, and the task group gets a less running time than what it could expect. ( This situation can be detected with /proc/sched_debug. The ".tg_load_avg" of the task group will be much higher than sum of ".tg_load_avg_contrib" of online cfs_rqs of the task group. ) The load of group entities don't have to be intialized to something else than 0 because their load will increase when an entity is attached. Reported-by: Joseph Salisbury <joseph.salisbury@canonical.com> Tested-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: <stable@vger.kernel.org> # 4.8.x Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: joonwoop@codeaurora.org Fixes: 3d30544f0212 ("sched/fair: Apply more PELT fixes) Link: http://lkml.kernel.org/r/1476881123-10159-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-10-19 12:45:23 +00:00
* they get a chance to stabilize to their real load level.
* Group entities are initialized with zero load to reflect the fact that
sched/fair: Fix incorrect task group ->load_avg A scheduler performance regression has been reported by Joseph Salisbury, which he bisected back to: 3d30544f0212 ("sched/fair: Apply more PELT fixes) The regression triggers when several levels of task groups are involved (read: SystemD) and cpu_possible_mask != cpu_present_mask. The root cause is that group entity's load (tg_child->se[i]->avg.load_avg) is initialized to scale_load_down(se->load.weight). During the creation of a child task group, its group entities on possible CPUs are attached to parent's cfs_rq (tg_parent) and their loads are added to the parent's load (tg_parent->load_avg) with update_tg_load_avg(). But only the load on online CPUs will then be updated to reflect real load, whereas load on other CPUs will stay at the initial value. The result is a tg_parent->load_avg that is higher than the real load, the weight of group entities (tg_parent->se[i]->load.weight) on online CPUs is smaller than it should be, and the task group gets a less running time than what it could expect. ( This situation can be detected with /proc/sched_debug. The ".tg_load_avg" of the task group will be much higher than sum of ".tg_load_avg_contrib" of online cfs_rqs of the task group. ) The load of group entities don't have to be intialized to something else than 0 because their load will increase when an entity is attached. Reported-by: Joseph Salisbury <joseph.salisbury@canonical.com> Tested-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: <stable@vger.kernel.org> # 4.8.x Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: joonwoop@codeaurora.org Fixes: 3d30544f0212 ("sched/fair: Apply more PELT fixes) Link: http://lkml.kernel.org/r/1476881123-10159-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-10-19 12:45:23 +00:00
* nothing has been attached to the task group yet.
*/
if (entity_is_task(se))
sa->load_avg = scale_load_down(se->load.weight);
sched/fair: Rewrite runnable load and utilization average tracking The idea of runnable load average (let runnable time contribute to weight) was proposed by Paul Turner and Ben Segall, and it is still followed by this rewrite. This rewrite aims to solve the following issues: 1. cfs_rq's load average (namely runnable_load_avg and blocked_load_avg) is updated at the granularity of an entity at a time, which results in the cfs_rq's load average is stale or partially updated: at any time, only one entity is up to date, all other entities are effectively lagging behind. This is undesirable. To illustrate, if we have n runnable entities in the cfs_rq, as time elapses, they certainly become outdated: t0: cfs_rq { e1_old, e2_old, ..., en_old } and when we update: t1: update e1, then we have cfs_rq { e1_new, e2_old, ..., en_old } t2: update e2, then we have cfs_rq { e1_old, e2_new, ..., en_old } ... We solve this by combining all runnable entities' load averages together in cfs_rq's avg, and update the cfs_rq's avg as a whole. This is based on the fact that if we regard the update as a function, then: w * update(e) = update(w * e) and update(e1) + update(e2) = update(e1 + e2), then w1 * update(e1) + w2 * update(e2) = update(w1 * e1 + w2 * e2) therefore, by this rewrite, we have an entirely updated cfs_rq at the time we update it: t1: update cfs_rq { e1_new, e2_new, ..., en_new } t2: update cfs_rq { e1_new, e2_new, ..., en_new } ... 2. cfs_rq's load average is different between top rq->cfs_rq and other task_group's per CPU cfs_rqs in whether or not blocked_load_average contributes to the load. The basic idea behind runnable load average (the same for utilization) is that the blocked state is taken into account as opposed to only accounting for the currently runnable state. Therefore, the average should include both the runnable/running and blocked load averages. This rewrite does that. In addition, we also combine runnable/running and blocked averages of all entities into the cfs_rq's average, and update it together at once. This is based on the fact that: update(runnable) + update(blocked) = update(runnable + blocked) This significantly reduces the code as we don't need to separately maintain/update runnable/running load and blocked load. 3. How task_group entities' share is calculated is complex and imprecise. We reduce the complexity in this rewrite to allow a very simple rule: the task_group's load_avg is aggregated from its per CPU cfs_rqs's load_avgs. Then group entity's weight is simply proportional to its own cfs_rq's load_avg / task_group's load_avg. To illustrate, if a task_group has { cfs_rq1, cfs_rq2, ..., cfs_rqn }, then, task_group_avg = cfs_rq1_avg + cfs_rq2_avg + ... + cfs_rqn_avg, then cfs_rqx's entity's share = cfs_rqx_avg / task_group_avg * task_group's share To sum up, this rewrite in principle is equivalent to the current one, but fixes the issues described above. Turns out, it significantly reduces the code complexity and hence increases clarity and efficiency. In addition, the new averages are more smooth/continuous (no spurious spikes and valleys) and updated more consistently and quickly to reflect the load dynamics. As a result, we have less load tracking overhead, better performance, and especially better power efficiency due to more balanced load. Signed-off-by: Yuyang Du <yuyang.du@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: arjan@linux.intel.com Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: fengguang.wu@intel.com Cc: len.brown@intel.com Cc: morten.rasmussen@arm.com Cc: pjt@google.com Cc: rafael.j.wysocki@intel.com Cc: umgwanakikbuti@gmail.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1436918682-4971-3-git-send-email-yuyang.du@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-07-15 00:04:37 +00:00
/* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
sched: Set an initial value of runnable avg for new forked task We need to initialize the se.avg.{decay_count, load_avg_contrib} for a new forked task. Otherwise random values of above variables cause a mess when a new task is enqueued: enqueue_task_fair enqueue_entity enqueue_entity_load_avg and make fork balancing imbalance due to incorrect load_avg_contrib. Further more, Morten Rasmussen notice some tasks were not launched at once after created. So Paul and Peter suggest giving a start value for new task runnable avg time same as sched_slice(). PeterZ said: > So the 'problem' is that our running avg is a 'floating' average; ie. it > decays with time. Now we have to guess about the future of our newly > spawned task -- something that is nigh impossible seeing these CPU > vendors keep refusing to implement the crystal ball instruction. > > So there's two asymptotic cases we want to deal well with; 1) the case > where the newly spawned program will be 'nearly' idle for its lifetime; > and 2) the case where its cpu-bound. > > Since we have to guess, we'll go for worst case and assume its > cpu-bound; now we don't want to make the avg so heavy adjusting to the > near-idle case takes forever. We want to be able to quickly adjust and > lower our running avg. > > Now we also don't want to make our avg too light, such that it gets > decremented just for the new task not having had a chance to run yet -- > even if when it would run, it would be more cpu-bound than not. > > So what we do is we make the initial avg of the same duration as that we > guess it takes to run each task on the system at least once -- aka > sched_slice(). > > Of course we can defeat this with wakeup/fork bombs, but in the 'normal' > case it should be good enough. Paul also contributed most of the code comments in this commit. Signed-off-by: Alex Shi <alex.shi@intel.com> Reviewed-by: Gu Zheng <guz.fnst@cn.fujitsu.com> Reviewed-by: Paul Turner <pjt@google.com> [peterz; added explanation of sched_slice() usage] Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1371694737-29336-4-git-send-email-alex.shi@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-06-20 02:18:47 +00:00
}
static void attach_entity_cfs_rq(struct sched_entity *se);
sched/fair: Fix PELT integrity for new tasks Vincent and Yuyang found another few scenarios in which entity tracking goes wobbly. The scenarios are basically due to the fact that new tasks are not immediately attached and thereby differ from the normal situation -- a task is always attached to a cfs_rq load average (such that it includes its blocked contribution) and are explicitly detached/attached on migration to another cfs_rq. Scenario 1: switch to fair class p->sched_class = fair_class; if (queued) enqueue_task(p); ... enqueue_entity() enqueue_entity_load_avg() migrated = !sa->last_update_time (true) if (migrated) attach_entity_load_avg() check_class_changed() switched_from() (!fair) switched_to() (fair) switched_to_fair() attach_entity_load_avg() If @p is a new task that hasn't been fair before, it will have !last_update_time and, per the above, end up in attach_entity_load_avg() _twice_. Scenario 2: change between cgroups sched_move_group(p) if (queued) dequeue_task() task_move_group_fair() detach_task_cfs_rq() detach_entity_load_avg() set_task_rq() attach_task_cfs_rq() attach_entity_load_avg() if (queued) enqueue_task(); ... enqueue_entity() enqueue_entity_load_avg() migrated = !sa->last_update_time (true) if (migrated) attach_entity_load_avg() Similar as with scenario 1, if @p is a new task, it will have !load_update_time and we'll end up in attach_entity_load_avg() _twice_. Furthermore, notice how we do a detach_entity_load_avg() on something that wasn't attached to begin with. As stated above; the problem is that the new task isn't yet attached to the load tracking and thereby violates the invariant assumption. This patch remedies this by ensuring a new task is indeed properly attached to the load tracking on creation, through post_init_entity_util_avg(). Of course, this isn't entirely as straightforward as one might think, since the task is hashed before we call wake_up_new_task() and thus can be poked at. We avoid this by adding TASK_NEW and teaching cpu_cgroup_can_attach() to refuse such tasks. Reported-by: Yuyang Du <yuyang.du@intel.com> Reported-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-06-16 11:29:28 +00:00
sched/fair: Initiate a new task's util avg to a bounded value A new task's util_avg is set to full utilization of a CPU (100% time running). This accelerates a new task's utilization ramp-up, useful to boost its execution in early time. However, it may result in (insanely) high utilization for a transient time period when a flood of tasks are spawned. Importantly, it violates the "fundamentally bounded" CPU utilization, and its side effect is negative if we don't take any measure to bound it. This patch proposes an algorithm to address this issue. It has two methods to approach a sensible initial util_avg: (1) An expected (or average) util_avg based on its cfs_rq's util_avg: util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight (2) A trajectory of how successive new tasks' util develops, which gives 1/2 of the left utilization budget to a new task such that the additional util is noticeably large (when overall util is low) or unnoticeably small (when overall util is high enough). In the meantime, the aggregate utilization is well bounded: util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n where n denotes the nth task. If util_avg is larger than util_avg_cap, then the effective util is clamped to the util_avg_cap. Reported-by: Andrey Ryabinin <aryabinin@virtuozzo.com> Signed-off-by: Yuyang Du <yuyang.du@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: bsegall@google.com Cc: morten.rasmussen@arm.com Cc: pjt@google.com Cc: steve.muckle@linaro.org Link: http://lkml.kernel.org/r/1459283456-21682-1-git-send-email-yuyang.du@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-03-29 20:30:56 +00:00
/*
* With new tasks being created, their initial util_avgs are extrapolated
* based on the cfs_rq's current util_avg:
*
* util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
*
* However, in many cases, the above util_avg does not give a desired
* value. Moreover, the sum of the util_avgs may be divergent, such
* as when the series is a harmonic series.
*
* To solve this problem, we also cap the util_avg of successive tasks to
* only 1/2 of the left utilization budget:
*
* util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
sched/fair: Initiate a new task's util avg to a bounded value A new task's util_avg is set to full utilization of a CPU (100% time running). This accelerates a new task's utilization ramp-up, useful to boost its execution in early time. However, it may result in (insanely) high utilization for a transient time period when a flood of tasks are spawned. Importantly, it violates the "fundamentally bounded" CPU utilization, and its side effect is negative if we don't take any measure to bound it. This patch proposes an algorithm to address this issue. It has two methods to approach a sensible initial util_avg: (1) An expected (or average) util_avg based on its cfs_rq's util_avg: util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight (2) A trajectory of how successive new tasks' util develops, which gives 1/2 of the left utilization budget to a new task such that the additional util is noticeably large (when overall util is low) or unnoticeably small (when overall util is high enough). In the meantime, the aggregate utilization is well bounded: util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n where n denotes the nth task. If util_avg is larger than util_avg_cap, then the effective util is clamped to the util_avg_cap. Reported-by: Andrey Ryabinin <aryabinin@virtuozzo.com> Signed-off-by: Yuyang Du <yuyang.du@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: bsegall@google.com Cc: morten.rasmussen@arm.com Cc: pjt@google.com Cc: steve.muckle@linaro.org Link: http://lkml.kernel.org/r/1459283456-21682-1-git-send-email-yuyang.du@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-03-29 20:30:56 +00:00
*
* where n denotes the nth task and cpu_scale the CPU capacity.
sched/fair: Initiate a new task's util avg to a bounded value A new task's util_avg is set to full utilization of a CPU (100% time running). This accelerates a new task's utilization ramp-up, useful to boost its execution in early time. However, it may result in (insanely) high utilization for a transient time period when a flood of tasks are spawned. Importantly, it violates the "fundamentally bounded" CPU utilization, and its side effect is negative if we don't take any measure to bound it. This patch proposes an algorithm to address this issue. It has two methods to approach a sensible initial util_avg: (1) An expected (or average) util_avg based on its cfs_rq's util_avg: util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight (2) A trajectory of how successive new tasks' util develops, which gives 1/2 of the left utilization budget to a new task such that the additional util is noticeably large (when overall util is low) or unnoticeably small (when overall util is high enough). In the meantime, the aggregate utilization is well bounded: util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n where n denotes the nth task. If util_avg is larger than util_avg_cap, then the effective util is clamped to the util_avg_cap. Reported-by: Andrey Ryabinin <aryabinin@virtuozzo.com> Signed-off-by: Yuyang Du <yuyang.du@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: bsegall@google.com Cc: morten.rasmussen@arm.com Cc: pjt@google.com Cc: steve.muckle@linaro.org Link: http://lkml.kernel.org/r/1459283456-21682-1-git-send-email-yuyang.du@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-03-29 20:30:56 +00:00
*
* For example, for a CPU with 1024 of capacity, a simplest series from
* the beginning would be like:
sched/fair: Initiate a new task's util avg to a bounded value A new task's util_avg is set to full utilization of a CPU (100% time running). This accelerates a new task's utilization ramp-up, useful to boost its execution in early time. However, it may result in (insanely) high utilization for a transient time period when a flood of tasks are spawned. Importantly, it violates the "fundamentally bounded" CPU utilization, and its side effect is negative if we don't take any measure to bound it. This patch proposes an algorithm to address this issue. It has two methods to approach a sensible initial util_avg: (1) An expected (or average) util_avg based on its cfs_rq's util_avg: util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight (2) A trajectory of how successive new tasks' util develops, which gives 1/2 of the left utilization budget to a new task such that the additional util is noticeably large (when overall util is low) or unnoticeably small (when overall util is high enough). In the meantime, the aggregate utilization is well bounded: util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n where n denotes the nth task. If util_avg is larger than util_avg_cap, then the effective util is clamped to the util_avg_cap. Reported-by: Andrey Ryabinin <aryabinin@virtuozzo.com> Signed-off-by: Yuyang Du <yuyang.du@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: bsegall@google.com Cc: morten.rasmussen@arm.com Cc: pjt@google.com Cc: steve.muckle@linaro.org Link: http://lkml.kernel.org/r/1459283456-21682-1-git-send-email-yuyang.du@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-03-29 20:30:56 +00:00
*
* task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
* cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
*
* Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
* if util_avg > util_avg_cap.
*/
void post_init_entity_util_avg(struct task_struct *p)
sched/fair: Initiate a new task's util avg to a bounded value A new task's util_avg is set to full utilization of a CPU (100% time running). This accelerates a new task's utilization ramp-up, useful to boost its execution in early time. However, it may result in (insanely) high utilization for a transient time period when a flood of tasks are spawned. Importantly, it violates the "fundamentally bounded" CPU utilization, and its side effect is negative if we don't take any measure to bound it. This patch proposes an algorithm to address this issue. It has two methods to approach a sensible initial util_avg: (1) An expected (or average) util_avg based on its cfs_rq's util_avg: util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight (2) A trajectory of how successive new tasks' util develops, which gives 1/2 of the left utilization budget to a new task such that the additional util is noticeably large (when overall util is low) or unnoticeably small (when overall util is high enough). In the meantime, the aggregate utilization is well bounded: util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n where n denotes the nth task. If util_avg is larger than util_avg_cap, then the effective util is clamped to the util_avg_cap. Reported-by: Andrey Ryabinin <aryabinin@virtuozzo.com> Signed-off-by: Yuyang Du <yuyang.du@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: bsegall@google.com Cc: morten.rasmussen@arm.com Cc: pjt@google.com Cc: steve.muckle@linaro.org Link: http://lkml.kernel.org/r/1459283456-21682-1-git-send-email-yuyang.du@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-03-29 20:30:56 +00:00
{
struct sched_entity *se = &p->se;
sched/fair: Initiate a new task's util avg to a bounded value A new task's util_avg is set to full utilization of a CPU (100% time running). This accelerates a new task's utilization ramp-up, useful to boost its execution in early time. However, it may result in (insanely) high utilization for a transient time period when a flood of tasks are spawned. Importantly, it violates the "fundamentally bounded" CPU utilization, and its side effect is negative if we don't take any measure to bound it. This patch proposes an algorithm to address this issue. It has two methods to approach a sensible initial util_avg: (1) An expected (or average) util_avg based on its cfs_rq's util_avg: util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight (2) A trajectory of how successive new tasks' util develops, which gives 1/2 of the left utilization budget to a new task such that the additional util is noticeably large (when overall util is low) or unnoticeably small (when overall util is high enough). In the meantime, the aggregate utilization is well bounded: util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n where n denotes the nth task. If util_avg is larger than util_avg_cap, then the effective util is clamped to the util_avg_cap. Reported-by: Andrey Ryabinin <aryabinin@virtuozzo.com> Signed-off-by: Yuyang Du <yuyang.du@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: bsegall@google.com Cc: morten.rasmussen@arm.com Cc: pjt@google.com Cc: steve.muckle@linaro.org Link: http://lkml.kernel.org/r/1459283456-21682-1-git-send-email-yuyang.du@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-03-29 20:30:56 +00:00
struct cfs_rq *cfs_rq = cfs_rq_of(se);
struct sched_avg *sa = &se->avg;
long cpu_scale = arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq)));
long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2;
sched/fair: Initiate a new task's util avg to a bounded value A new task's util_avg is set to full utilization of a CPU (100% time running). This accelerates a new task's utilization ramp-up, useful to boost its execution in early time. However, it may result in (insanely) high utilization for a transient time period when a flood of tasks are spawned. Importantly, it violates the "fundamentally bounded" CPU utilization, and its side effect is negative if we don't take any measure to bound it. This patch proposes an algorithm to address this issue. It has two methods to approach a sensible initial util_avg: (1) An expected (or average) util_avg based on its cfs_rq's util_avg: util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight (2) A trajectory of how successive new tasks' util develops, which gives 1/2 of the left utilization budget to a new task such that the additional util is noticeably large (when overall util is low) or unnoticeably small (when overall util is high enough). In the meantime, the aggregate utilization is well bounded: util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n where n denotes the nth task. If util_avg is larger than util_avg_cap, then the effective util is clamped to the util_avg_cap. Reported-by: Andrey Ryabinin <aryabinin@virtuozzo.com> Signed-off-by: Yuyang Du <yuyang.du@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: bsegall@google.com Cc: morten.rasmussen@arm.com Cc: pjt@google.com Cc: steve.muckle@linaro.org Link: http://lkml.kernel.org/r/1459283456-21682-1-git-send-email-yuyang.du@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-03-29 20:30:56 +00:00
if (cap > 0) {
if (cfs_rq->avg.util_avg != 0) {
sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
sa->util_avg /= (cfs_rq->avg.load_avg + 1);
if (sa->util_avg > cap)
sa->util_avg = cap;
} else {
sa->util_avg = cap;
}
}
sched/fair: Fix PELT integrity for new tasks Vincent and Yuyang found another few scenarios in which entity tracking goes wobbly. The scenarios are basically due to the fact that new tasks are not immediately attached and thereby differ from the normal situation -- a task is always attached to a cfs_rq load average (such that it includes its blocked contribution) and are explicitly detached/attached on migration to another cfs_rq. Scenario 1: switch to fair class p->sched_class = fair_class; if (queued) enqueue_task(p); ... enqueue_entity() enqueue_entity_load_avg() migrated = !sa->last_update_time (true) if (migrated) attach_entity_load_avg() check_class_changed() switched_from() (!fair) switched_to() (fair) switched_to_fair() attach_entity_load_avg() If @p is a new task that hasn't been fair before, it will have !last_update_time and, per the above, end up in attach_entity_load_avg() _twice_. Scenario 2: change between cgroups sched_move_group(p) if (queued) dequeue_task() task_move_group_fair() detach_task_cfs_rq() detach_entity_load_avg() set_task_rq() attach_task_cfs_rq() attach_entity_load_avg() if (queued) enqueue_task(); ... enqueue_entity() enqueue_entity_load_avg() migrated = !sa->last_update_time (true) if (migrated) attach_entity_load_avg() Similar as with scenario 1, if @p is a new task, it will have !load_update_time and we'll end up in attach_entity_load_avg() _twice_. Furthermore, notice how we do a detach_entity_load_avg() on something that wasn't attached to begin with. As stated above; the problem is that the new task isn't yet attached to the load tracking and thereby violates the invariant assumption. This patch remedies this by ensuring a new task is indeed properly attached to the load tracking on creation, through post_init_entity_util_avg(). Of course, this isn't entirely as straightforward as one might think, since the task is hashed before we call wake_up_new_task() and thus can be poked at. We avoid this by adding TASK_NEW and teaching cpu_cgroup_can_attach() to refuse such tasks. Reported-by: Yuyang Du <yuyang.du@intel.com> Reported-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-06-16 11:29:28 +00:00
sched/pelt: Add a new runnable average signal Now that runnable_load_avg has been removed, we can replace it by a new signal that will highlight the runnable pressure on a cfs_rq. This signal track the waiting time of tasks on rq and can help to better define the state of rqs. At now, only util_avg is used to define the state of a rq: A rq with more that around 80% of utilization and more than 1 tasks is considered as overloaded. But the util_avg signal of a rq can become temporaly low after that a task migrated onto another rq which can bias the classification of the rq. When tasks compete for the same rq, their runnable average signal will be higher than util_avg as it will include the waiting time and we can use this signal to better classify cfs_rqs. The new runnable_avg will track the runnable time of a task which simply adds the waiting time to the running time. The runnable _avg of cfs_rq will be the /Sum of se's runnable_avg and the runnable_avg of group entity will follow the one of the rq similarly to util_avg. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: "Dietmar Eggemann <dietmar.eggemann@arm.com>" Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Valentin Schneider <valentin.schneider@arm.com> Cc: Phil Auld <pauld@redhat.com> Cc: Hillf Danton <hdanton@sina.com> Link: https://lore.kernel.org/r/20200224095223.13361-9-mgorman@techsingularity.net
2020-02-24 09:52:18 +00:00
sa->runnable_avg = cpu_scale;
if (p->sched_class != &fair_sched_class) {
/*
* For !fair tasks do:
*
update_cfs_rq_load_avg(now, cfs_rq);
attach_entity_load_avg(cfs_rq, se);
switched_from_fair(rq, p);
*
* such that the next switched_to_fair() has the
* expected state.
*/
se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq);
return;
sched/fair: Fix PELT integrity for new tasks Vincent and Yuyang found another few scenarios in which entity tracking goes wobbly. The scenarios are basically due to the fact that new tasks are not immediately attached and thereby differ from the normal situation -- a task is always attached to a cfs_rq load average (such that it includes its blocked contribution) and are explicitly detached/attached on migration to another cfs_rq. Scenario 1: switch to fair class p->sched_class = fair_class; if (queued) enqueue_task(p); ... enqueue_entity() enqueue_entity_load_avg() migrated = !sa->last_update_time (true) if (migrated) attach_entity_load_avg() check_class_changed() switched_from() (!fair) switched_to() (fair) switched_to_fair() attach_entity_load_avg() If @p is a new task that hasn't been fair before, it will have !last_update_time and, per the above, end up in attach_entity_load_avg() _twice_. Scenario 2: change between cgroups sched_move_group(p) if (queued) dequeue_task() task_move_group_fair() detach_task_cfs_rq() detach_entity_load_avg() set_task_rq() attach_task_cfs_rq() attach_entity_load_avg() if (queued) enqueue_task(); ... enqueue_entity() enqueue_entity_load_avg() migrated = !sa->last_update_time (true) if (migrated) attach_entity_load_avg() Similar as with scenario 1, if @p is a new task, it will have !load_update_time and we'll end up in attach_entity_load_avg() _twice_. Furthermore, notice how we do a detach_entity_load_avg() on something that wasn't attached to begin with. As stated above; the problem is that the new task isn't yet attached to the load tracking and thereby violates the invariant assumption. This patch remedies this by ensuring a new task is indeed properly attached to the load tracking on creation, through post_init_entity_util_avg(). Of course, this isn't entirely as straightforward as one might think, since the task is hashed before we call wake_up_new_task() and thus can be poked at. We avoid this by adding TASK_NEW and teaching cpu_cgroup_can_attach() to refuse such tasks. Reported-by: Yuyang Du <yuyang.du@intel.com> Reported-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-06-16 11:29:28 +00:00
}
attach_entity_cfs_rq(se);
sched/fair: Initiate a new task's util avg to a bounded value A new task's util_avg is set to full utilization of a CPU (100% time running). This accelerates a new task's utilization ramp-up, useful to boost its execution in early time. However, it may result in (insanely) high utilization for a transient time period when a flood of tasks are spawned. Importantly, it violates the "fundamentally bounded" CPU utilization, and its side effect is negative if we don't take any measure to bound it. This patch proposes an algorithm to address this issue. It has two methods to approach a sensible initial util_avg: (1) An expected (or average) util_avg based on its cfs_rq's util_avg: util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight (2) A trajectory of how successive new tasks' util develops, which gives 1/2 of the left utilization budget to a new task such that the additional util is noticeably large (when overall util is low) or unnoticeably small (when overall util is high enough). In the meantime, the aggregate utilization is well bounded: util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n where n denotes the nth task. If util_avg is larger than util_avg_cap, then the effective util is clamped to the util_avg_cap. Reported-by: Andrey Ryabinin <aryabinin@virtuozzo.com> Signed-off-by: Yuyang Du <yuyang.du@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: bsegall@google.com Cc: morten.rasmussen@arm.com Cc: pjt@google.com Cc: steve.muckle@linaro.org Link: http://lkml.kernel.org/r/1459283456-21682-1-git-send-email-yuyang.du@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-03-29 20:30:56 +00:00
}
sched/fair: Fix PELT integrity for new tasks Vincent and Yuyang found another few scenarios in which entity tracking goes wobbly. The scenarios are basically due to the fact that new tasks are not immediately attached and thereby differ from the normal situation -- a task is always attached to a cfs_rq load average (such that it includes its blocked contribution) and are explicitly detached/attached on migration to another cfs_rq. Scenario 1: switch to fair class p->sched_class = fair_class; if (queued) enqueue_task(p); ... enqueue_entity() enqueue_entity_load_avg() migrated = !sa->last_update_time (true) if (migrated) attach_entity_load_avg() check_class_changed() switched_from() (!fair) switched_to() (fair) switched_to_fair() attach_entity_load_avg() If @p is a new task that hasn't been fair before, it will have !last_update_time and, per the above, end up in attach_entity_load_avg() _twice_. Scenario 2: change between cgroups sched_move_group(p) if (queued) dequeue_task() task_move_group_fair() detach_task_cfs_rq() detach_entity_load_avg() set_task_rq() attach_task_cfs_rq() attach_entity_load_avg() if (queued) enqueue_task(); ... enqueue_entity() enqueue_entity_load_avg() migrated = !sa->last_update_time (true) if (migrated) attach_entity_load_avg() Similar as with scenario 1, if @p is a new task, it will have !load_update_time and we'll end up in attach_entity_load_avg() _twice_. Furthermore, notice how we do a detach_entity_load_avg() on something that wasn't attached to begin with. As stated above; the problem is that the new task isn't yet attached to the load tracking and thereby violates the invariant assumption. This patch remedies this by ensuring a new task is indeed properly attached to the load tracking on creation, through post_init_entity_util_avg(). Of course, this isn't entirely as straightforward as one might think, since the task is hashed before we call wake_up_new_task() and thus can be poked at. We avoid this by adding TASK_NEW and teaching cpu_cgroup_can_attach() to refuse such tasks. Reported-by: Yuyang Du <yuyang.du@intel.com> Reported-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-06-16 11:29:28 +00:00
#else /* !CONFIG_SMP */
void init_entity_runnable_average(struct sched_entity *se)
sched: Set an initial value of runnable avg for new forked task We need to initialize the se.avg.{decay_count, load_avg_contrib} for a new forked task. Otherwise random values of above variables cause a mess when a new task is enqueued: enqueue_task_fair enqueue_entity enqueue_entity_load_avg and make fork balancing imbalance due to incorrect load_avg_contrib. Further more, Morten Rasmussen notice some tasks were not launched at once after created. So Paul and Peter suggest giving a start value for new task runnable avg time same as sched_slice(). PeterZ said: > So the 'problem' is that our running avg is a 'floating' average; ie. it > decays with time. Now we have to guess about the future of our newly > spawned task -- something that is nigh impossible seeing these CPU > vendors keep refusing to implement the crystal ball instruction. > > So there's two asymptotic cases we want to deal well with; 1) the case > where the newly spawned program will be 'nearly' idle for its lifetime; > and 2) the case where its cpu-bound. > > Since we have to guess, we'll go for worst case and assume its > cpu-bound; now we don't want to make the avg so heavy adjusting to the > near-idle case takes forever. We want to be able to quickly adjust and > lower our running avg. > > Now we also don't want to make our avg too light, such that it gets > decremented just for the new task not having had a chance to run yet -- > even if when it would run, it would be more cpu-bound than not. > > So what we do is we make the initial avg of the same duration as that we > guess it takes to run each task on the system at least once -- aka > sched_slice(). > > Of course we can defeat this with wakeup/fork bombs, but in the 'normal' > case it should be good enough. Paul also contributed most of the code comments in this commit. Signed-off-by: Alex Shi <alex.shi@intel.com> Reviewed-by: Gu Zheng <guz.fnst@cn.fujitsu.com> Reviewed-by: Paul Turner <pjt@google.com> [peterz; added explanation of sched_slice() usage] Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1371694737-29336-4-git-send-email-alex.shi@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-06-20 02:18:47 +00:00
{
}
void post_init_entity_util_avg(struct task_struct *p)
sched/fair: Initiate a new task's util avg to a bounded value A new task's util_avg is set to full utilization of a CPU (100% time running). This accelerates a new task's utilization ramp-up, useful to boost its execution in early time. However, it may result in (insanely) high utilization for a transient time period when a flood of tasks are spawned. Importantly, it violates the "fundamentally bounded" CPU utilization, and its side effect is negative if we don't take any measure to bound it. This patch proposes an algorithm to address this issue. It has two methods to approach a sensible initial util_avg: (1) An expected (or average) util_avg based on its cfs_rq's util_avg: util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight (2) A trajectory of how successive new tasks' util develops, which gives 1/2 of the left utilization budget to a new task such that the additional util is noticeably large (when overall util is low) or unnoticeably small (when overall util is high enough). In the meantime, the aggregate utilization is well bounded: util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n where n denotes the nth task. If util_avg is larger than util_avg_cap, then the effective util is clamped to the util_avg_cap. Reported-by: Andrey Ryabinin <aryabinin@virtuozzo.com> Signed-off-by: Yuyang Du <yuyang.du@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: bsegall@google.com Cc: morten.rasmussen@arm.com Cc: pjt@google.com Cc: steve.muckle@linaro.org Link: http://lkml.kernel.org/r/1459283456-21682-1-git-send-email-yuyang.du@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-03-29 20:30:56 +00:00
{
}
static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
{
}
sched/fair: Fix PELT integrity for new tasks Vincent and Yuyang found another few scenarios in which entity tracking goes wobbly. The scenarios are basically due to the fact that new tasks are not immediately attached and thereby differ from the normal situation -- a task is always attached to a cfs_rq load average (such that it includes its blocked contribution) and are explicitly detached/attached on migration to another cfs_rq. Scenario 1: switch to fair class p->sched_class = fair_class; if (queued) enqueue_task(p); ... enqueue_entity() enqueue_entity_load_avg() migrated = !sa->last_update_time (true) if (migrated) attach_entity_load_avg() check_class_changed() switched_from() (!fair) switched_to() (fair) switched_to_fair() attach_entity_load_avg() If @p is a new task that hasn't been fair before, it will have !last_update_time and, per the above, end up in attach_entity_load_avg() _twice_. Scenario 2: change between cgroups sched_move_group(p) if (queued) dequeue_task() task_move_group_fair() detach_task_cfs_rq() detach_entity_load_avg() set_task_rq() attach_task_cfs_rq() attach_entity_load_avg() if (queued) enqueue_task(); ... enqueue_entity() enqueue_entity_load_avg() migrated = !sa->last_update_time (true) if (migrated) attach_entity_load_avg() Similar as with scenario 1, if @p is a new task, it will have !load_update_time and we'll end up in attach_entity_load_avg() _twice_. Furthermore, notice how we do a detach_entity_load_avg() on something that wasn't attached to begin with. As stated above; the problem is that the new task isn't yet attached to the load tracking and thereby violates the invariant assumption. This patch remedies this by ensuring a new task is indeed properly attached to the load tracking on creation, through post_init_entity_util_avg(). Of course, this isn't entirely as straightforward as one might think, since the task is hashed before we call wake_up_new_task() and thus can be poked at. We avoid this by adding TASK_NEW and teaching cpu_cgroup_can_attach() to refuse such tasks. Reported-by: Yuyang Du <yuyang.du@intel.com> Reported-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-06-16 11:29:28 +00:00
#endif /* CONFIG_SMP */
sched: Set an initial value of runnable avg for new forked task We need to initialize the se.avg.{decay_count, load_avg_contrib} for a new forked task. Otherwise random values of above variables cause a mess when a new task is enqueued: enqueue_task_fair enqueue_entity enqueue_entity_load_avg and make fork balancing imbalance due to incorrect load_avg_contrib. Further more, Morten Rasmussen notice some tasks were not launched at once after created. So Paul and Peter suggest giving a start value for new task runnable avg time same as sched_slice(). PeterZ said: > So the 'problem' is that our running avg is a 'floating' average; ie. it > decays with time. Now we have to guess about the future of our newly > spawned task -- something that is nigh impossible seeing these CPU > vendors keep refusing to implement the crystal ball instruction. > > So there's two asymptotic cases we want to deal well with; 1) the case > where the newly spawned program will be 'nearly' idle for its lifetime; > and 2) the case where its cpu-bound. > > Since we have to guess, we'll go for worst case and assume its > cpu-bound; now we don't want to make the avg so heavy adjusting to the > near-idle case takes forever. We want to be able to quickly adjust and > lower our running avg. > > Now we also don't want to make our avg too light, such that it gets > decremented just for the new task not having had a chance to run yet -- > even if when it would run, it would be more cpu-bound than not. > > So what we do is we make the initial avg of the same duration as that we > guess it takes to run each task on the system at least once -- aka > sched_slice(). > > Of course we can defeat this with wakeup/fork bombs, but in the 'normal' > case it should be good enough. Paul also contributed most of the code comments in this commit. Signed-off-by: Alex Shi <alex.shi@intel.com> Reviewed-by: Gu Zheng <guz.fnst@cn.fujitsu.com> Reviewed-by: Paul Turner <pjt@google.com> [peterz; added explanation of sched_slice() usage] Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1371694737-29336-4-git-send-email-alex.shi@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-06-20 02:18:47 +00:00
/*
sched/fair: Rework sched_fair time accounting Christian suffers from a bad BIOS that wrecks his i5's TSC sync. This results in him occasionally seeing time going backwards - which crashes the scheduler ... Most of our time accounting can actually handle that except the most common one; the tick time update of sched_fair. There is a further problem with that code; previously we assumed that because we get a tick every TICK_NSEC our time delta could never exceed 32bits and math was simpler. However, ever since Frederic managed to get NO_HZ_FULL merged; this is no longer the case since now a task can run for a long time indeed without getting a tick. It only takes about ~4.2 seconds to overflow our u32 in nanoseconds. This means we not only need to better deal with time going backwards; but also means we need to be able to deal with large deltas. This patch reworks the entire code and uses mul_u64_u32_shr() as proposed by Andy a long while ago. We express our virtual time scale factor in a u32 multiplier and shift right and the 32bit mul_u64_u32_shr() implementation reduces to a single 32x32->64 multiply if the time delta is still short (common case). For 64bit a 64x64->128 multiply can be used if ARCH_SUPPORTS_INT128. Reported-and-Tested-by: Christian Engelmayer <cengelma@gmx.at> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Cc: fweisbec@gmail.com Cc: Paul Turner <pjt@google.com> Cc: Stanislaw Gruszka <sgruszka@redhat.com> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Link: http://lkml.kernel.org/r/20131118172706.GI3866@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-11-18 17:27:06 +00:00
* Update the current task's runtime statistics.
*/
static void update_curr(struct cfs_rq *cfs_rq)
{
struct sched_entity *curr = cfs_rq->curr;
u64 now = rq_clock_task(rq_of(cfs_rq));
sched/fair: Rework sched_fair time accounting Christian suffers from a bad BIOS that wrecks his i5's TSC sync. This results in him occasionally seeing time going backwards - which crashes the scheduler ... Most of our time accounting can actually handle that except the most common one; the tick time update of sched_fair. There is a further problem with that code; previously we assumed that because we get a tick every TICK_NSEC our time delta could never exceed 32bits and math was simpler. However, ever since Frederic managed to get NO_HZ_FULL merged; this is no longer the case since now a task can run for a long time indeed without getting a tick. It only takes about ~4.2 seconds to overflow our u32 in nanoseconds. This means we not only need to better deal with time going backwards; but also means we need to be able to deal with large deltas. This patch reworks the entire code and uses mul_u64_u32_shr() as proposed by Andy a long while ago. We express our virtual time scale factor in a u32 multiplier and shift right and the 32bit mul_u64_u32_shr() implementation reduces to a single 32x32->64 multiply if the time delta is still short (common case). For 64bit a 64x64->128 multiply can be used if ARCH_SUPPORTS_INT128. Reported-and-Tested-by: Christian Engelmayer <cengelma@gmx.at> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Cc: fweisbec@gmail.com Cc: Paul Turner <pjt@google.com> Cc: Stanislaw Gruszka <sgruszka@redhat.com> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Link: http://lkml.kernel.org/r/20131118172706.GI3866@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-11-18 17:27:06 +00:00
u64 delta_exec;
if (unlikely(!curr))
return;
sched/fair: Rework sched_fair time accounting Christian suffers from a bad BIOS that wrecks his i5's TSC sync. This results in him occasionally seeing time going backwards - which crashes the scheduler ... Most of our time accounting can actually handle that except the most common one; the tick time update of sched_fair. There is a further problem with that code; previously we assumed that because we get a tick every TICK_NSEC our time delta could never exceed 32bits and math was simpler. However, ever since Frederic managed to get NO_HZ_FULL merged; this is no longer the case since now a task can run for a long time indeed without getting a tick. It only takes about ~4.2 seconds to overflow our u32 in nanoseconds. This means we not only need to better deal with time going backwards; but also means we need to be able to deal with large deltas. This patch reworks the entire code and uses mul_u64_u32_shr() as proposed by Andy a long while ago. We express our virtual time scale factor in a u32 multiplier and shift right and the 32bit mul_u64_u32_shr() implementation reduces to a single 32x32->64 multiply if the time delta is still short (common case). For 64bit a 64x64->128 multiply can be used if ARCH_SUPPORTS_INT128. Reported-and-Tested-by: Christian Engelmayer <cengelma@gmx.at> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Cc: fweisbec@gmail.com Cc: Paul Turner <pjt@google.com> Cc: Stanislaw Gruszka <sgruszka@redhat.com> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Link: http://lkml.kernel.org/r/20131118172706.GI3866@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-11-18 17:27:06 +00:00
delta_exec = now - curr->exec_start;
if (unlikely((s64)delta_exec <= 0))
return;
curr->exec_start = now;
sched: cpu accounting controller (V2) Commit cfb5285660aad4931b2ebbfa902ea48a37dfffa1 removed a useful feature for us, which provided a cpu accounting resource controller. This feature would be useful if someone wants to group tasks only for accounting purpose and doesnt really want to exercise any control over their cpu consumption. The patch below reintroduces the feature. It is based on Paul Menage's original patch (Commit 62d0df64065e7c135d0002f069444fbdfc64768f), with these differences: - Removed load average information. I felt it needs more thought (esp to deal with SMP and virtualized platforms) and can be added for 2.6.25 after more discussions. - Convert group cpu usage to be nanosecond accurate (as rest of the cfs stats are) and invoke cpuacct_charge() from the respective scheduler classes - Make accounting scalable on SMP systems by splitting the usage counter to be per-cpu - Move the code from kernel/cpu_acct.c to kernel/sched.c (since the code is not big enough to warrant a new file and also this rightly needs to live inside the scheduler. Also things like accessing rq->lock while reading cpu usage becomes easier if the code lived in kernel/sched.c) The patch also modifies the cpu controller not to provide the same accounting information. Tested-by: Balbir Singh <balbir@linux.vnet.ibm.com> Tested the patches on top of 2.6.24-rc3. The patches work fine. Ran some simple tests like cpuspin (spin on the cpu), ran several tasks in the same group and timed them. Compared their time stamps with cpuacct.usage. Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Signed-off-by: Balbir Singh <balbir@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2007-12-02 19:04:49 +00:00
sched/fair: Rework sched_fair time accounting Christian suffers from a bad BIOS that wrecks his i5's TSC sync. This results in him occasionally seeing time going backwards - which crashes the scheduler ... Most of our time accounting can actually handle that except the most common one; the tick time update of sched_fair. There is a further problem with that code; previously we assumed that because we get a tick every TICK_NSEC our time delta could never exceed 32bits and math was simpler. However, ever since Frederic managed to get NO_HZ_FULL merged; this is no longer the case since now a task can run for a long time indeed without getting a tick. It only takes about ~4.2 seconds to overflow our u32 in nanoseconds. This means we not only need to better deal with time going backwards; but also means we need to be able to deal with large deltas. This patch reworks the entire code and uses mul_u64_u32_shr() as proposed by Andy a long while ago. We express our virtual time scale factor in a u32 multiplier and shift right and the 32bit mul_u64_u32_shr() implementation reduces to a single 32x32->64 multiply if the time delta is still short (common case). For 64bit a 64x64->128 multiply can be used if ARCH_SUPPORTS_INT128. Reported-and-Tested-by: Christian Engelmayer <cengelma@gmx.at> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Cc: fweisbec@gmail.com Cc: Paul Turner <pjt@google.com> Cc: Stanislaw Gruszka <sgruszka@redhat.com> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Link: http://lkml.kernel.org/r/20131118172706.GI3866@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-11-18 17:27:06 +00:00
schedstat_set(curr->statistics.exec_max,
max(delta_exec, curr->statistics.exec_max));
curr->sum_exec_runtime += delta_exec;
schedstat_add(cfs_rq->exec_clock, delta_exec);
sched/fair: Rework sched_fair time accounting Christian suffers from a bad BIOS that wrecks his i5's TSC sync. This results in him occasionally seeing time going backwards - which crashes the scheduler ... Most of our time accounting can actually handle that except the most common one; the tick time update of sched_fair. There is a further problem with that code; previously we assumed that because we get a tick every TICK_NSEC our time delta could never exceed 32bits and math was simpler. However, ever since Frederic managed to get NO_HZ_FULL merged; this is no longer the case since now a task can run for a long time indeed without getting a tick. It only takes about ~4.2 seconds to overflow our u32 in nanoseconds. This means we not only need to better deal with time going backwards; but also means we need to be able to deal with large deltas. This patch reworks the entire code and uses mul_u64_u32_shr() as proposed by Andy a long while ago. We express our virtual time scale factor in a u32 multiplier and shift right and the 32bit mul_u64_u32_shr() implementation reduces to a single 32x32->64 multiply if the time delta is still short (common case). For 64bit a 64x64->128 multiply can be used if ARCH_SUPPORTS_INT128. Reported-and-Tested-by: Christian Engelmayer <cengelma@gmx.at> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Cc: fweisbec@gmail.com Cc: Paul Turner <pjt@google.com> Cc: Stanislaw Gruszka <sgruszka@redhat.com> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Link: http://lkml.kernel.org/r/20131118172706.GI3866@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-11-18 17:27:06 +00:00
curr->vruntime += calc_delta_fair(delta_exec, curr);
update_min_vruntime(cfs_rq);
sched: cpu accounting controller (V2) Commit cfb5285660aad4931b2ebbfa902ea48a37dfffa1 removed a useful feature for us, which provided a cpu accounting resource controller. This feature would be useful if someone wants to group tasks only for accounting purpose and doesnt really want to exercise any control over their cpu consumption. The patch below reintroduces the feature. It is based on Paul Menage's original patch (Commit 62d0df64065e7c135d0002f069444fbdfc64768f), with these differences: - Removed load average information. I felt it needs more thought (esp to deal with SMP and virtualized platforms) and can be added for 2.6.25 after more discussions. - Convert group cpu usage to be nanosecond accurate (as rest of the cfs stats are) and invoke cpuacct_charge() from the respective scheduler classes - Make accounting scalable on SMP systems by splitting the usage counter to be per-cpu - Move the code from kernel/cpu_acct.c to kernel/sched.c (since the code is not big enough to warrant a new file and also this rightly needs to live inside the scheduler. Also things like accessing rq->lock while reading cpu usage becomes easier if the code lived in kernel/sched.c) The patch also modifies the cpu controller not to provide the same accounting information. Tested-by: Balbir Singh <balbir@linux.vnet.ibm.com> Tested the patches on top of 2.6.24-rc3. The patches work fine. Ran some simple tests like cpuspin (spin on the cpu), ran several tasks in the same group and timed them. Compared their time stamps with cpuacct.usage. Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Signed-off-by: Balbir Singh <balbir@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2007-12-02 19:04:49 +00:00
if (entity_is_task(curr)) {
struct task_struct *curtask = task_of(curr);
trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
cgroup_account_cputime(curtask, delta_exec);
timers: fix itimer/many thread hang Overview This patch reworks the handling of POSIX CPU timers, including the ITIMER_PROF, ITIMER_VIRT timers and rlimit handling. It was put together with the help of Roland McGrath, the owner and original writer of this code. The problem we ran into, and the reason for this rework, has to do with using a profiling timer in a process with a large number of threads. It appears that the performance of the old implementation of run_posix_cpu_timers() was at least O(n*3) (where "n" is the number of threads in a process) or worse. Everything is fine with an increasing number of threads until the time taken for that routine to run becomes the same as or greater than the tick time, at which point things degrade rather quickly. This patch fixes bug 9906, "Weird hang with NPTL and SIGPROF." Code Changes This rework corrects the implementation of run_posix_cpu_timers() to make it run in constant time for a particular machine. (Performance may vary between one machine and another depending upon whether the kernel is built as single- or multiprocessor and, in the latter case, depending upon the number of running processors.) To do this, at each tick we now update fields in signal_struct as well as task_struct. The run_posix_cpu_timers() function uses those fields to make its decisions. We define a new structure, "task_cputime," to contain user, system and scheduler times and use these in appropriate places: struct task_cputime { cputime_t utime; cputime_t stime; unsigned long long sum_exec_runtime; }; This is included in the structure "thread_group_cputime," which is a new substructure of signal_struct and which varies for uniprocessor versus multiprocessor kernels. For uniprocessor kernels, it uses "task_cputime" as a simple substructure, while for multiprocessor kernels it is a pointer: struct thread_group_cputime { struct task_cputime totals; }; struct thread_group_cputime { struct task_cputime *totals; }; We also add a new task_cputime substructure directly to signal_struct, to cache the earliest expiration of process-wide timers, and task_cputime also replaces the it_*_expires fields of task_struct (used for earliest expiration of thread timers). The "thread_group_cputime" structure contains process-wide timers that are updated via account_user_time() and friends. In the non-SMP case the structure is a simple aggregator; unfortunately in the SMP case that simplicity was not achievable due to cache-line contention between CPUs (in one measured case performance was actually _worse_ on a 16-cpu system than the same test on a 4-cpu system, due to this contention). For SMP, the thread_group_cputime counters are maintained as a per-cpu structure allocated using alloc_percpu(). The timer functions update only the timer field in the structure corresponding to the running CPU, obtained using per_cpu_ptr(). We define a set of inline functions in sched.h that we use to maintain the thread_group_cputime structure and hide the differences between UP and SMP implementations from the rest of the kernel. The thread_group_cputime_init() function initializes the thread_group_cputime structure for the given task. The thread_group_cputime_alloc() is a no-op for UP; for SMP it calls the out-of-line function thread_group_cputime_alloc_smp() to allocate and fill in the per-cpu structures and fields. The thread_group_cputime_free() function, also a no-op for UP, in SMP frees the per-cpu structures. The thread_group_cputime_clone_thread() function (also a UP no-op) for SMP calls thread_group_cputime_alloc() if the per-cpu structures haven't yet been allocated. The thread_group_cputime() function fills the task_cputime structure it is passed with the contents of the thread_group_cputime fields; in UP it's that simple but in SMP it must also safely check that tsk->signal is non-NULL (if it is it just uses the appropriate fields of task_struct) and, if so, sums the per-cpu values for each online CPU. Finally, the three functions account_group_user_time(), account_group_system_time() and account_group_exec_runtime() are used by timer functions to update the respective fields of the thread_group_cputime structure. Non-SMP operation is trivial and will not be mentioned further. The per-cpu structure is always allocated when a task creates its first new thread, via a call to thread_group_cputime_clone_thread() from copy_signal(). It is freed at process exit via a call to thread_group_cputime_free() from cleanup_signal(). All functions that formerly summed utime/stime/sum_sched_runtime values from from all threads in the thread group now use thread_group_cputime() to snapshot the values in the thread_group_cputime structure or the values in the task structure itself if the per-cpu structure hasn't been allocated. Finally, the code in kernel/posix-cpu-timers.c has changed quite a bit. The run_posix_cpu_timers() function has been split into a fast path and a slow path; the former safely checks whether there are any expired thread timers and, if not, just returns, while the slow path does the heavy lifting. With the dedicated thread group fields, timers are no longer "rebalanced" and the process_timer_rebalance() function and related code has gone away. All summing loops are gone and all code that used them now uses the thread_group_cputime() inline. When process-wide timers are set, the new task_cputime structure in signal_struct is used to cache the earliest expiration; this is checked in the fast path. Performance The fix appears not to add significant overhead to existing operations. It generally performs the same as the current code except in two cases, one in which it performs slightly worse (Case 5 below) and one in which it performs very significantly better (Case 2 below). Overall it's a wash except in those two cases. I've since done somewhat more involved testing on a dual-core Opteron system. Case 1: With no itimer running, for a test with 100,000 threads, the fixed kernel took 1428.5 seconds, 513 seconds more than the unfixed system, all of which was spent in the system. There were twice as many voluntary context switches with the fix as without it. Case 2: With an itimer running at .01 second ticks and 4000 threads (the most an unmodified kernel can handle), the fixed kernel ran the test in eight percent of the time (5.8 seconds as opposed to 70 seconds) and had better tick accuracy (.012 seconds per tick as opposed to .023 seconds per tick). Case 3: A 4000-thread test with an initial timer tick of .01 second and an interval of 10,000 seconds (i.e. a timer that ticks only once) had very nearly the same performance in both cases: 6.3 seconds elapsed for the fixed kernel versus 5.5 seconds for the unfixed kernel. With fewer threads (eight in these tests), the Case 1 test ran in essentially the same time on both the modified and unmodified kernels (5.2 seconds versus 5.8 seconds). The Case 2 test ran in about the same time as well, 5.9 seconds versus 5.4 seconds but again with much better tick accuracy, .013 seconds per tick versus .025 seconds per tick for the unmodified kernel. Since the fix affected the rlimit code, I also tested soft and hard CPU limits. Case 4: With a hard CPU limit of 20 seconds and eight threads (and an itimer running), the modified kernel was very slightly favored in that while it killed the process in 19.997 seconds of CPU time (5.002 seconds of wall time), only .003 seconds of that was system time, the rest was user time. The unmodified kernel killed the process in 20.001 seconds of CPU (5.014 seconds of wall time) of which .016 seconds was system time. Really, though, the results were too close to call. The results were essentially the same with no itimer running. Case 5: With a soft limit of 20 seconds and a hard limit of 2000 seconds (where the hard limit would never be reached) and an itimer running, the modified kernel exhibited worse tick accuracy than the unmodified kernel: .050 seconds/tick versus .028 seconds/tick. Otherwise, performance was almost indistinguishable. With no itimer running this test exhibited virtually identical behavior and times in both cases. In times past I did some limited performance testing. those results are below. On a four-cpu Opteron system without this fix, a sixteen-thread test executed in 3569.991 seconds, of which user was 3568.435s and system was 1.556s. On the same system with the fix, user and elapsed time were about the same, but system time dropped to 0.007 seconds. Performance with eight, four and one thread were comparable. Interestingly, the timer ticks with the fix seemed more accurate: The sixteen-thread test with the fix received 149543 ticks for 0.024 seconds per tick, while the same test without the fix received 58720 for 0.061 seconds per tick. Both cases were configured for an interval of 0.01 seconds. Again, the other tests were comparable. Each thread in this test computed the primes up to 25,000,000. I also did a test with a large number of threads, 100,000 threads, which is impossible without the fix. In this case each thread computed the primes only up to 10,000 (to make the runtime manageable). System time dominated, at 1546.968 seconds out of a total 2176.906 seconds (giving a user time of 629.938s). It received 147651 ticks for 0.015 seconds per tick, still quite accurate. There is obviously no comparable test without the fix. Signed-off-by: Frank Mayhar <fmayhar@google.com> Cc: Roland McGrath <roland@redhat.com> Cc: Alexey Dobriyan <adobriyan@gmail.com> Cc: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-09-12 16:54:39 +00:00
account_group_exec_runtime(curtask, delta_exec);
sched: cpu accounting controller (V2) Commit cfb5285660aad4931b2ebbfa902ea48a37dfffa1 removed a useful feature for us, which provided a cpu accounting resource controller. This feature would be useful if someone wants to group tasks only for accounting purpose and doesnt really want to exercise any control over their cpu consumption. The patch below reintroduces the feature. It is based on Paul Menage's original patch (Commit 62d0df64065e7c135d0002f069444fbdfc64768f), with these differences: - Removed load average information. I felt it needs more thought (esp to deal with SMP and virtualized platforms) and can be added for 2.6.25 after more discussions. - Convert group cpu usage to be nanosecond accurate (as rest of the cfs stats are) and invoke cpuacct_charge() from the respective scheduler classes - Make accounting scalable on SMP systems by splitting the usage counter to be per-cpu - Move the code from kernel/cpu_acct.c to kernel/sched.c (since the code is not big enough to warrant a new file and also this rightly needs to live inside the scheduler. Also things like accessing rq->lock while reading cpu usage becomes easier if the code lived in kernel/sched.c) The patch also modifies the cpu controller not to provide the same accounting information. Tested-by: Balbir Singh <balbir@linux.vnet.ibm.com> Tested the patches on top of 2.6.24-rc3. The patches work fine. Ran some simple tests like cpuspin (spin on the cpu), ran several tasks in the same group and timed them. Compared their time stamps with cpuacct.usage. Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Signed-off-by: Balbir Singh <balbir@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2007-12-02 19:04:49 +00:00
}
account_cfs_rq_runtime(cfs_rq, delta_exec);
}
sched/cputime: Fix clock_nanosleep()/clock_gettime() inconsistency Commit d670ec13178d0 "posix-cpu-timers: Cure SMP wobbles" fixes one glibc test case in cost of breaking another one. After that commit, calling clock_nanosleep(TIMER_ABSTIME, X) and then clock_gettime(&Y) can result of Y time being smaller than X time. Reproducer/tester can be found further below, it can be compiled and ran by: gcc -o tst-cpuclock2 tst-cpuclock2.c -pthread while ./tst-cpuclock2 ; do : ; done This reproducer, when running on a buggy kernel, will complain about "clock_gettime difference too small". Issue happens because on start in thread_group_cputimer() we initialize sum_exec_runtime of cputimer with threads runtime not yet accounted and then add the threads runtime to running cputimer again on scheduler tick, making it's sum_exec_runtime bigger than actual threads runtime. KOSAKI Motohiro posted a fix for this problem, but that patch was never applied: https://lkml.org/lkml/2013/5/26/191 . This patch takes different approach to cure the problem. It calls update_curr() when cputimer starts, that assure we will have updated stats of running threads and on the next schedule tick we will account only the runtime that elapsed from cputimer start. That also assure we have consistent state between cpu times of individual threads and cpu time of the process consisted by those threads. Full reproducer (tst-cpuclock2.c): #define _GNU_SOURCE #include <unistd.h> #include <sys/syscall.h> #include <stdio.h> #include <time.h> #include <pthread.h> #include <stdint.h> #include <inttypes.h> /* Parameters for the Linux kernel ABI for CPU clocks. */ #define CPUCLOCK_SCHED 2 #define MAKE_PROCESS_CPUCLOCK(pid, clock) \ ((~(clockid_t) (pid) << 3) | (clockid_t) (clock)) static pthread_barrier_t barrier; /* Help advance the clock. */ static void *chew_cpu(void *arg) { pthread_barrier_wait(&barrier); while (1) ; return NULL; } /* Don't use the glibc wrapper. */ static int do_nanosleep(int flags, const struct timespec *req) { clockid_t clock_id = MAKE_PROCESS_CPUCLOCK(0, CPUCLOCK_SCHED); return syscall(SYS_clock_nanosleep, clock_id, flags, req, NULL); } static int64_t tsdiff(const struct timespec *before, const struct timespec *after) { int64_t before_i = before->tv_sec * 1000000000ULL + before->tv_nsec; int64_t after_i = after->tv_sec * 1000000000ULL + after->tv_nsec; return after_i - before_i; } int main(void) { int result = 0; pthread_t th; pthread_barrier_init(&barrier, NULL, 2); if (pthread_create(&th, NULL, chew_cpu, NULL) != 0) { perror("pthread_create"); return 1; } pthread_barrier_wait(&barrier); /* The test. */ struct timespec before, after, sleeptimeabs; int64_t sleepdiff, diffabs; const struct timespec sleeptime = {.tv_sec = 0,.tv_nsec = 100000000 }; /* The relative nanosleep. Not sure why this is needed, but its presence seems to make it easier to reproduce the problem. */ if (do_nanosleep(0, &sleeptime) != 0) { perror("clock_nanosleep"); return 1; } /* Get the current time. */ if (clock_gettime(CLOCK_PROCESS_CPUTIME_ID, &before) < 0) { perror("clock_gettime[2]"); return 1; } /* Compute the absolute sleep time based on the current time. */ uint64_t nsec = before.tv_nsec + sleeptime.tv_nsec; sleeptimeabs.tv_sec = before.tv_sec + nsec / 1000000000; sleeptimeabs.tv_nsec = nsec % 1000000000; /* Sleep for the computed time. */ if (do_nanosleep(TIMER_ABSTIME, &sleeptimeabs) != 0) { perror("absolute clock_nanosleep"); return 1; } /* Get the time after the sleep. */ if (clock_gettime(CLOCK_PROCESS_CPUTIME_ID, &after) < 0) { perror("clock_gettime[3]"); return 1; } /* The time after sleep should always be equal to or after the absolute sleep time passed to clock_nanosleep. */ sleepdiff = tsdiff(&sleeptimeabs, &after); if (sleepdiff < 0) { printf("absolute clock_nanosleep woke too early: %" PRId64 "\n", sleepdiff); result = 1; printf("Before %llu.%09llu\n", before.tv_sec, before.tv_nsec); printf("After %llu.%09llu\n", after.tv_sec, after.tv_nsec); printf("Sleep %llu.%09llu\n", sleeptimeabs.tv_sec, sleeptimeabs.tv_nsec); } /* The difference between the timestamps taken before and after the clock_nanosleep call should be equal to or more than the duration of the sleep. */ diffabs = tsdiff(&before, &after); if (diffabs < sleeptime.tv_nsec) { printf("clock_gettime difference too small: %" PRId64 "\n", diffabs); result = 1; } pthread_cancel(th); return result; } Signed-off-by: Stanislaw Gruszka <sgruszka@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Frederic Weisbecker <fweisbec@gmail.com> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/20141112155843.GA24803@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-11-12 15:58:44 +00:00
static void update_curr_fair(struct rq *rq)
{
update_curr(cfs_rq_of(&rq->curr->se));
}
static inline void
update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
{
u64 wait_start, prev_wait_start;
if (!schedstat_enabled())
return;
wait_start = rq_clock(rq_of(cfs_rq));
prev_wait_start = schedstat_val(se->statistics.wait_start);
if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
likely(wait_start > prev_wait_start))
wait_start -= prev_wait_start;
__schedstat_set(se->statistics.wait_start, wait_start);
}
static inline void
update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
{
struct task_struct *p;
sched/debug: Make schedstats a runtime tunable that is disabled by default schedstats is very useful during debugging and performance tuning but it incurs overhead to calculate the stats. As such, even though it can be disabled at build time, it is often enabled as the information is useful. This patch adds a kernel command-line and sysctl tunable to enable or disable schedstats on demand (when it's built in). It is disabled by default as someone who knows they need it can also learn to enable it when necessary. The benefits are dependent on how scheduler-intensive the workload is. If it is then the patch reduces the number of cycles spent calculating the stats with a small benefit from reducing the cache footprint of the scheduler. These measurements were taken from a 48-core 2-socket machine with Xeon(R) E5-2670 v3 cpus although they were also tested on a single socket machine 8-core machine with Intel i7-3770 processors. netperf-tcp 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v3r1 Hmean 64 560.45 ( 0.00%) 575.98 ( 2.77%) Hmean 128 766.66 ( 0.00%) 795.79 ( 3.80%) Hmean 256 950.51 ( 0.00%) 981.50 ( 3.26%) Hmean 1024 1433.25 ( 0.00%) 1466.51 ( 2.32%) Hmean 2048 2810.54 ( 0.00%) 2879.75 ( 2.46%) Hmean 3312 4618.18 ( 0.00%) 4682.09 ( 1.38%) Hmean 4096 5306.42 ( 0.00%) 5346.39 ( 0.75%) Hmean 8192 10581.44 ( 0.00%) 10698.15 ( 1.10%) Hmean 16384 18857.70 ( 0.00%) 18937.61 ( 0.42%) Small gains here, UDP_STREAM showed nothing intresting and neither did the TCP_RR tests. The gains on the 8-core machine were very similar. tbench4 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v3r1 Hmean mb/sec-1 500.85 ( 0.00%) 522.43 ( 4.31%) Hmean mb/sec-2 984.66 ( 0.00%) 1018.19 ( 3.41%) Hmean mb/sec-4 1827.91 ( 0.00%) 1847.78 ( 1.09%) Hmean mb/sec-8 3561.36 ( 0.00%) 3611.28 ( 1.40%) Hmean mb/sec-16 5824.52 ( 0.00%) 5929.03 ( 1.79%) Hmean mb/sec-32 10943.10 ( 0.00%) 10802.83 ( -1.28%) Hmean mb/sec-64 15950.81 ( 0.00%) 16211.31 ( 1.63%) Hmean mb/sec-128 15302.17 ( 0.00%) 15445.11 ( 0.93%) Hmean mb/sec-256 14866.18 ( 0.00%) 15088.73 ( 1.50%) Hmean mb/sec-512 15223.31 ( 0.00%) 15373.69 ( 0.99%) Hmean mb/sec-1024 14574.25 ( 0.00%) 14598.02 ( 0.16%) Hmean mb/sec-2048 13569.02 ( 0.00%) 13733.86 ( 1.21%) Hmean mb/sec-3072 12865.98 ( 0.00%) 13209.23 ( 2.67%) Small gains of 2-4% at low thread counts and otherwise flat. The gains on the 8-core machine were slightly different tbench4 on 8-core i7-3770 single socket machine Hmean mb/sec-1 442.59 ( 0.00%) 448.73 ( 1.39%) Hmean mb/sec-2 796.68 ( 0.00%) 794.39 ( -0.29%) Hmean mb/sec-4 1322.52 ( 0.00%) 1343.66 ( 1.60%) Hmean mb/sec-8 2611.65 ( 0.00%) 2694.86 ( 3.19%) Hmean mb/sec-16 2537.07 ( 0.00%) 2609.34 ( 2.85%) Hmean mb/sec-32 2506.02 ( 0.00%) 2578.18 ( 2.88%) Hmean mb/sec-64 2511.06 ( 0.00%) 2569.16 ( 2.31%) Hmean mb/sec-128 2313.38 ( 0.00%) 2395.50 ( 3.55%) Hmean mb/sec-256 2110.04 ( 0.00%) 2177.45 ( 3.19%) Hmean mb/sec-512 2072.51 ( 0.00%) 2053.97 ( -0.89%) In constract, this shows a relatively steady 2-3% gain at higher thread counts. Due to the nature of the patch and the type of workload, it's not a surprise that the result will depend on the CPU used. hackbench-pipes 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v3r1 Amean 1 0.0637 ( 0.00%) 0.0660 ( -3.59%) Amean 4 0.1229 ( 0.00%) 0.1181 ( 3.84%) Amean 7 0.1921 ( 0.00%) 0.1911 ( 0.52%) Amean 12 0.3117 ( 0.00%) 0.2923 ( 6.23%) Amean 21 0.4050 ( 0.00%) 0.3899 ( 3.74%) Amean 30 0.4586 ( 0.00%) 0.4433 ( 3.33%) Amean 48 0.5910 ( 0.00%) 0.5694 ( 3.65%) Amean 79 0.8663 ( 0.00%) 0.8626 ( 0.43%) Amean 110 1.1543 ( 0.00%) 1.1517 ( 0.22%) Amean 141 1.4457 ( 0.00%) 1.4290 ( 1.16%) Amean 172 1.7090 ( 0.00%) 1.6924 ( 0.97%) Amean 192 1.9126 ( 0.00%) 1.9089 ( 0.19%) Some small gains and losses and while the variance data is not included, it's close to the noise. The UMA machine did not show anything particularly different pipetest 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v2r2 Min Time 4.13 ( 0.00%) 3.99 ( 3.39%) 1st-qrtle Time 4.38 ( 0.00%) 4.27 ( 2.51%) 2nd-qrtle Time 4.46 ( 0.00%) 4.39 ( 1.57%) 3rd-qrtle Time 4.56 ( 0.00%) 4.51 ( 1.10%) Max-90% Time 4.67 ( 0.00%) 4.60 ( 1.50%) Max-93% Time 4.71 ( 0.00%) 4.65 ( 1.27%) Max-95% Time 4.74 ( 0.00%) 4.71 ( 0.63%) Max-99% Time 4.88 ( 0.00%) 4.79 ( 1.84%) Max Time 4.93 ( 0.00%) 4.83 ( 2.03%) Mean Time 4.48 ( 0.00%) 4.39 ( 1.91%) Best99%Mean Time 4.47 ( 0.00%) 4.39 ( 1.91%) Best95%Mean Time 4.46 ( 0.00%) 4.38 ( 1.93%) Best90%Mean Time 4.45 ( 0.00%) 4.36 ( 1.98%) Best50%Mean Time 4.36 ( 0.00%) 4.25 ( 2.49%) Best10%Mean Time 4.23 ( 0.00%) 4.10 ( 3.13%) Best5%Mean Time 4.19 ( 0.00%) 4.06 ( 3.20%) Best1%Mean Time 4.13 ( 0.00%) 4.00 ( 3.39%) Small improvement and similar gains were seen on the UMA machine. The gain is small but it stands to reason that doing less work in the scheduler is a good thing. The downside is that the lack of schedstats and tracepoints may be surprising to experts doing performance analysis until they find the existence of the schedstats= parameter or schedstats sysctl. It will be automatically activated for latencytop and sleep profiling to alleviate the problem. For tracepoints, there is a simple warning as it's not safe to activate schedstats in the context when it's known the tracepoint may be wanted but is unavailable. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Reviewed-by: Matt Fleming <matt@codeblueprint.co.uk> Reviewed-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <mgalbraith@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1454663316-22048-1-git-send-email-mgorman@techsingularity.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-02-05 09:08:36 +00:00
u64 delta;
if (!schedstat_enabled())
return;
delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
if (entity_is_task(se)) {
p = task_of(se);
if (task_on_rq_migrating(p)) {
/*
* Preserve migrating task's wait time so wait_start
* time stamp can be adjusted to accumulate wait time
* prior to migration.
*/
__schedstat_set(se->statistics.wait_start, delta);
return;
}
trace_sched_stat_wait(p, delta);
}
__schedstat_set(se->statistics.wait_max,
max(schedstat_val(se->statistics.wait_max), delta));
__schedstat_inc(se->statistics.wait_count);
__schedstat_add(se->statistics.wait_sum, delta);
__schedstat_set(se->statistics.wait_start, 0);
}
static inline void
update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
{
struct task_struct *tsk = NULL;
u64 sleep_start, block_start;
if (!schedstat_enabled())
return;
sleep_start = schedstat_val(se->statistics.sleep_start);
block_start = schedstat_val(se->statistics.block_start);
if (entity_is_task(se))
tsk = task_of(se);
if (sleep_start) {
u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start;
if ((s64)delta < 0)
delta = 0;
if (unlikely(delta > schedstat_val(se->statistics.sleep_max)))
__schedstat_set(se->statistics.sleep_max, delta);
__schedstat_set(se->statistics.sleep_start, 0);
__schedstat_add(se->statistics.sum_sleep_runtime, delta);
if (tsk) {
account_scheduler_latency(tsk, delta >> 10, 1);
trace_sched_stat_sleep(tsk, delta);
}
}
if (block_start) {
u64 delta = rq_clock(rq_of(cfs_rq)) - block_start;
if ((s64)delta < 0)
delta = 0;
if (unlikely(delta > schedstat_val(se->statistics.block_max)))
__schedstat_set(se->statistics.block_max, delta);
__schedstat_set(se->statistics.block_start, 0);
__schedstat_add(se->statistics.sum_sleep_runtime, delta);
if (tsk) {
if (tsk->in_iowait) {
__schedstat_add(se->statistics.iowait_sum, delta);
__schedstat_inc(se->statistics.iowait_count);
trace_sched_stat_iowait(tsk, delta);
}
trace_sched_stat_blocked(tsk, delta);
/*
* Blocking time is in units of nanosecs, so shift by
* 20 to get a milliseconds-range estimation of the
* amount of time that the task spent sleeping:
*/
if (unlikely(prof_on == SLEEP_PROFILING)) {
profile_hits(SLEEP_PROFILING,
(void *)get_wchan(tsk),
delta >> 20);
}
account_scheduler_latency(tsk, delta >> 10, 0);
}
}
}
/*
* Task is being enqueued - update stats:
*/
sched/debug: Make schedstats a runtime tunable that is disabled by default schedstats is very useful during debugging and performance tuning but it incurs overhead to calculate the stats. As such, even though it can be disabled at build time, it is often enabled as the information is useful. This patch adds a kernel command-line and sysctl tunable to enable or disable schedstats on demand (when it's built in). It is disabled by default as someone who knows they need it can also learn to enable it when necessary. The benefits are dependent on how scheduler-intensive the workload is. If it is then the patch reduces the number of cycles spent calculating the stats with a small benefit from reducing the cache footprint of the scheduler. These measurements were taken from a 48-core 2-socket machine with Xeon(R) E5-2670 v3 cpus although they were also tested on a single socket machine 8-core machine with Intel i7-3770 processors. netperf-tcp 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v3r1 Hmean 64 560.45 ( 0.00%) 575.98 ( 2.77%) Hmean 128 766.66 ( 0.00%) 795.79 ( 3.80%) Hmean 256 950.51 ( 0.00%) 981.50 ( 3.26%) Hmean 1024 1433.25 ( 0.00%) 1466.51 ( 2.32%) Hmean 2048 2810.54 ( 0.00%) 2879.75 ( 2.46%) Hmean 3312 4618.18 ( 0.00%) 4682.09 ( 1.38%) Hmean 4096 5306.42 ( 0.00%) 5346.39 ( 0.75%) Hmean 8192 10581.44 ( 0.00%) 10698.15 ( 1.10%) Hmean 16384 18857.70 ( 0.00%) 18937.61 ( 0.42%) Small gains here, UDP_STREAM showed nothing intresting and neither did the TCP_RR tests. The gains on the 8-core machine were very similar. tbench4 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v3r1 Hmean mb/sec-1 500.85 ( 0.00%) 522.43 ( 4.31%) Hmean mb/sec-2 984.66 ( 0.00%) 1018.19 ( 3.41%) Hmean mb/sec-4 1827.91 ( 0.00%) 1847.78 ( 1.09%) Hmean mb/sec-8 3561.36 ( 0.00%) 3611.28 ( 1.40%) Hmean mb/sec-16 5824.52 ( 0.00%) 5929.03 ( 1.79%) Hmean mb/sec-32 10943.10 ( 0.00%) 10802.83 ( -1.28%) Hmean mb/sec-64 15950.81 ( 0.00%) 16211.31 ( 1.63%) Hmean mb/sec-128 15302.17 ( 0.00%) 15445.11 ( 0.93%) Hmean mb/sec-256 14866.18 ( 0.00%) 15088.73 ( 1.50%) Hmean mb/sec-512 15223.31 ( 0.00%) 15373.69 ( 0.99%) Hmean mb/sec-1024 14574.25 ( 0.00%) 14598.02 ( 0.16%) Hmean mb/sec-2048 13569.02 ( 0.00%) 13733.86 ( 1.21%) Hmean mb/sec-3072 12865.98 ( 0.00%) 13209.23 ( 2.67%) Small gains of 2-4% at low thread counts and otherwise flat. The gains on the 8-core machine were slightly different tbench4 on 8-core i7-3770 single socket machine Hmean mb/sec-1 442.59 ( 0.00%) 448.73 ( 1.39%) Hmean mb/sec-2 796.68 ( 0.00%) 794.39 ( -0.29%) Hmean mb/sec-4 1322.52 ( 0.00%) 1343.66 ( 1.60%) Hmean mb/sec-8 2611.65 ( 0.00%) 2694.86 ( 3.19%) Hmean mb/sec-16 2537.07 ( 0.00%) 2609.34 ( 2.85%) Hmean mb/sec-32 2506.02 ( 0.00%) 2578.18 ( 2.88%) Hmean mb/sec-64 2511.06 ( 0.00%) 2569.16 ( 2.31%) Hmean mb/sec-128 2313.38 ( 0.00%) 2395.50 ( 3.55%) Hmean mb/sec-256 2110.04 ( 0.00%) 2177.45 ( 3.19%) Hmean mb/sec-512 2072.51 ( 0.00%) 2053.97 ( -0.89%) In constract, this shows a relatively steady 2-3% gain at higher thread counts. Due to the nature of the patch and the type of workload, it's not a surprise that the result will depend on the CPU used. hackbench-pipes 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v3r1 Amean 1 0.0637 ( 0.00%) 0.0660 ( -3.59%) Amean 4 0.1229 ( 0.00%) 0.1181 ( 3.84%) Amean 7 0.1921 ( 0.00%) 0.1911 ( 0.52%) Amean 12 0.3117 ( 0.00%) 0.2923 ( 6.23%) Amean 21 0.4050 ( 0.00%) 0.3899 ( 3.74%) Amean 30 0.4586 ( 0.00%) 0.4433 ( 3.33%) Amean 48 0.5910 ( 0.00%) 0.5694 ( 3.65%) Amean 79 0.8663 ( 0.00%) 0.8626 ( 0.43%) Amean 110 1.1543 ( 0.00%) 1.1517 ( 0.22%) Amean 141 1.4457 ( 0.00%) 1.4290 ( 1.16%) Amean 172 1.7090 ( 0.00%) 1.6924 ( 0.97%) Amean 192 1.9126 ( 0.00%) 1.9089 ( 0.19%) Some small gains and losses and while the variance data is not included, it's close to the noise. The UMA machine did not show anything particularly different pipetest 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v2r2 Min Time 4.13 ( 0.00%) 3.99 ( 3.39%) 1st-qrtle Time 4.38 ( 0.00%) 4.27 ( 2.51%) 2nd-qrtle Time 4.46 ( 0.00%) 4.39 ( 1.57%) 3rd-qrtle Time 4.56 ( 0.00%) 4.51 ( 1.10%) Max-90% Time 4.67 ( 0.00%) 4.60 ( 1.50%) Max-93% Time 4.71 ( 0.00%) 4.65 ( 1.27%) Max-95% Time 4.74 ( 0.00%) 4.71 ( 0.63%) Max-99% Time 4.88 ( 0.00%) 4.79 ( 1.84%) Max Time 4.93 ( 0.00%) 4.83 ( 2.03%) Mean Time 4.48 ( 0.00%) 4.39 ( 1.91%) Best99%Mean Time 4.47 ( 0.00%) 4.39 ( 1.91%) Best95%Mean Time 4.46 ( 0.00%) 4.38 ( 1.93%) Best90%Mean Time 4.45 ( 0.00%) 4.36 ( 1.98%) Best50%Mean Time 4.36 ( 0.00%) 4.25 ( 2.49%) Best10%Mean Time 4.23 ( 0.00%) 4.10 ( 3.13%) Best5%Mean Time 4.19 ( 0.00%) 4.06 ( 3.20%) Best1%Mean Time 4.13 ( 0.00%) 4.00 ( 3.39%) Small improvement and similar gains were seen on the UMA machine. The gain is small but it stands to reason that doing less work in the scheduler is a good thing. The downside is that the lack of schedstats and tracepoints may be surprising to experts doing performance analysis until they find the existence of the schedstats= parameter or schedstats sysctl. It will be automatically activated for latencytop and sleep profiling to alleviate the problem. For tracepoints, there is a simple warning as it's not safe to activate schedstats in the context when it's known the tracepoint may be wanted but is unavailable. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Reviewed-by: Matt Fleming <matt@codeblueprint.co.uk> Reviewed-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <mgalbraith@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1454663316-22048-1-git-send-email-mgorman@techsingularity.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-02-05 09:08:36 +00:00
static inline void
update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
{
if (!schedstat_enabled())
return;
/*
* Are we enqueueing a waiting task? (for current tasks
* a dequeue/enqueue event is a NOP)
*/
if (se != cfs_rq->curr)
update_stats_wait_start(cfs_rq, se);
if (flags & ENQUEUE_WAKEUP)
update_stats_enqueue_sleeper(cfs_rq, se);
}
static inline void
sched/debug: Make schedstats a runtime tunable that is disabled by default schedstats is very useful during debugging and performance tuning but it incurs overhead to calculate the stats. As such, even though it can be disabled at build time, it is often enabled as the information is useful. This patch adds a kernel command-line and sysctl tunable to enable or disable schedstats on demand (when it's built in). It is disabled by default as someone who knows they need it can also learn to enable it when necessary. The benefits are dependent on how scheduler-intensive the workload is. If it is then the patch reduces the number of cycles spent calculating the stats with a small benefit from reducing the cache footprint of the scheduler. These measurements were taken from a 48-core 2-socket machine with Xeon(R) E5-2670 v3 cpus although they were also tested on a single socket machine 8-core machine with Intel i7-3770 processors. netperf-tcp 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v3r1 Hmean 64 560.45 ( 0.00%) 575.98 ( 2.77%) Hmean 128 766.66 ( 0.00%) 795.79 ( 3.80%) Hmean 256 950.51 ( 0.00%) 981.50 ( 3.26%) Hmean 1024 1433.25 ( 0.00%) 1466.51 ( 2.32%) Hmean 2048 2810.54 ( 0.00%) 2879.75 ( 2.46%) Hmean 3312 4618.18 ( 0.00%) 4682.09 ( 1.38%) Hmean 4096 5306.42 ( 0.00%) 5346.39 ( 0.75%) Hmean 8192 10581.44 ( 0.00%) 10698.15 ( 1.10%) Hmean 16384 18857.70 ( 0.00%) 18937.61 ( 0.42%) Small gains here, UDP_STREAM showed nothing intresting and neither did the TCP_RR tests. The gains on the 8-core machine were very similar. tbench4 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v3r1 Hmean mb/sec-1 500.85 ( 0.00%) 522.43 ( 4.31%) Hmean mb/sec-2 984.66 ( 0.00%) 1018.19 ( 3.41%) Hmean mb/sec-4 1827.91 ( 0.00%) 1847.78 ( 1.09%) Hmean mb/sec-8 3561.36 ( 0.00%) 3611.28 ( 1.40%) Hmean mb/sec-16 5824.52 ( 0.00%) 5929.03 ( 1.79%) Hmean mb/sec-32 10943.10 ( 0.00%) 10802.83 ( -1.28%) Hmean mb/sec-64 15950.81 ( 0.00%) 16211.31 ( 1.63%) Hmean mb/sec-128 15302.17 ( 0.00%) 15445.11 ( 0.93%) Hmean mb/sec-256 14866.18 ( 0.00%) 15088.73 ( 1.50%) Hmean mb/sec-512 15223.31 ( 0.00%) 15373.69 ( 0.99%) Hmean mb/sec-1024 14574.25 ( 0.00%) 14598.02 ( 0.16%) Hmean mb/sec-2048 13569.02 ( 0.00%) 13733.86 ( 1.21%) Hmean mb/sec-3072 12865.98 ( 0.00%) 13209.23 ( 2.67%) Small gains of 2-4% at low thread counts and otherwise flat. The gains on the 8-core machine were slightly different tbench4 on 8-core i7-3770 single socket machine Hmean mb/sec-1 442.59 ( 0.00%) 448.73 ( 1.39%) Hmean mb/sec-2 796.68 ( 0.00%) 794.39 ( -0.29%) Hmean mb/sec-4 1322.52 ( 0.00%) 1343.66 ( 1.60%) Hmean mb/sec-8 2611.65 ( 0.00%) 2694.86 ( 3.19%) Hmean mb/sec-16 2537.07 ( 0.00%) 2609.34 ( 2.85%) Hmean mb/sec-32 2506.02 ( 0.00%) 2578.18 ( 2.88%) Hmean mb/sec-64 2511.06 ( 0.00%) 2569.16 ( 2.31%) Hmean mb/sec-128 2313.38 ( 0.00%) 2395.50 ( 3.55%) Hmean mb/sec-256 2110.04 ( 0.00%) 2177.45 ( 3.19%) Hmean mb/sec-512 2072.51 ( 0.00%) 2053.97 ( -0.89%) In constract, this shows a relatively steady 2-3% gain at higher thread counts. Due to the nature of the patch and the type of workload, it's not a surprise that the result will depend on the CPU used. hackbench-pipes 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v3r1 Amean 1 0.0637 ( 0.00%) 0.0660 ( -3.59%) Amean 4 0.1229 ( 0.00%) 0.1181 ( 3.84%) Amean 7 0.1921 ( 0.00%) 0.1911 ( 0.52%) Amean 12 0.3117 ( 0.00%) 0.2923 ( 6.23%) Amean 21 0.4050 ( 0.00%) 0.3899 ( 3.74%) Amean 30 0.4586 ( 0.00%) 0.4433 ( 3.33%) Amean 48 0.5910 ( 0.00%) 0.5694 ( 3.65%) Amean 79 0.8663 ( 0.00%) 0.8626 ( 0.43%) Amean 110 1.1543 ( 0.00%) 1.1517 ( 0.22%) Amean 141 1.4457 ( 0.00%) 1.4290 ( 1.16%) Amean 172 1.7090 ( 0.00%) 1.6924 ( 0.97%) Amean 192 1.9126 ( 0.00%) 1.9089 ( 0.19%) Some small gains and losses and while the variance data is not included, it's close to the noise. The UMA machine did not show anything particularly different pipetest 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v2r2 Min Time 4.13 ( 0.00%) 3.99 ( 3.39%) 1st-qrtle Time 4.38 ( 0.00%) 4.27 ( 2.51%) 2nd-qrtle Time 4.46 ( 0.00%) 4.39 ( 1.57%) 3rd-qrtle Time 4.56 ( 0.00%) 4.51 ( 1.10%) Max-90% Time 4.67 ( 0.00%) 4.60 ( 1.50%) Max-93% Time 4.71 ( 0.00%) 4.65 ( 1.27%) Max-95% Time 4.74 ( 0.00%) 4.71 ( 0.63%) Max-99% Time 4.88 ( 0.00%) 4.79 ( 1.84%) Max Time 4.93 ( 0.00%) 4.83 ( 2.03%) Mean Time 4.48 ( 0.00%) 4.39 ( 1.91%) Best99%Mean Time 4.47 ( 0.00%) 4.39 ( 1.91%) Best95%Mean Time 4.46 ( 0.00%) 4.38 ( 1.93%) Best90%Mean Time 4.45 ( 0.00%) 4.36 ( 1.98%) Best50%Mean Time 4.36 ( 0.00%) 4.25 ( 2.49%) Best10%Mean Time 4.23 ( 0.00%) 4.10 ( 3.13%) Best5%Mean Time 4.19 ( 0.00%) 4.06 ( 3.20%) Best1%Mean Time 4.13 ( 0.00%) 4.00 ( 3.39%) Small improvement and similar gains were seen on the UMA machine. The gain is small but it stands to reason that doing less work in the scheduler is a good thing. The downside is that the lack of schedstats and tracepoints may be surprising to experts doing performance analysis until they find the existence of the schedstats= parameter or schedstats sysctl. It will be automatically activated for latencytop and sleep profiling to alleviate the problem. For tracepoints, there is a simple warning as it's not safe to activate schedstats in the context when it's known the tracepoint may be wanted but is unavailable. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Reviewed-by: Matt Fleming <matt@codeblueprint.co.uk> Reviewed-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <mgalbraith@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1454663316-22048-1-git-send-email-mgorman@techsingularity.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-02-05 09:08:36 +00:00
update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
{
if (!schedstat_enabled())
return;
/*
* Mark the end of the wait period if dequeueing a
* waiting task:
*/
if (se != cfs_rq->curr)
update_stats_wait_end(cfs_rq, se);
sched/debug: Make schedstats a runtime tunable that is disabled by default schedstats is very useful during debugging and performance tuning but it incurs overhead to calculate the stats. As such, even though it can be disabled at build time, it is often enabled as the information is useful. This patch adds a kernel command-line and sysctl tunable to enable or disable schedstats on demand (when it's built in). It is disabled by default as someone who knows they need it can also learn to enable it when necessary. The benefits are dependent on how scheduler-intensive the workload is. If it is then the patch reduces the number of cycles spent calculating the stats with a small benefit from reducing the cache footprint of the scheduler. These measurements were taken from a 48-core 2-socket machine with Xeon(R) E5-2670 v3 cpus although they were also tested on a single socket machine 8-core machine with Intel i7-3770 processors. netperf-tcp 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v3r1 Hmean 64 560.45 ( 0.00%) 575.98 ( 2.77%) Hmean 128 766.66 ( 0.00%) 795.79 ( 3.80%) Hmean 256 950.51 ( 0.00%) 981.50 ( 3.26%) Hmean 1024 1433.25 ( 0.00%) 1466.51 ( 2.32%) Hmean 2048 2810.54 ( 0.00%) 2879.75 ( 2.46%) Hmean 3312 4618.18 ( 0.00%) 4682.09 ( 1.38%) Hmean 4096 5306.42 ( 0.00%) 5346.39 ( 0.75%) Hmean 8192 10581.44 ( 0.00%) 10698.15 ( 1.10%) Hmean 16384 18857.70 ( 0.00%) 18937.61 ( 0.42%) Small gains here, UDP_STREAM showed nothing intresting and neither did the TCP_RR tests. The gains on the 8-core machine were very similar. tbench4 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v3r1 Hmean mb/sec-1 500.85 ( 0.00%) 522.43 ( 4.31%) Hmean mb/sec-2 984.66 ( 0.00%) 1018.19 ( 3.41%) Hmean mb/sec-4 1827.91 ( 0.00%) 1847.78 ( 1.09%) Hmean mb/sec-8 3561.36 ( 0.00%) 3611.28 ( 1.40%) Hmean mb/sec-16 5824.52 ( 0.00%) 5929.03 ( 1.79%) Hmean mb/sec-32 10943.10 ( 0.00%) 10802.83 ( -1.28%) Hmean mb/sec-64 15950.81 ( 0.00%) 16211.31 ( 1.63%) Hmean mb/sec-128 15302.17 ( 0.00%) 15445.11 ( 0.93%) Hmean mb/sec-256 14866.18 ( 0.00%) 15088.73 ( 1.50%) Hmean mb/sec-512 15223.31 ( 0.00%) 15373.69 ( 0.99%) Hmean mb/sec-1024 14574.25 ( 0.00%) 14598.02 ( 0.16%) Hmean mb/sec-2048 13569.02 ( 0.00%) 13733.86 ( 1.21%) Hmean mb/sec-3072 12865.98 ( 0.00%) 13209.23 ( 2.67%) Small gains of 2-4% at low thread counts and otherwise flat. The gains on the 8-core machine were slightly different tbench4 on 8-core i7-3770 single socket machine Hmean mb/sec-1 442.59 ( 0.00%) 448.73 ( 1.39%) Hmean mb/sec-2 796.68 ( 0.00%) 794.39 ( -0.29%) Hmean mb/sec-4 1322.52 ( 0.00%) 1343.66 ( 1.60%) Hmean mb/sec-8 2611.65 ( 0.00%) 2694.86 ( 3.19%) Hmean mb/sec-16 2537.07 ( 0.00%) 2609.34 ( 2.85%) Hmean mb/sec-32 2506.02 ( 0.00%) 2578.18 ( 2.88%) Hmean mb/sec-64 2511.06 ( 0.00%) 2569.16 ( 2.31%) Hmean mb/sec-128 2313.38 ( 0.00%) 2395.50 ( 3.55%) Hmean mb/sec-256 2110.04 ( 0.00%) 2177.45 ( 3.19%) Hmean mb/sec-512 2072.51 ( 0.00%) 2053.97 ( -0.89%) In constract, this shows a relatively steady 2-3% gain at higher thread counts. Due to the nature of the patch and the type of workload, it's not a surprise that the result will depend on the CPU used. hackbench-pipes 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v3r1 Amean 1 0.0637 ( 0.00%) 0.0660 ( -3.59%) Amean 4 0.1229 ( 0.00%) 0.1181 ( 3.84%) Amean 7 0.1921 ( 0.00%) 0.1911 ( 0.52%) Amean 12 0.3117 ( 0.00%) 0.2923 ( 6.23%) Amean 21 0.4050 ( 0.00%) 0.3899 ( 3.74%) Amean 30 0.4586 ( 0.00%) 0.4433 ( 3.33%) Amean 48 0.5910 ( 0.00%) 0.5694 ( 3.65%) Amean 79 0.8663 ( 0.00%) 0.8626 ( 0.43%) Amean 110 1.1543 ( 0.00%) 1.1517 ( 0.22%) Amean 141 1.4457 ( 0.00%) 1.4290 ( 1.16%) Amean 172 1.7090 ( 0.00%) 1.6924 ( 0.97%) Amean 192 1.9126 ( 0.00%) 1.9089 ( 0.19%) Some small gains and losses and while the variance data is not included, it's close to the noise. The UMA machine did not show anything particularly different pipetest 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v2r2 Min Time 4.13 ( 0.00%) 3.99 ( 3.39%) 1st-qrtle Time 4.38 ( 0.00%) 4.27 ( 2.51%) 2nd-qrtle Time 4.46 ( 0.00%) 4.39 ( 1.57%) 3rd-qrtle Time 4.56 ( 0.00%) 4.51 ( 1.10%) Max-90% Time 4.67 ( 0.00%) 4.60 ( 1.50%) Max-93% Time 4.71 ( 0.00%) 4.65 ( 1.27%) Max-95% Time 4.74 ( 0.00%) 4.71 ( 0.63%) Max-99% Time 4.88 ( 0.00%) 4.79 ( 1.84%) Max Time 4.93 ( 0.00%) 4.83 ( 2.03%) Mean Time 4.48 ( 0.00%) 4.39 ( 1.91%) Best99%Mean Time 4.47 ( 0.00%) 4.39 ( 1.91%) Best95%Mean Time 4.46 ( 0.00%) 4.38 ( 1.93%) Best90%Mean Time 4.45 ( 0.00%) 4.36 ( 1.98%) Best50%Mean Time 4.36 ( 0.00%) 4.25 ( 2.49%) Best10%Mean Time 4.23 ( 0.00%) 4.10 ( 3.13%) Best5%Mean Time 4.19 ( 0.00%) 4.06 ( 3.20%) Best1%Mean Time 4.13 ( 0.00%) 4.00 ( 3.39%) Small improvement and similar gains were seen on the UMA machine. The gain is small but it stands to reason that doing less work in the scheduler is a good thing. The downside is that the lack of schedstats and tracepoints may be surprising to experts doing performance analysis until they find the existence of the schedstats= parameter or schedstats sysctl. It will be automatically activated for latencytop and sleep profiling to alleviate the problem. For tracepoints, there is a simple warning as it's not safe to activate schedstats in the context when it's known the tracepoint may be wanted but is unavailable. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Reviewed-by: Matt Fleming <matt@codeblueprint.co.uk> Reviewed-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <mgalbraith@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1454663316-22048-1-git-send-email-mgorman@techsingularity.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-02-05 09:08:36 +00:00
if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
struct task_struct *tsk = task_of(se);
sched/debug: Make schedstats a runtime tunable that is disabled by default schedstats is very useful during debugging and performance tuning but it incurs overhead to calculate the stats. As such, even though it can be disabled at build time, it is often enabled as the information is useful. This patch adds a kernel command-line and sysctl tunable to enable or disable schedstats on demand (when it's built in). It is disabled by default as someone who knows they need it can also learn to enable it when necessary. The benefits are dependent on how scheduler-intensive the workload is. If it is then the patch reduces the number of cycles spent calculating the stats with a small benefit from reducing the cache footprint of the scheduler. These measurements were taken from a 48-core 2-socket machine with Xeon(R) E5-2670 v3 cpus although they were also tested on a single socket machine 8-core machine with Intel i7-3770 processors. netperf-tcp 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v3r1 Hmean 64 560.45 ( 0.00%) 575.98 ( 2.77%) Hmean 128 766.66 ( 0.00%) 795.79 ( 3.80%) Hmean 256 950.51 ( 0.00%) 981.50 ( 3.26%) Hmean 1024 1433.25 ( 0.00%) 1466.51 ( 2.32%) Hmean 2048 2810.54 ( 0.00%) 2879.75 ( 2.46%) Hmean 3312 4618.18 ( 0.00%) 4682.09 ( 1.38%) Hmean 4096 5306.42 ( 0.00%) 5346.39 ( 0.75%) Hmean 8192 10581.44 ( 0.00%) 10698.15 ( 1.10%) Hmean 16384 18857.70 ( 0.00%) 18937.61 ( 0.42%) Small gains here, UDP_STREAM showed nothing intresting and neither did the TCP_RR tests. The gains on the 8-core machine were very similar. tbench4 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v3r1 Hmean mb/sec-1 500.85 ( 0.00%) 522.43 ( 4.31%) Hmean mb/sec-2 984.66 ( 0.00%) 1018.19 ( 3.41%) Hmean mb/sec-4 1827.91 ( 0.00%) 1847.78 ( 1.09%) Hmean mb/sec-8 3561.36 ( 0.00%) 3611.28 ( 1.40%) Hmean mb/sec-16 5824.52 ( 0.00%) 5929.03 ( 1.79%) Hmean mb/sec-32 10943.10 ( 0.00%) 10802.83 ( -1.28%) Hmean mb/sec-64 15950.81 ( 0.00%) 16211.31 ( 1.63%) Hmean mb/sec-128 15302.17 ( 0.00%) 15445.11 ( 0.93%) Hmean mb/sec-256 14866.18 ( 0.00%) 15088.73 ( 1.50%) Hmean mb/sec-512 15223.31 ( 0.00%) 15373.69 ( 0.99%) Hmean mb/sec-1024 14574.25 ( 0.00%) 14598.02 ( 0.16%) Hmean mb/sec-2048 13569.02 ( 0.00%) 13733.86 ( 1.21%) Hmean mb/sec-3072 12865.98 ( 0.00%) 13209.23 ( 2.67%) Small gains of 2-4% at low thread counts and otherwise flat. The gains on the 8-core machine were slightly different tbench4 on 8-core i7-3770 single socket machine Hmean mb/sec-1 442.59 ( 0.00%) 448.73 ( 1.39%) Hmean mb/sec-2 796.68 ( 0.00%) 794.39 ( -0.29%) Hmean mb/sec-4 1322.52 ( 0.00%) 1343.66 ( 1.60%) Hmean mb/sec-8 2611.65 ( 0.00%) 2694.86 ( 3.19%) Hmean mb/sec-16 2537.07 ( 0.00%) 2609.34 ( 2.85%) Hmean mb/sec-32 2506.02 ( 0.00%) 2578.18 ( 2.88%) Hmean mb/sec-64 2511.06 ( 0.00%) 2569.16 ( 2.31%) Hmean mb/sec-128 2313.38 ( 0.00%) 2395.50 ( 3.55%) Hmean mb/sec-256 2110.04 ( 0.00%) 2177.45 ( 3.19%) Hmean mb/sec-512 2072.51 ( 0.00%) 2053.97 ( -0.89%) In constract, this shows a relatively steady 2-3% gain at higher thread counts. Due to the nature of the patch and the type of workload, it's not a surprise that the result will depend on the CPU used. hackbench-pipes 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v3r1 Amean 1 0.0637 ( 0.00%) 0.0660 ( -3.59%) Amean 4 0.1229 ( 0.00%) 0.1181 ( 3.84%) Amean 7 0.1921 ( 0.00%) 0.1911 ( 0.52%) Amean 12 0.3117 ( 0.00%) 0.2923 ( 6.23%) Amean 21 0.4050 ( 0.00%) 0.3899 ( 3.74%) Amean 30 0.4586 ( 0.00%) 0.4433 ( 3.33%) Amean 48 0.5910 ( 0.00%) 0.5694 ( 3.65%) Amean 79 0.8663 ( 0.00%) 0.8626 ( 0.43%) Amean 110 1.1543 ( 0.00%) 1.1517 ( 0.22%) Amean 141 1.4457 ( 0.00%) 1.4290 ( 1.16%) Amean 172 1.7090 ( 0.00%) 1.6924 ( 0.97%) Amean 192 1.9126 ( 0.00%) 1.9089 ( 0.19%) Some small gains and losses and while the variance data is not included, it's close to the noise. The UMA machine did not show anything particularly different pipetest 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v2r2 Min Time 4.13 ( 0.00%) 3.99 ( 3.39%) 1st-qrtle Time 4.38 ( 0.00%) 4.27 ( 2.51%) 2nd-qrtle Time 4.46 ( 0.00%) 4.39 ( 1.57%) 3rd-qrtle Time 4.56 ( 0.00%) 4.51 ( 1.10%) Max-90% Time 4.67 ( 0.00%) 4.60 ( 1.50%) Max-93% Time 4.71 ( 0.00%) 4.65 ( 1.27%) Max-95% Time 4.74 ( 0.00%) 4.71 ( 0.63%) Max-99% Time 4.88 ( 0.00%) 4.79 ( 1.84%) Max Time 4.93 ( 0.00%) 4.83 ( 2.03%) Mean Time 4.48 ( 0.00%) 4.39 ( 1.91%) Best99%Mean Time 4.47 ( 0.00%) 4.39 ( 1.91%) Best95%Mean Time 4.46 ( 0.00%) 4.38 ( 1.93%) Best90%Mean Time 4.45 ( 0.00%) 4.36 ( 1.98%) Best50%Mean Time 4.36 ( 0.00%) 4.25 ( 2.49%) Best10%Mean Time 4.23 ( 0.00%) 4.10 ( 3.13%) Best5%Mean Time 4.19 ( 0.00%) 4.06 ( 3.20%) Best1%Mean Time 4.13 ( 0.00%) 4.00 ( 3.39%) Small improvement and similar gains were seen on the UMA machine. The gain is small but it stands to reason that doing less work in the scheduler is a good thing. The downside is that the lack of schedstats and tracepoints may be surprising to experts doing performance analysis until they find the existence of the schedstats= parameter or schedstats sysctl. It will be automatically activated for latencytop and sleep profiling to alleviate the problem. For tracepoints, there is a simple warning as it's not safe to activate schedstats in the context when it's known the tracepoint may be wanted but is unavailable. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Reviewed-by: Matt Fleming <matt@codeblueprint.co.uk> Reviewed-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <mgalbraith@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1454663316-22048-1-git-send-email-mgorman@techsingularity.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-02-05 09:08:36 +00:00
if (tsk->state & TASK_INTERRUPTIBLE)
__schedstat_set(se->statistics.sleep_start,
rq_clock(rq_of(cfs_rq)));
if (tsk->state & TASK_UNINTERRUPTIBLE)
__schedstat_set(se->statistics.block_start,
rq_clock(rq_of(cfs_rq)));
sched/debug: Make schedstats a runtime tunable that is disabled by default schedstats is very useful during debugging and performance tuning but it incurs overhead to calculate the stats. As such, even though it can be disabled at build time, it is often enabled as the information is useful. This patch adds a kernel command-line and sysctl tunable to enable or disable schedstats on demand (when it's built in). It is disabled by default as someone who knows they need it can also learn to enable it when necessary. The benefits are dependent on how scheduler-intensive the workload is. If it is then the patch reduces the number of cycles spent calculating the stats with a small benefit from reducing the cache footprint of the scheduler. These measurements were taken from a 48-core 2-socket machine with Xeon(R) E5-2670 v3 cpus although they were also tested on a single socket machine 8-core machine with Intel i7-3770 processors. netperf-tcp 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v3r1 Hmean 64 560.45 ( 0.00%) 575.98 ( 2.77%) Hmean 128 766.66 ( 0.00%) 795.79 ( 3.80%) Hmean 256 950.51 ( 0.00%) 981.50 ( 3.26%) Hmean 1024 1433.25 ( 0.00%) 1466.51 ( 2.32%) Hmean 2048 2810.54 ( 0.00%) 2879.75 ( 2.46%) Hmean 3312 4618.18 ( 0.00%) 4682.09 ( 1.38%) Hmean 4096 5306.42 ( 0.00%) 5346.39 ( 0.75%) Hmean 8192 10581.44 ( 0.00%) 10698.15 ( 1.10%) Hmean 16384 18857.70 ( 0.00%) 18937.61 ( 0.42%) Small gains here, UDP_STREAM showed nothing intresting and neither did the TCP_RR tests. The gains on the 8-core machine were very similar. tbench4 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v3r1 Hmean mb/sec-1 500.85 ( 0.00%) 522.43 ( 4.31%) Hmean mb/sec-2 984.66 ( 0.00%) 1018.19 ( 3.41%) Hmean mb/sec-4 1827.91 ( 0.00%) 1847.78 ( 1.09%) Hmean mb/sec-8 3561.36 ( 0.00%) 3611.28 ( 1.40%) Hmean mb/sec-16 5824.52 ( 0.00%) 5929.03 ( 1.79%) Hmean mb/sec-32 10943.10 ( 0.00%) 10802.83 ( -1.28%) Hmean mb/sec-64 15950.81 ( 0.00%) 16211.31 ( 1.63%) Hmean mb/sec-128 15302.17 ( 0.00%) 15445.11 ( 0.93%) Hmean mb/sec-256 14866.18 ( 0.00%) 15088.73 ( 1.50%) Hmean mb/sec-512 15223.31 ( 0.00%) 15373.69 ( 0.99%) Hmean mb/sec-1024 14574.25 ( 0.00%) 14598.02 ( 0.16%) Hmean mb/sec-2048 13569.02 ( 0.00%) 13733.86 ( 1.21%) Hmean mb/sec-3072 12865.98 ( 0.00%) 13209.23 ( 2.67%) Small gains of 2-4% at low thread counts and otherwise flat. The gains on the 8-core machine were slightly different tbench4 on 8-core i7-3770 single socket machine Hmean mb/sec-1 442.59 ( 0.00%) 448.73 ( 1.39%) Hmean mb/sec-2 796.68 ( 0.00%) 794.39 ( -0.29%) Hmean mb/sec-4 1322.52 ( 0.00%) 1343.66 ( 1.60%) Hmean mb/sec-8 2611.65 ( 0.00%) 2694.86 ( 3.19%) Hmean mb/sec-16 2537.07 ( 0.00%) 2609.34 ( 2.85%) Hmean mb/sec-32 2506.02 ( 0.00%) 2578.18 ( 2.88%) Hmean mb/sec-64 2511.06 ( 0.00%) 2569.16 ( 2.31%) Hmean mb/sec-128 2313.38 ( 0.00%) 2395.50 ( 3.55%) Hmean mb/sec-256 2110.04 ( 0.00%) 2177.45 ( 3.19%) Hmean mb/sec-512 2072.51 ( 0.00%) 2053.97 ( -0.89%) In constract, this shows a relatively steady 2-3% gain at higher thread counts. Due to the nature of the patch and the type of workload, it's not a surprise that the result will depend on the CPU used. hackbench-pipes 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v3r1 Amean 1 0.0637 ( 0.00%) 0.0660 ( -3.59%) Amean 4 0.1229 ( 0.00%) 0.1181 ( 3.84%) Amean 7 0.1921 ( 0.00%) 0.1911 ( 0.52%) Amean 12 0.3117 ( 0.00%) 0.2923 ( 6.23%) Amean 21 0.4050 ( 0.00%) 0.3899 ( 3.74%) Amean 30 0.4586 ( 0.00%) 0.4433 ( 3.33%) Amean 48 0.5910 ( 0.00%) 0.5694 ( 3.65%) Amean 79 0.8663 ( 0.00%) 0.8626 ( 0.43%) Amean 110 1.1543 ( 0.00%) 1.1517 ( 0.22%) Amean 141 1.4457 ( 0.00%) 1.4290 ( 1.16%) Amean 172 1.7090 ( 0.00%) 1.6924 ( 0.97%) Amean 192 1.9126 ( 0.00%) 1.9089 ( 0.19%) Some small gains and losses and while the variance data is not included, it's close to the noise. The UMA machine did not show anything particularly different pipetest 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v2r2 Min Time 4.13 ( 0.00%) 3.99 ( 3.39%) 1st-qrtle Time 4.38 ( 0.00%) 4.27 ( 2.51%) 2nd-qrtle Time 4.46 ( 0.00%) 4.39 ( 1.57%) 3rd-qrtle Time 4.56 ( 0.00%) 4.51 ( 1.10%) Max-90% Time 4.67 ( 0.00%) 4.60 ( 1.50%) Max-93% Time 4.71 ( 0.00%) 4.65 ( 1.27%) Max-95% Time 4.74 ( 0.00%) 4.71 ( 0.63%) Max-99% Time 4.88 ( 0.00%) 4.79 ( 1.84%) Max Time 4.93 ( 0.00%) 4.83 ( 2.03%) Mean Time 4.48 ( 0.00%) 4.39 ( 1.91%) Best99%Mean Time 4.47 ( 0.00%) 4.39 ( 1.91%) Best95%Mean Time 4.46 ( 0.00%) 4.38 ( 1.93%) Best90%Mean Time 4.45 ( 0.00%) 4.36 ( 1.98%) Best50%Mean Time 4.36 ( 0.00%) 4.25 ( 2.49%) Best10%Mean Time 4.23 ( 0.00%) 4.10 ( 3.13%) Best5%Mean Time 4.19 ( 0.00%) 4.06 ( 3.20%) Best1%Mean Time 4.13 ( 0.00%) 4.00 ( 3.39%) Small improvement and similar gains were seen on the UMA machine. The gain is small but it stands to reason that doing less work in the scheduler is a good thing. The downside is that the lack of schedstats and tracepoints may be surprising to experts doing performance analysis until they find the existence of the schedstats= parameter or schedstats sysctl. It will be automatically activated for latencytop and sleep profiling to alleviate the problem. For tracepoints, there is a simple warning as it's not safe to activate schedstats in the context when it's known the tracepoint may be wanted but is unavailable. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Reviewed-by: Matt Fleming <matt@codeblueprint.co.uk> Reviewed-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <mgalbraith@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1454663316-22048-1-git-send-email-mgorman@techsingularity.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-02-05 09:08:36 +00:00
}
}
/*
* We are picking a new current task - update its stats:
*/
static inline void
update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
{
/*
* We are starting a new run period:
*/
se->exec_start = rq_clock_task(rq_of(cfs_rq));
}
/**************************************************
* Scheduling class queueing methods:
*/
#ifdef CONFIG_NUMA_BALANCING
/*
sched/numa: Set the scan rate proportional to the memory usage of the task being scanned The NUMA PTE scan rate is controlled with a combination of the numa_balancing_scan_period_min, numa_balancing_scan_period_max and numa_balancing_scan_size. This scan rate is independent of the size of the task and as an aside it is further complicated by the fact that numa_balancing_scan_size controls how many pages are marked pte_numa and not how much virtual memory is scanned. In combination, it is almost impossible to meaningfully tune the min and max scan periods and reasoning about performance is complex when the time to complete a full scan is is partially a function of the tasks memory size. This patch alters the semantic of the min and max tunables to be about tuning the length time it takes to complete a scan of a tasks occupied virtual address space. Conceptually this is a lot easier to understand. There is a "sanity" check to ensure the scan rate is never extremely fast based on the amount of virtual memory that should be scanned in a second. The default of 2.5G seems arbitrary but it is to have the maximum scan rate after the patch roughly match the maximum scan rate before the patch was applied. On a similar note, numa_scan_period is in milliseconds and not jiffies. Properly placed pages slow the scanning rate but adding 10 jiffies to numa_scan_period means that the rate scanning slows depends on HZ which is confusing. Get rid of the jiffies_to_msec conversion and treat it as ms. Signed-off-by: Mel Gorman <mgorman@suse.de> Reviewed-by: Rik van Riel <riel@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1381141781-10992-18-git-send-email-mgorman@suse.de Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-10-07 10:28:55 +00:00
* Approximate time to scan a full NUMA task in ms. The task scan period is
* calculated based on the tasks virtual memory size and
* numa_balancing_scan_size.
*/
sched/numa: Set the scan rate proportional to the memory usage of the task being scanned The NUMA PTE scan rate is controlled with a combination of the numa_balancing_scan_period_min, numa_balancing_scan_period_max and numa_balancing_scan_size. This scan rate is independent of the size of the task and as an aside it is further complicated by the fact that numa_balancing_scan_size controls how many pages are marked pte_numa and not how much virtual memory is scanned. In combination, it is almost impossible to meaningfully tune the min and max scan periods and reasoning about performance is complex when the time to complete a full scan is is partially a function of the tasks memory size. This patch alters the semantic of the min and max tunables to be about tuning the length time it takes to complete a scan of a tasks occupied virtual address space. Conceptually this is a lot easier to understand. There is a "sanity" check to ensure the scan rate is never extremely fast based on the amount of virtual memory that should be scanned in a second. The default of 2.5G seems arbitrary but it is to have the maximum scan rate after the patch roughly match the maximum scan rate before the patch was applied. On a similar note, numa_scan_period is in milliseconds and not jiffies. Properly placed pages slow the scanning rate but adding 10 jiffies to numa_scan_period means that the rate scanning slows depends on HZ which is confusing. Get rid of the jiffies_to_msec conversion and treat it as ms. Signed-off-by: Mel Gorman <mgorman@suse.de> Reviewed-by: Rik van Riel <riel@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1381141781-10992-18-git-send-email-mgorman@suse.de Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-10-07 10:28:55 +00:00
unsigned int sysctl_numa_balancing_scan_period_min = 1000;
unsigned int sysctl_numa_balancing_scan_period_max = 60000;
mm: sched: numa: Implement constant, per task Working Set Sampling (WSS) rate Previously, to probe the working set of a task, we'd use a very simple and crude method: mark all of its address space PROT_NONE. That method has various (obvious) disadvantages: - it samples the working set at dissimilar rates, giving some tasks a sampling quality advantage over others. - creates performance problems for tasks with very large working sets - over-samples processes with large address spaces but which only very rarely execute Improve that method by keeping a rotating offset into the address space that marks the current position of the scan, and advance it by a constant rate (in a CPU cycles execution proportional manner). If the offset reaches the last mapped address of the mm then it then it starts over at the first address. The per-task nature of the working set sampling functionality in this tree allows such constant rate, per task, execution-weight proportional sampling of the working set, with an adaptive sampling interval/frequency that goes from once per 100ms up to just once per 8 seconds. The current sampling volume is 256 MB per interval. As tasks mature and converge their working set, so does the sampling rate slow down to just a trickle, 256 MB per 8 seconds of CPU time executed. This, beyond being adaptive, also rate-limits rarely executing systems and does not over-sample on overloaded systems. [ In AutoNUMA speak, this patch deals with the effective sampling rate of the 'hinting page fault'. AutoNUMA's scanning is currently rate-limited, but it is also fundamentally single-threaded, executing in the knuma_scand kernel thread, so the limit in AutoNUMA is global and does not scale up with the number of CPUs, nor does it scan tasks in an execution proportional manner. So the idea of rate-limiting the scanning was first implemented in the AutoNUMA tree via a global rate limit. This patch goes beyond that by implementing an execution rate proportional working set sampling rate that is not implemented via a single global scanning daemon. ] [ Dan Carpenter pointed out a possible NULL pointer dereference in the first version of this patch. ] Based-on-idea-by: Andrea Arcangeli <aarcange@redhat.com> Bug-Found-By: Dan Carpenter <dan.carpenter@oracle.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Rik van Riel <riel@redhat.com> [ Wrote changelog and fixed bug. ] Signed-off-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Mel Gorman <mgorman@suse.de> Reviewed-by: Rik van Riel <riel@redhat.com>
2012-10-25 12:16:45 +00:00
/* Portion of address space to scan in MB */
unsigned int sysctl_numa_balancing_scan_size = 256;
mm: sched: numa: Implement slow start for working set sampling Add a 1 second delay before starting to scan the working set of a task and starting to balance it amongst nodes. [ note that before the constant per task WSS sampling rate patch the initial scan would happen much later still, in effect that patch caused this regression. ] The theory is that short-run tasks benefit very little from NUMA placement: they come and go, and they better stick to the node they were started on. As tasks mature and rebalance to other CPUs and nodes, so does their NUMA placement have to change and so does it start to matter more and more. In practice this change fixes an observable kbuild regression: # [ a perf stat --null --repeat 10 test of ten bzImage builds to /dev/shm ] !NUMA: 45.291088843 seconds time elapsed ( +- 0.40% ) 45.154231752 seconds time elapsed ( +- 0.36% ) +NUMA, no slow start: 46.172308123 seconds time elapsed ( +- 0.30% ) 46.343168745 seconds time elapsed ( +- 0.25% ) +NUMA, 1 sec slow start: 45.224189155 seconds time elapsed ( +- 0.25% ) 45.160866532 seconds time elapsed ( +- 0.17% ) and it also fixes an observable perf bench (hackbench) regression: # perf stat --null --repeat 10 perf bench sched messaging -NUMA: -NUMA: 0.246225691 seconds time elapsed ( +- 1.31% ) +NUMA no slow start: 0.252620063 seconds time elapsed ( +- 1.13% ) +NUMA 1sec delay: 0.248076230 seconds time elapsed ( +- 1.35% ) The implementation is simple and straightforward, most of the patch deals with adding the /proc/sys/kernel/numa_balancing_scan_delay_ms tunable knob. Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Rik van Riel <riel@redhat.com> [ Wrote the changelog, ran measurements, tuned the default. ] Signed-off-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Mel Gorman <mgorman@suse.de> Reviewed-by: Rik van Riel <riel@redhat.com>
2012-10-25 12:16:47 +00:00
/* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
unsigned int sysctl_numa_balancing_scan_delay = 1000;
struct numa_group {
sched/fair: Convert numa_group.refcount to refcount_t atomic_t variables are currently used to implement reference counters with the following properties: - counter is initialized to 1 using atomic_set() - a resource is freed upon counter reaching zero - once counter reaches zero, its further increments aren't allowed - counter schema uses basic atomic operations (set, inc, inc_not_zero, dec_and_test, etc.) Such atomic variables should be converted to a newly provided refcount_t type and API that prevents accidental counter overflows and underflows. This is important since overflows and underflows can lead to use-after-free situation and be exploitable. The variable numa_group.refcount is used as pure reference counter. Convert it to refcount_t and fix up the operations. ** Important note for maintainers: Some functions from refcount_t API defined in lib/refcount.c have different memory ordering guarantees than their atomic counterparts. The full comparison can be seen in https://lkml.org/lkml/2017/11/15/57 and it is hopefully soon in state to be merged to the documentation tree. Normally the differences should not matter since refcount_t provides enough guarantees to satisfy the refcounting use cases, but in some rare cases it might matter. Please double check that you don't have some undocumented memory guarantees for this variable usage. For the numa_group.refcount it might make a difference in following places: - get_numa_group(): increment in refcount_inc_not_zero() only guarantees control dependency on success vs. fully ordered atomic counterpart - put_numa_group(): decrement in refcount_dec_and_test() only provides RELEASE ordering and control dependency on success vs. fully ordered atomic counterpart Suggested-by: Kees Cook <keescook@chromium.org> Signed-off-by: Elena Reshetova <elena.reshetova@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: David Windsor <dwindsor@gmail.com> Reviewed-by: Hans Liljestrand <ishkamiel@gmail.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: akpm@linux-foundation.org Cc: viro@zeniv.linux.org.uk Link: https://lkml.kernel.org/r/1547814450-18902-4-git-send-email-elena.reshetova@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-01-18 12:27:28 +00:00
refcount_t refcount;
spinlock_t lock; /* nr_tasks, tasks */
int nr_tasks;
pid_t gid;
int active_nodes;
struct rcu_head rcu;
unsigned long total_faults;
unsigned long max_faults_cpu;
/*
* Faults_cpu is used to decide whether memory should move
* towards the CPU. As a consequence, these stats are weighted
* more by CPU use than by memory faults.
*/
unsigned long *faults_cpu;
unsigned long faults[0];
};
/*
* For functions that can be called in multiple contexts that permit reading
* ->numa_group (see struct task_struct for locking rules).
*/
static struct numa_group *deref_task_numa_group(struct task_struct *p)
{
return rcu_dereference_check(p->numa_group, p == current ||
(lockdep_is_held(&task_rq(p)->lock) && !READ_ONCE(p->on_cpu)));
}
static struct numa_group *deref_curr_numa_group(struct task_struct *p)
{
return rcu_dereference_protected(p->numa_group, p == current);
}
static inline unsigned long group_faults_priv(struct numa_group *ng);
static inline unsigned long group_faults_shared(struct numa_group *ng);
sched/numa: Set the scan rate proportional to the memory usage of the task being scanned The NUMA PTE scan rate is controlled with a combination of the numa_balancing_scan_period_min, numa_balancing_scan_period_max and numa_balancing_scan_size. This scan rate is independent of the size of the task and as an aside it is further complicated by the fact that numa_balancing_scan_size controls how many pages are marked pte_numa and not how much virtual memory is scanned. In combination, it is almost impossible to meaningfully tune the min and max scan periods and reasoning about performance is complex when the time to complete a full scan is is partially a function of the tasks memory size. This patch alters the semantic of the min and max tunables to be about tuning the length time it takes to complete a scan of a tasks occupied virtual address space. Conceptually this is a lot easier to understand. There is a "sanity" check to ensure the scan rate is never extremely fast based on the amount of virtual memory that should be scanned in a second. The default of 2.5G seems arbitrary but it is to have the maximum scan rate after the patch roughly match the maximum scan rate before the patch was applied. On a similar note, numa_scan_period is in milliseconds and not jiffies. Properly placed pages slow the scanning rate but adding 10 jiffies to numa_scan_period means that the rate scanning slows depends on HZ which is confusing. Get rid of the jiffies_to_msec conversion and treat it as ms. Signed-off-by: Mel Gorman <mgorman@suse.de> Reviewed-by: Rik van Riel <riel@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1381141781-10992-18-git-send-email-mgorman@suse.de Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-10-07 10:28:55 +00:00
static unsigned int task_nr_scan_windows(struct task_struct *p)
{
unsigned long rss = 0;
unsigned long nr_scan_pages;
/*
* Calculations based on RSS as non-present and empty pages are skipped
* by the PTE scanner and NUMA hinting faults should be trapped based
* on resident pages
*/
nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
rss = get_mm_rss(p->mm);
if (!rss)
rss = nr_scan_pages;
rss = round_up(rss, nr_scan_pages);
return rss / nr_scan_pages;
}
/* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
#define MAX_SCAN_WINDOW 2560
static unsigned int task_scan_min(struct task_struct *p)
{
unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
sched/numa: Set the scan rate proportional to the memory usage of the task being scanned The NUMA PTE scan rate is controlled with a combination of the numa_balancing_scan_period_min, numa_balancing_scan_period_max and numa_balancing_scan_size. This scan rate is independent of the size of the task and as an aside it is further complicated by the fact that numa_balancing_scan_size controls how many pages are marked pte_numa and not how much virtual memory is scanned. In combination, it is almost impossible to meaningfully tune the min and max scan periods and reasoning about performance is complex when the time to complete a full scan is is partially a function of the tasks memory size. This patch alters the semantic of the min and max tunables to be about tuning the length time it takes to complete a scan of a tasks occupied virtual address space. Conceptually this is a lot easier to understand. There is a "sanity" check to ensure the scan rate is never extremely fast based on the amount of virtual memory that should be scanned in a second. The default of 2.5G seems arbitrary but it is to have the maximum scan rate after the patch roughly match the maximum scan rate before the patch was applied. On a similar note, numa_scan_period is in milliseconds and not jiffies. Properly placed pages slow the scanning rate but adding 10 jiffies to numa_scan_period means that the rate scanning slows depends on HZ which is confusing. Get rid of the jiffies_to_msec conversion and treat it as ms. Signed-off-by: Mel Gorman <mgorman@suse.de> Reviewed-by: Rik van Riel <riel@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1381141781-10992-18-git-send-email-mgorman@suse.de Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-10-07 10:28:55 +00:00
unsigned int scan, floor;
unsigned int windows = 1;
sched/fair: Fix division by zero sysctl_numa_balancing_scan_size File /proc/sys/kernel/numa_balancing_scan_size_mb allows writing of zero. This bash command reproduces problem: $ while :; do echo 0 > /proc/sys/kernel/numa_balancing_scan_size_mb; \ echo 256 > /proc/sys/kernel/numa_balancing_scan_size_mb; done divide error: 0000 [#1] SMP Modules linked in: CPU: 0 PID: 24112 Comm: bash Not tainted 3.17.0+ #8 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Bochs 01/01/2011 task: ffff88013c852600 ti: ffff880037a68000 task.ti: ffff880037a68000 RIP: 0010:[<ffffffff81074191>] [<ffffffff81074191>] task_scan_min+0x21/0x50 RSP: 0000:ffff880037a6bce0 EFLAGS: 00010246 RAX: 0000000000000a00 RBX: 00000000000003e8 RCX: 0000000000000000 RDX: 0000000000000000 RSI: 0000000000000000 RDI: ffff88013c852600 RBP: ffff880037a6bcf0 R08: 0000000000000001 R09: 0000000000015c90 R10: ffff880239bf6c00 R11: 0000000000000016 R12: 0000000000003fff R13: ffff88013c852600 R14: ffffea0008d1b000 R15: 0000000000000003 FS: 00007f12bb048700(0000) GS:ffff88007da00000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 000000008005003b CR2: 0000000001505678 CR3: 0000000234770000 CR4: 00000000000006f0 Stack: ffff88013c852600 0000000000003fff ffff880037a6bd18 ffffffff810741d1 ffff88013c852600 0000000000003fff 000000000002bfff ffff880037a6bda8 ffffffff81077ef7 ffffea0008a56d40 0000000000000001 0000000000000001 Call Trace: [<ffffffff810741d1>] task_scan_max+0x11/0x40 [<ffffffff81077ef7>] task_numa_fault+0x1f7/0xae0 [<ffffffff8115a896>] ? migrate_misplaced_page+0x276/0x300 [<ffffffff81134a4d>] handle_mm_fault+0x62d/0xba0 [<ffffffff8103e2f1>] __do_page_fault+0x191/0x510 [<ffffffff81030122>] ? native_smp_send_reschedule+0x42/0x60 [<ffffffff8106dc00>] ? check_preempt_curr+0x80/0xa0 [<ffffffff8107092c>] ? wake_up_new_task+0x11c/0x1a0 [<ffffffff8104887d>] ? do_fork+0x14d/0x340 [<ffffffff811799bb>] ? get_unused_fd_flags+0x2b/0x30 [<ffffffff811799df>] ? __fd_install+0x1f/0x60 [<ffffffff8103e67c>] do_page_fault+0xc/0x10 [<ffffffff8150d322>] page_fault+0x22/0x30 RIP [<ffffffff81074191>] task_scan_min+0x21/0x50 RSP <ffff880037a6bce0> ---[ end trace 9a826d16936c04de ]--- Also fix race in task_scan_min (it depends on compiler behaviour). Signed-off-by: Kirill Tkhai <ktkhai@parallels.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Aaron Tomlin <atomlin@redhat.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Dario Faggioli <raistlin@linux.it> Cc: David Rientjes <rientjes@google.com> Cc: Jens Axboe <axboe@fb.com> Cc: Kees Cook <keescook@chromium.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Rik van Riel <riel@redhat.com> Link: http://lkml.kernel.org/r/1413455977.24793.78.camel@tkhai Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-10-16 10:39:37 +00:00
if (scan_size < MAX_SCAN_WINDOW)
windows = MAX_SCAN_WINDOW / scan_size;
sched/numa: Set the scan rate proportional to the memory usage of the task being scanned The NUMA PTE scan rate is controlled with a combination of the numa_balancing_scan_period_min, numa_balancing_scan_period_max and numa_balancing_scan_size. This scan rate is independent of the size of the task and as an aside it is further complicated by the fact that numa_balancing_scan_size controls how many pages are marked pte_numa and not how much virtual memory is scanned. In combination, it is almost impossible to meaningfully tune the min and max scan periods and reasoning about performance is complex when the time to complete a full scan is is partially a function of the tasks memory size. This patch alters the semantic of the min and max tunables to be about tuning the length time it takes to complete a scan of a tasks occupied virtual address space. Conceptually this is a lot easier to understand. There is a "sanity" check to ensure the scan rate is never extremely fast based on the amount of virtual memory that should be scanned in a second. The default of 2.5G seems arbitrary but it is to have the maximum scan rate after the patch roughly match the maximum scan rate before the patch was applied. On a similar note, numa_scan_period is in milliseconds and not jiffies. Properly placed pages slow the scanning rate but adding 10 jiffies to numa_scan_period means that the rate scanning slows depends on HZ which is confusing. Get rid of the jiffies_to_msec conversion and treat it as ms. Signed-off-by: Mel Gorman <mgorman@suse.de> Reviewed-by: Rik van Riel <riel@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1381141781-10992-18-git-send-email-mgorman@suse.de Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-10-07 10:28:55 +00:00
floor = 1000 / windows;
scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
return max_t(unsigned int, floor, scan);
}
static unsigned int task_scan_start(struct task_struct *p)
{
unsigned long smin = task_scan_min(p);
unsigned long period = smin;
struct numa_group *ng;
/* Scale the maximum scan period with the amount of shared memory. */
rcu_read_lock();
ng = rcu_dereference(p->numa_group);
if (ng) {
unsigned long shared = group_faults_shared(ng);
unsigned long private = group_faults_priv(ng);
sched/fair: Convert numa_group.refcount to refcount_t atomic_t variables are currently used to implement reference counters with the following properties: - counter is initialized to 1 using atomic_set() - a resource is freed upon counter reaching zero - once counter reaches zero, its further increments aren't allowed - counter schema uses basic atomic operations (set, inc, inc_not_zero, dec_and_test, etc.) Such atomic variables should be converted to a newly provided refcount_t type and API that prevents accidental counter overflows and underflows. This is important since overflows and underflows can lead to use-after-free situation and be exploitable. The variable numa_group.refcount is used as pure reference counter. Convert it to refcount_t and fix up the operations. ** Important note for maintainers: Some functions from refcount_t API defined in lib/refcount.c have different memory ordering guarantees than their atomic counterparts. The full comparison can be seen in https://lkml.org/lkml/2017/11/15/57 and it is hopefully soon in state to be merged to the documentation tree. Normally the differences should not matter since refcount_t provides enough guarantees to satisfy the refcounting use cases, but in some rare cases it might matter. Please double check that you don't have some undocumented memory guarantees for this variable usage. For the numa_group.refcount it might make a difference in following places: - get_numa_group(): increment in refcount_inc_not_zero() only guarantees control dependency on success vs. fully ordered atomic counterpart - put_numa_group(): decrement in refcount_dec_and_test() only provides RELEASE ordering and control dependency on success vs. fully ordered atomic counterpart Suggested-by: Kees Cook <keescook@chromium.org> Signed-off-by: Elena Reshetova <elena.reshetova@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: David Windsor <dwindsor@gmail.com> Reviewed-by: Hans Liljestrand <ishkamiel@gmail.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: akpm@linux-foundation.org Cc: viro@zeniv.linux.org.uk Link: https://lkml.kernel.org/r/1547814450-18902-4-git-send-email-elena.reshetova@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-01-18 12:27:28 +00:00
period *= refcount_read(&ng->refcount);
period *= shared + 1;
period /= private + shared + 1;
}
rcu_read_unlock();
return max(smin, period);
}
sched/numa: Set the scan rate proportional to the memory usage of the task being scanned The NUMA PTE scan rate is controlled with a combination of the numa_balancing_scan_period_min, numa_balancing_scan_period_max and numa_balancing_scan_size. This scan rate is independent of the size of the task and as an aside it is further complicated by the fact that numa_balancing_scan_size controls how many pages are marked pte_numa and not how much virtual memory is scanned. In combination, it is almost impossible to meaningfully tune the min and max scan periods and reasoning about performance is complex when the time to complete a full scan is is partially a function of the tasks memory size. This patch alters the semantic of the min and max tunables to be about tuning the length time it takes to complete a scan of a tasks occupied virtual address space. Conceptually this is a lot easier to understand. There is a "sanity" check to ensure the scan rate is never extremely fast based on the amount of virtual memory that should be scanned in a second. The default of 2.5G seems arbitrary but it is to have the maximum scan rate after the patch roughly match the maximum scan rate before the patch was applied. On a similar note, numa_scan_period is in milliseconds and not jiffies. Properly placed pages slow the scanning rate but adding 10 jiffies to numa_scan_period means that the rate scanning slows depends on HZ which is confusing. Get rid of the jiffies_to_msec conversion and treat it as ms. Signed-off-by: Mel Gorman <mgorman@suse.de> Reviewed-by: Rik van Riel <riel@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1381141781-10992-18-git-send-email-mgorman@suse.de Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-10-07 10:28:55 +00:00
static unsigned int task_scan_max(struct task_struct *p)
{
unsigned long smin = task_scan_min(p);
unsigned long smax;
struct numa_group *ng;
sched/numa: Set the scan rate proportional to the memory usage of the task being scanned The NUMA PTE scan rate is controlled with a combination of the numa_balancing_scan_period_min, numa_balancing_scan_period_max and numa_balancing_scan_size. This scan rate is independent of the size of the task and as an aside it is further complicated by the fact that numa_balancing_scan_size controls how many pages are marked pte_numa and not how much virtual memory is scanned. In combination, it is almost impossible to meaningfully tune the min and max scan periods and reasoning about performance is complex when the time to complete a full scan is is partially a function of the tasks memory size. This patch alters the semantic of the min and max tunables to be about tuning the length time it takes to complete a scan of a tasks occupied virtual address space. Conceptually this is a lot easier to understand. There is a "sanity" check to ensure the scan rate is never extremely fast based on the amount of virtual memory that should be scanned in a second. The default of 2.5G seems arbitrary but it is to have the maximum scan rate after the patch roughly match the maximum scan rate before the patch was applied. On a similar note, numa_scan_period is in milliseconds and not jiffies. Properly placed pages slow the scanning rate but adding 10 jiffies to numa_scan_period means that the rate scanning slows depends on HZ which is confusing. Get rid of the jiffies_to_msec conversion and treat it as ms. Signed-off-by: Mel Gorman <mgorman@suse.de> Reviewed-by: Rik van Riel <riel@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1381141781-10992-18-git-send-email-mgorman@suse.de Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-10-07 10:28:55 +00:00
/* Watch for min being lower than max due to floor calculations */
smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
/* Scale the maximum scan period with the amount of shared memory. */
ng = deref_curr_numa_group(p);
if (ng) {
unsigned long shared = group_faults_shared(ng);
unsigned long private = group_faults_priv(ng);
unsigned long period = smax;
sched/fair: Convert numa_group.refcount to refcount_t atomic_t variables are currently used to implement reference counters with the following properties: - counter is initialized to 1 using atomic_set() - a resource is freed upon counter reaching zero - once counter reaches zero, its further increments aren't allowed - counter schema uses basic atomic operations (set, inc, inc_not_zero, dec_and_test, etc.) Such atomic variables should be converted to a newly provided refcount_t type and API that prevents accidental counter overflows and underflows. This is important since overflows and underflows can lead to use-after-free situation and be exploitable. The variable numa_group.refcount is used as pure reference counter. Convert it to refcount_t and fix up the operations. ** Important note for maintainers: Some functions from refcount_t API defined in lib/refcount.c have different memory ordering guarantees than their atomic counterparts. The full comparison can be seen in https://lkml.org/lkml/2017/11/15/57 and it is hopefully soon in state to be merged to the documentation tree. Normally the differences should not matter since refcount_t provides enough guarantees to satisfy the refcounting use cases, but in some rare cases it might matter. Please double check that you don't have some undocumented memory guarantees for this variable usage. For the numa_group.refcount it might make a difference in following places: - get_numa_group(): increment in refcount_inc_not_zero() only guarantees control dependency on success vs. fully ordered atomic counterpart - put_numa_group(): decrement in refcount_dec_and_test() only provides RELEASE ordering and control dependency on success vs. fully ordered atomic counterpart Suggested-by: Kees Cook <keescook@chromium.org> Signed-off-by: Elena Reshetova <elena.reshetova@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: David Windsor <dwindsor@gmail.com> Reviewed-by: Hans Liljestrand <ishkamiel@gmail.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: akpm@linux-foundation.org Cc: viro@zeniv.linux.org.uk Link: https://lkml.kernel.org/r/1547814450-18902-4-git-send-email-elena.reshetova@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-01-18 12:27:28 +00:00
period *= refcount_read(&ng->refcount);
period *= shared + 1;
period /= private + shared + 1;
smax = max(smax, period);
}
sched/numa: Set the scan rate proportional to the memory usage of the task being scanned The NUMA PTE scan rate is controlled with a combination of the numa_balancing_scan_period_min, numa_balancing_scan_period_max and numa_balancing_scan_size. This scan rate is independent of the size of the task and as an aside it is further complicated by the fact that numa_balancing_scan_size controls how many pages are marked pte_numa and not how much virtual memory is scanned. In combination, it is almost impossible to meaningfully tune the min and max scan periods and reasoning about performance is complex when the time to complete a full scan is is partially a function of the tasks memory size. This patch alters the semantic of the min and max tunables to be about tuning the length time it takes to complete a scan of a tasks occupied virtual address space. Conceptually this is a lot easier to understand. There is a "sanity" check to ensure the scan rate is never extremely fast based on the amount of virtual memory that should be scanned in a second. The default of 2.5G seems arbitrary but it is to have the maximum scan rate after the patch roughly match the maximum scan rate before the patch was applied. On a similar note, numa_scan_period is in milliseconds and not jiffies. Properly placed pages slow the scanning rate but adding 10 jiffies to numa_scan_period means that the rate scanning slows depends on HZ which is confusing. Get rid of the jiffies_to_msec conversion and treat it as ms. Signed-off-by: Mel Gorman <mgorman@suse.de> Reviewed-by: Rik van Riel <riel@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1381141781-10992-18-git-send-email-mgorman@suse.de Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-10-07 10:28:55 +00:00
return max(smin, smax);
}
static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
{
mm: replace all open encodings for NUMA_NO_NODE Patch series "Replace all open encodings for NUMA_NO_NODE", v3. All these places for replacement were found by running the following grep patterns on the entire kernel code. Please let me know if this might have missed some instances. This might also have replaced some false positives. I will appreciate suggestions, inputs and review. 1. git grep "nid == -1" 2. git grep "node == -1" 3. git grep "nid = -1" 4. git grep "node = -1" This patch (of 2): At present there are multiple places where invalid node number is encoded as -1. Even though implicitly understood it is always better to have macros in there. Replace these open encodings for an invalid node number with the global macro NUMA_NO_NODE. This helps remove NUMA related assumptions like 'invalid node' from various places redirecting them to a common definition. Link: http://lkml.kernel.org/r/1545127933-10711-2-git-send-email-anshuman.khandual@arm.com Signed-off-by: Anshuman Khandual <anshuman.khandual@arm.com> Reviewed-by: David Hildenbrand <david@redhat.com> Acked-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com> [ixgbe] Acked-by: Jens Axboe <axboe@kernel.dk> [mtip32xx] Acked-by: Vinod Koul <vkoul@kernel.org> [dmaengine.c] Acked-by: Michael Ellerman <mpe@ellerman.id.au> [powerpc] Acked-by: Doug Ledford <dledford@redhat.com> [drivers/infiniband] Cc: Joseph Qi <jiangqi903@gmail.com> Cc: Hans Verkuil <hverkuil@xs4all.nl> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:42:58 +00:00
rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE);
rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
}
static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
{
mm: replace all open encodings for NUMA_NO_NODE Patch series "Replace all open encodings for NUMA_NO_NODE", v3. All these places for replacement were found by running the following grep patterns on the entire kernel code. Please let me know if this might have missed some instances. This might also have replaced some false positives. I will appreciate suggestions, inputs and review. 1. git grep "nid == -1" 2. git grep "node == -1" 3. git grep "nid = -1" 4. git grep "node = -1" This patch (of 2): At present there are multiple places where invalid node number is encoded as -1. Even though implicitly understood it is always better to have macros in there. Replace these open encodings for an invalid node number with the global macro NUMA_NO_NODE. This helps remove NUMA related assumptions like 'invalid node' from various places redirecting them to a common definition. Link: http://lkml.kernel.org/r/1545127933-10711-2-git-send-email-anshuman.khandual@arm.com Signed-off-by: Anshuman Khandual <anshuman.khandual@arm.com> Reviewed-by: David Hildenbrand <david@redhat.com> Acked-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com> [ixgbe] Acked-by: Jens Axboe <axboe@kernel.dk> [mtip32xx] Acked-by: Vinod Koul <vkoul@kernel.org> [dmaengine.c] Acked-by: Michael Ellerman <mpe@ellerman.id.au> [powerpc] Acked-by: Doug Ledford <dledford@redhat.com> [drivers/infiniband] Cc: Joseph Qi <jiangqi903@gmail.com> Cc: Hans Verkuil <hverkuil@xs4all.nl> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:42:58 +00:00
rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE);
rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
}
/* Shared or private faults. */
#define NR_NUMA_HINT_FAULT_TYPES 2
/* Memory and CPU locality */
#define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
/* Averaged statistics, and temporary buffers. */
#define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
pid_t task_numa_group_id(struct task_struct *p)
{
struct numa_group *ng;
pid_t gid = 0;
rcu_read_lock();
ng = rcu_dereference(p->numa_group);
if (ng)
gid = ng->gid;
rcu_read_unlock();
return gid;
}
/*
* The averaged statistics, shared & private, memory & CPU,
* occupy the first half of the array. The second half of the
* array is for current counters, which are averaged into the
* first set by task_numa_placement.
*/
static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
{
return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
}
static inline unsigned long task_faults(struct task_struct *p, int nid)
{
if (!p->numa_faults)
return 0;
return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
}
static inline unsigned long group_faults(struct task_struct *p, int nid)
{
struct numa_group *ng = deref_task_numa_group(p);
if (!ng)
return 0;
return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
ng->faults[task_faults_idx(NUMA_MEM, nid, 1)];
}
static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
{
return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
}
static inline unsigned long group_faults_priv(struct numa_group *ng)
{
unsigned long faults = 0;
int node;
for_each_online_node(node) {
faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
}
return faults;
}
static inline unsigned long group_faults_shared(struct numa_group *ng)
{
unsigned long faults = 0;
int node;
for_each_online_node(node) {
faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
}
return faults;
}
sched/numa: Spread memory according to CPU and memory use The pseudo-interleaving in NUMA placement has a fundamental problem: using hard usage thresholds to spread memory equally between nodes can prevent workloads from converging, or keep memory "trapped" on nodes where the workload is barely running any more. In order for workloads to properly converge, the memory migration should not be stopped when nodes reach parity, but instead be distributed according to how heavily memory is used from each node. This way memory migration and task migration reinforce each other, instead of one putting the brakes on the other. Remove the hard thresholds from the pseudo-interleaving code, and instead use a more gradual policy on memory placement. This also seems to improve convergence of workloads that do not run flat out, but sleep in between bursts of activity. We still want to slow down NUMA scanning and migration once a workload has settled on a few actively used nodes, so keep the 3/4 hysteresis in place. Keep track of whether a workload is actively running on multiple nodes, so task_numa_migrate does a full scan of the system for better task placement. In the case of running 3 SPECjbb2005 instances on a 4 node system, this code seems to result in fairer distribution of memory between nodes, with more memory bandwidth for each instance. Signed-off-by: Rik van Riel <riel@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: mgorman@suse.de Link: http://lkml.kernel.org/r/20160125170739.2fc9a641@annuminas.surriel.com [ Minor readability tweaks. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-01-25 22:07:39 +00:00
/*
* A node triggering more than 1/3 as many NUMA faults as the maximum is
* considered part of a numa group's pseudo-interleaving set. Migrations
* between these nodes are slowed down, to allow things to settle down.
*/
#define ACTIVE_NODE_FRACTION 3
static bool numa_is_active_node(int nid, struct numa_group *ng)
{
return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
}
sched/numa: Calculate node scores in complex NUMA topologies In order to do task placement on systems with complex NUMA topologies, it is necessary to count the faults on nodes nearby the node that is being examined for a potential move. In case of a system with a backplane interconnect, we are dealing with groups of NUMA nodes; each of the nodes within a group is the same number of hops away from nodes in other groups in the system. Optimal placement on this topology is achieved by counting all nearby nodes equally. When comparing nodes A and B at distance N, nearby nodes are those at distances smaller than N from nodes A or B. Placement strategy on a system with a glueless mesh NUMA topology needs to be different, because there are no natural groups of nodes determined by the hardware. Instead, when dealing with two nodes A and B at distance N, N >= 2, there will be intermediate nodes at distance < N from both nodes A and B. Good placement can be achieved by right shifting the faults on nearby nodes by the number of hops from the node being scored. In this context, a nearby node is any node less than the maximum distance in the system away from the node. Those nodes are skipped for efficiency reasons, there is no real policy reason to do so. Placement policy on directly connected NUMA systems is not affected. Signed-off-by: Rik van Riel <riel@redhat.com> Tested-by: Chegu Vinod <chegu_vinod@hp.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: mgorman@suse.de Cc: chegu_vinod@hp.com Link: http://lkml.kernel.org/r/1413530994-9732-5-git-send-email-riel@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-10-17 07:29:52 +00:00
/* Handle placement on systems where not all nodes are directly connected. */
static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
int maxdist, bool task)
{
unsigned long score = 0;
int node;
/*
* All nodes are directly connected, and the same distance
* from each other. No need for fancy placement algorithms.
*/
if (sched_numa_topology_type == NUMA_DIRECT)
return 0;
/*
* This code is called for each node, introducing N^2 complexity,
* which should be ok given the number of nodes rarely exceeds 8.
*/
for_each_online_node(node) {
unsigned long faults;
int dist = node_distance(nid, node);
/*
* The furthest away nodes in the system are not interesting
* for placement; nid was already counted.
*/
if (dist == sched_max_numa_distance || node == nid)
continue;
/*
* On systems with a backplane NUMA topology, compare groups
* of nodes, and move tasks towards the group with the most
* memory accesses. When comparing two nodes at distance
* "hoplimit", only nodes closer by than "hoplimit" are part
* of each group. Skip other nodes.
*/
if (sched_numa_topology_type == NUMA_BACKPLANE &&
sched/numa: Skip nodes that are at 'hoplimit' When comparing two nodes at a distance of 'hoplimit', we should consider nodes only up to 'hoplimit'. Currently we also consider nodes at 'oplimit' distance too. Hence two nodes at a distance of 'hoplimit' will have same groupweight. Fix this by skipping nodes at hoplimit. Running SPECjbb2005 on a 4 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 16 25375.3 25308.6 -0.26 1 72617 72964 0.477 Running SPECjbb2005 on a 16 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 8 113372 108750 -4.07684 1 177403 183115 3.21979 (numbers from v1 based on v4.17-rc5) Testcase Time: Min Max Avg StdDev numa01.sh Real: 478.45 565.90 515.11 30.87 numa01.sh Sys: 207.79 271.04 232.94 21.33 numa01.sh User: 39763.93 47303.12 43210.73 2644.86 numa02.sh Real: 60.00 61.46 60.78 0.49 numa02.sh Sys: 15.71 25.31 20.69 3.42 numa02.sh User: 5175.92 5265.86 5235.97 32.82 numa03.sh Real: 776.42 834.85 806.01 23.22 numa03.sh Sys: 114.43 128.75 121.65 5.49 numa03.sh User: 60773.93 64855.25 62616.91 1576.39 numa04.sh Real: 456.93 511.95 482.91 20.88 numa04.sh Sys: 178.09 460.89 356.86 94.58 numa04.sh User: 36312.09 42553.24 39623.21 2247.96 numa05.sh Real: 393.98 493.48 436.61 35.59 numa05.sh Sys: 164.49 329.15 265.87 61.78 numa05.sh User: 33182.65 36654.53 35074.51 1187.71 Testcase Time: Min Max Avg StdDev %Change numa01.sh Real: 414.64 819.20 556.08 147.70 -7.36% numa01.sh Sys: 77.52 205.04 139.40 52.05 67.10% numa01.sh User: 37043.24 61757.88 45517.48 9290.38 -5.06% numa02.sh Real: 60.80 63.32 61.63 0.88 -1.37% numa02.sh Sys: 17.35 39.37 25.71 7.33 -19.5% numa02.sh User: 5213.79 5374.73 5268.90 55.09 -0.62% numa03.sh Real: 780.09 948.64 831.43 63.02 -3.05% numa03.sh Sys: 104.96 136.92 116.31 11.34 4.591% numa03.sh User: 60465.42 73339.78 64368.03 4700.14 -2.72% numa04.sh Real: 412.60 681.92 521.29 96.64 -7.36% numa04.sh Sys: 210.32 314.10 251.77 37.71 41.74% numa04.sh User: 34026.38 45581.20 38534.49 4198.53 2.825% numa05.sh Real: 394.79 439.63 411.35 16.87 6.140% numa05.sh Sys: 238.32 330.09 292.31 38.32 -9.04% numa05.sh User: 33456.45 34876.07 34138.62 609.45 2.741% While there is a regression with this change, this change is needed from a correctness perspective. Also it helps consolidation as seen from perf bench output. Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Rik van Riel <riel@surriel.com> Acked-by: Mel Gorman <mgorman@techsingularity.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1529514181-9842-8-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-20 17:02:48 +00:00
dist >= maxdist)
sched/numa: Calculate node scores in complex NUMA topologies In order to do task placement on systems with complex NUMA topologies, it is necessary to count the faults on nodes nearby the node that is being examined for a potential move. In case of a system with a backplane interconnect, we are dealing with groups of NUMA nodes; each of the nodes within a group is the same number of hops away from nodes in other groups in the system. Optimal placement on this topology is achieved by counting all nearby nodes equally. When comparing nodes A and B at distance N, nearby nodes are those at distances smaller than N from nodes A or B. Placement strategy on a system with a glueless mesh NUMA topology needs to be different, because there are no natural groups of nodes determined by the hardware. Instead, when dealing with two nodes A and B at distance N, N >= 2, there will be intermediate nodes at distance < N from both nodes A and B. Good placement can be achieved by right shifting the faults on nearby nodes by the number of hops from the node being scored. In this context, a nearby node is any node less than the maximum distance in the system away from the node. Those nodes are skipped for efficiency reasons, there is no real policy reason to do so. Placement policy on directly connected NUMA systems is not affected. Signed-off-by: Rik van Riel <riel@redhat.com> Tested-by: Chegu Vinod <chegu_vinod@hp.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: mgorman@suse.de Cc: chegu_vinod@hp.com Link: http://lkml.kernel.org/r/1413530994-9732-5-git-send-email-riel@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-10-17 07:29:52 +00:00
continue;
/* Add up the faults from nearby nodes. */
if (task)
faults = task_faults(p, node);
else
faults = group_faults(p, node);
/*
* On systems with a glueless mesh NUMA topology, there are
* no fixed "groups of nodes". Instead, nodes that are not
* directly connected bounce traffic through intermediate
* nodes; a numa_group can occupy any set of nodes.
* The further away a node is, the less the faults count.
* This seems to result in good task placement.
*/
if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
faults *= (sched_max_numa_distance - dist);
faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
}
score += faults;
}
return score;
}
/*
* These return the fraction of accesses done by a particular task, or
* task group, on a particular numa node. The group weight is given a
* larger multiplier, in order to group tasks together that are almost
* evenly spread out between numa nodes.
*/
static inline unsigned long task_weight(struct task_struct *p, int nid,
int dist)
{
unsigned long faults, total_faults;
if (!p->numa_faults)
return 0;
total_faults = p->total_numa_faults;
if (!total_faults)
return 0;
faults = task_faults(p, nid);
sched/numa: Calculate node scores in complex NUMA topologies In order to do task placement on systems with complex NUMA topologies, it is necessary to count the faults on nodes nearby the node that is being examined for a potential move. In case of a system with a backplane interconnect, we are dealing with groups of NUMA nodes; each of the nodes within a group is the same number of hops away from nodes in other groups in the system. Optimal placement on this topology is achieved by counting all nearby nodes equally. When comparing nodes A and B at distance N, nearby nodes are those at distances smaller than N from nodes A or B. Placement strategy on a system with a glueless mesh NUMA topology needs to be different, because there are no natural groups of nodes determined by the hardware. Instead, when dealing with two nodes A and B at distance N, N >= 2, there will be intermediate nodes at distance < N from both nodes A and B. Good placement can be achieved by right shifting the faults on nearby nodes by the number of hops from the node being scored. In this context, a nearby node is any node less than the maximum distance in the system away from the node. Those nodes are skipped for efficiency reasons, there is no real policy reason to do so. Placement policy on directly connected NUMA systems is not affected. Signed-off-by: Rik van Riel <riel@redhat.com> Tested-by: Chegu Vinod <chegu_vinod@hp.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: mgorman@suse.de Cc: chegu_vinod@hp.com Link: http://lkml.kernel.org/r/1413530994-9732-5-git-send-email-riel@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-10-17 07:29:52 +00:00
faults += score_nearby_nodes(p, nid, dist, true);
return 1000 * faults / total_faults;
}
static inline unsigned long group_weight(struct task_struct *p, int nid,
int dist)
{
struct numa_group *ng = deref_task_numa_group(p);
unsigned long faults, total_faults;
if (!ng)
return 0;
total_faults = ng->total_faults;
if (!total_faults)
return 0;
faults = group_faults(p, nid);
sched/numa: Calculate node scores in complex NUMA topologies In order to do task placement on systems with complex NUMA topologies, it is necessary to count the faults on nodes nearby the node that is being examined for a potential move. In case of a system with a backplane interconnect, we are dealing with groups of NUMA nodes; each of the nodes within a group is the same number of hops away from nodes in other groups in the system. Optimal placement on this topology is achieved by counting all nearby nodes equally. When comparing nodes A and B at distance N, nearby nodes are those at distances smaller than N from nodes A or B. Placement strategy on a system with a glueless mesh NUMA topology needs to be different, because there are no natural groups of nodes determined by the hardware. Instead, when dealing with two nodes A and B at distance N, N >= 2, there will be intermediate nodes at distance < N from both nodes A and B. Good placement can be achieved by right shifting the faults on nearby nodes by the number of hops from the node being scored. In this context, a nearby node is any node less than the maximum distance in the system away from the node. Those nodes are skipped for efficiency reasons, there is no real policy reason to do so. Placement policy on directly connected NUMA systems is not affected. Signed-off-by: Rik van Riel <riel@redhat.com> Tested-by: Chegu Vinod <chegu_vinod@hp.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: mgorman@suse.de Cc: chegu_vinod@hp.com Link: http://lkml.kernel.org/r/1413530994-9732-5-git-send-email-riel@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-10-17 07:29:52 +00:00
faults += score_nearby_nodes(p, nid, dist, false);
return 1000 * faults / total_faults;
}
bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
int src_nid, int dst_cpu)
{
struct numa_group *ng = deref_curr_numa_group(p);
int dst_nid = cpu_to_node(dst_cpu);
int last_cpupid, this_cpupid;
this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
sched/numa: Migrate pages to local nodes quicker early in the lifetime of a task Automatic NUMA Balancing uses a multi-stage pass to decide whether a page should migrate to a local node. This filter avoids excessive ping-ponging if a page is shared or used by threads that migrate cross-node frequently. Threads inherit both page tables and the preferred node ID from the parent. This means that threads can trigger hinting faults earlier than a new task which delays scanning for a number of seconds. As it can be load balanced very early in its lifetime there can be an unnecessary delay before it starts migrating thread-local data. This patch migrates private pages faster early in the lifetime of a thread using the sequence counter as an identifier of new tasks. With this patch applied, STREAM performance is the same as 4.17 even though processes are not spread cross-node prematurely. Other workloads showed a mix of minor gains and losses. This is somewhat expected most workloads are not very sensitive to the starting conditions of a process. 4.19.0-rc5 4.19.0-rc5 4.17.0 numab-v1r1 fastmigrate-v1r1 vanilla MB/sec copy 43298.52 ( 0.00%) 47335.46 ( 9.32%) 47219.24 ( 9.06%) MB/sec scale 30115.06 ( 0.00%) 32568.12 ( 8.15%) 32527.56 ( 8.01%) MB/sec add 32825.12 ( 0.00%) 36078.94 ( 9.91%) 35928.02 ( 9.45%) MB/sec triad 32549.52 ( 0.00%) 35935.94 ( 10.40%) 35969.88 ( 10.51%) Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Reviewed-by: Rik van Riel <riel@surriel.com> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Jirka Hladky <jhladky@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Linux-MM <linux-mm@kvack.org> Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/20181001100525.29789-3-mgorman@techsingularity.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-10-01 10:05:25 +00:00
last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
/*
* Allow first faults or private faults to migrate immediately early in
* the lifetime of a task. The magic number 4 is based on waiting for
* two full passes of the "multi-stage node selection" test that is
* executed below.
*/
mm: replace all open encodings for NUMA_NO_NODE Patch series "Replace all open encodings for NUMA_NO_NODE", v3. All these places for replacement were found by running the following grep patterns on the entire kernel code. Please let me know if this might have missed some instances. This might also have replaced some false positives. I will appreciate suggestions, inputs and review. 1. git grep "nid == -1" 2. git grep "node == -1" 3. git grep "nid = -1" 4. git grep "node = -1" This patch (of 2): At present there are multiple places where invalid node number is encoded as -1. Even though implicitly understood it is always better to have macros in there. Replace these open encodings for an invalid node number with the global macro NUMA_NO_NODE. This helps remove NUMA related assumptions like 'invalid node' from various places redirecting them to a common definition. Link: http://lkml.kernel.org/r/1545127933-10711-2-git-send-email-anshuman.khandual@arm.com Signed-off-by: Anshuman Khandual <anshuman.khandual@arm.com> Reviewed-by: David Hildenbrand <david@redhat.com> Acked-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com> [ixgbe] Acked-by: Jens Axboe <axboe@kernel.dk> [mtip32xx] Acked-by: Vinod Koul <vkoul@kernel.org> [dmaengine.c] Acked-by: Michael Ellerman <mpe@ellerman.id.au> [powerpc] Acked-by: Doug Ledford <dledford@redhat.com> [drivers/infiniband] Cc: Joseph Qi <jiangqi903@gmail.com> Cc: Hans Verkuil <hverkuil@xs4all.nl> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:42:58 +00:00
if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) &&
sched/numa: Migrate pages to local nodes quicker early in the lifetime of a task Automatic NUMA Balancing uses a multi-stage pass to decide whether a page should migrate to a local node. This filter avoids excessive ping-ponging if a page is shared or used by threads that migrate cross-node frequently. Threads inherit both page tables and the preferred node ID from the parent. This means that threads can trigger hinting faults earlier than a new task which delays scanning for a number of seconds. As it can be load balanced very early in its lifetime there can be an unnecessary delay before it starts migrating thread-local data. This patch migrates private pages faster early in the lifetime of a thread using the sequence counter as an identifier of new tasks. With this patch applied, STREAM performance is the same as 4.17 even though processes are not spread cross-node prematurely. Other workloads showed a mix of minor gains and losses. This is somewhat expected most workloads are not very sensitive to the starting conditions of a process. 4.19.0-rc5 4.19.0-rc5 4.17.0 numab-v1r1 fastmigrate-v1r1 vanilla MB/sec copy 43298.52 ( 0.00%) 47335.46 ( 9.32%) 47219.24 ( 9.06%) MB/sec scale 30115.06 ( 0.00%) 32568.12 ( 8.15%) 32527.56 ( 8.01%) MB/sec add 32825.12 ( 0.00%) 36078.94 ( 9.91%) 35928.02 ( 9.45%) MB/sec triad 32549.52 ( 0.00%) 35935.94 ( 10.40%) 35969.88 ( 10.51%) Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Reviewed-by: Rik van Riel <riel@surriel.com> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Jirka Hladky <jhladky@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Linux-MM <linux-mm@kvack.org> Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/20181001100525.29789-3-mgorman@techsingularity.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-10-01 10:05:25 +00:00
(cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid)))
return true;
/*
* Multi-stage node selection is used in conjunction with a periodic
* migration fault to build a temporal task<->page relation. By using
* a two-stage filter we remove short/unlikely relations.
*
* Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
* a task's usage of a particular page (n_p) per total usage of this
* page (n_t) (in a given time-span) to a probability.
*
* Our periodic faults will sample this probability and getting the
* same result twice in a row, given these samples are fully
* independent, is then given by P(n)^2, provided our sample period
* is sufficiently short compared to the usage pattern.
*
* This quadric squishes small probabilities, making it less likely we
* act on an unlikely task<->page relation.
*/
if (!cpupid_pid_unset(last_cpupid) &&
cpupid_to_nid(last_cpupid) != dst_nid)
return false;
/* Always allow migrate on private faults */
if (cpupid_match_pid(p, last_cpupid))
return true;
/* A shared fault, but p->numa_group has not been set up yet. */
if (!ng)
return true;
/*
sched/numa: Spread memory according to CPU and memory use The pseudo-interleaving in NUMA placement has a fundamental problem: using hard usage thresholds to spread memory equally between nodes can prevent workloads from converging, or keep memory "trapped" on nodes where the workload is barely running any more. In order for workloads to properly converge, the memory migration should not be stopped when nodes reach parity, but instead be distributed according to how heavily memory is used from each node. This way memory migration and task migration reinforce each other, instead of one putting the brakes on the other. Remove the hard thresholds from the pseudo-interleaving code, and instead use a more gradual policy on memory placement. This also seems to improve convergence of workloads that do not run flat out, but sleep in between bursts of activity. We still want to slow down NUMA scanning and migration once a workload has settled on a few actively used nodes, so keep the 3/4 hysteresis in place. Keep track of whether a workload is actively running on multiple nodes, so task_numa_migrate does a full scan of the system for better task placement. In the case of running 3 SPECjbb2005 instances on a 4 node system, this code seems to result in fairer distribution of memory between nodes, with more memory bandwidth for each instance. Signed-off-by: Rik van Riel <riel@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: mgorman@suse.de Link: http://lkml.kernel.org/r/20160125170739.2fc9a641@annuminas.surriel.com [ Minor readability tweaks. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-01-25 22:07:39 +00:00
* Destination node is much more heavily used than the source
* node? Allow migration.
*/
sched/numa: Spread memory according to CPU and memory use The pseudo-interleaving in NUMA placement has a fundamental problem: using hard usage thresholds to spread memory equally between nodes can prevent workloads from converging, or keep memory "trapped" on nodes where the workload is barely running any more. In order for workloads to properly converge, the memory migration should not be stopped when nodes reach parity, but instead be distributed according to how heavily memory is used from each node. This way memory migration and task migration reinforce each other, instead of one putting the brakes on the other. Remove the hard thresholds from the pseudo-interleaving code, and instead use a more gradual policy on memory placement. This also seems to improve convergence of workloads that do not run flat out, but sleep in between bursts of activity. We still want to slow down NUMA scanning and migration once a workload has settled on a few actively used nodes, so keep the 3/4 hysteresis in place. Keep track of whether a workload is actively running on multiple nodes, so task_numa_migrate does a full scan of the system for better task placement. In the case of running 3 SPECjbb2005 instances on a 4 node system, this code seems to result in fairer distribution of memory between nodes, with more memory bandwidth for each instance. Signed-off-by: Rik van Riel <riel@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: mgorman@suse.de Link: http://lkml.kernel.org/r/20160125170739.2fc9a641@annuminas.surriel.com [ Minor readability tweaks. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-01-25 22:07:39 +00:00
if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
ACTIVE_NODE_FRACTION)
return true;
/*
sched/numa: Spread memory according to CPU and memory use The pseudo-interleaving in NUMA placement has a fundamental problem: using hard usage thresholds to spread memory equally between nodes can prevent workloads from converging, or keep memory "trapped" on nodes where the workload is barely running any more. In order for workloads to properly converge, the memory migration should not be stopped when nodes reach parity, but instead be distributed according to how heavily memory is used from each node. This way memory migration and task migration reinforce each other, instead of one putting the brakes on the other. Remove the hard thresholds from the pseudo-interleaving code, and instead use a more gradual policy on memory placement. This also seems to improve convergence of workloads that do not run flat out, but sleep in between bursts of activity. We still want to slow down NUMA scanning and migration once a workload has settled on a few actively used nodes, so keep the 3/4 hysteresis in place. Keep track of whether a workload is actively running on multiple nodes, so task_numa_migrate does a full scan of the system for better task placement. In the case of running 3 SPECjbb2005 instances on a 4 node system, this code seems to result in fairer distribution of memory between nodes, with more memory bandwidth for each instance. Signed-off-by: Rik van Riel <riel@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: mgorman@suse.de Link: http://lkml.kernel.org/r/20160125170739.2fc9a641@annuminas.surriel.com [ Minor readability tweaks. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-01-25 22:07:39 +00:00
* Distribute memory according to CPU & memory use on each node,
* with 3/4 hysteresis to avoid unnecessary memory migrations:
*
* faults_cpu(dst) 3 faults_cpu(src)
* --------------- * - > ---------------
* faults_mem(dst) 4 faults_mem(src)
*/
sched/numa: Spread memory according to CPU and memory use The pseudo-interleaving in NUMA placement has a fundamental problem: using hard usage thresholds to spread memory equally between nodes can prevent workloads from converging, or keep memory "trapped" on nodes where the workload is barely running any more. In order for workloads to properly converge, the memory migration should not be stopped when nodes reach parity, but instead be distributed according to how heavily memory is used from each node. This way memory migration and task migration reinforce each other, instead of one putting the brakes on the other. Remove the hard thresholds from the pseudo-interleaving code, and instead use a more gradual policy on memory placement. This also seems to improve convergence of workloads that do not run flat out, but sleep in between bursts of activity. We still want to slow down NUMA scanning and migration once a workload has settled on a few actively used nodes, so keep the 3/4 hysteresis in place. Keep track of whether a workload is actively running on multiple nodes, so task_numa_migrate does a full scan of the system for better task placement. In the case of running 3 SPECjbb2005 instances on a 4 node system, this code seems to result in fairer distribution of memory between nodes, with more memory bandwidth for each instance. Signed-off-by: Rik van Riel <riel@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: mgorman@suse.de Link: http://lkml.kernel.org/r/20160125170739.2fc9a641@annuminas.surriel.com [ Minor readability tweaks. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-01-25 22:07:39 +00:00
return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
}
/*
* 'numa_type' describes the node at the moment of load balancing.
*/
enum numa_type {
/* The node has spare capacity that can be used to run more tasks. */
node_has_spare = 0,
/*
* The node is fully used and the tasks don't compete for more CPU
* cycles. Nevertheless, some tasks might wait before running.
*/
node_fully_busy,
/*
* The node is overloaded and can't provide expected CPU cycles to all
* tasks.
*/
node_overloaded
};
/* Cached statistics for all CPUs within a node */
struct numa_stats {
unsigned long load;
unsigned long util;
/* Total compute capacity of CPUs on a node */
unsigned long compute_capacity;
unsigned int nr_running;
unsigned int weight;
enum numa_type node_type;
sched/numa: Prefer using an idle CPU as a migration target instead of comparing tasks task_numa_find_cpu() can scan a node multiple times. Minimally it scans to gather statistics and later to find a suitable target. In some cases, the second scan will simply pick an idle CPU if the load is not imbalanced. This patch caches information on an idle core while gathering statistics and uses it immediately if load is not imbalanced to avoid a second scan of the node runqueues. Preference is given to an idle core rather than an idle SMT sibling to avoid packing HT siblings due to linearly scanning the node cpumask. As a side-effect, even when the second scan is necessary, the importance of using select_idle_sibling is much reduced because information on idle CPUs is cached and can be reused. Note that this patch actually makes is harder to move to an idle CPU as multiple tasks can race for the same idle CPU due to a race checking numa_migrate_on. This is addressed in the next patch. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Ingo Molnar <mingo@kernel.org> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Valentin Schneider <valentin.schneider@arm.com> Cc: Phil Auld <pauld@redhat.com> Cc: Hillf Danton <hdanton@sina.com> Link: https://lore.kernel.org/r/20200224095223.13361-11-mgorman@techsingularity.net
2020-02-24 09:52:20 +00:00
int idle_cpu;
};
sched/numa: Prefer using an idle CPU as a migration target instead of comparing tasks task_numa_find_cpu() can scan a node multiple times. Minimally it scans to gather statistics and later to find a suitable target. In some cases, the second scan will simply pick an idle CPU if the load is not imbalanced. This patch caches information on an idle core while gathering statistics and uses it immediately if load is not imbalanced to avoid a second scan of the node runqueues. Preference is given to an idle core rather than an idle SMT sibling to avoid packing HT siblings due to linearly scanning the node cpumask. As a side-effect, even when the second scan is necessary, the importance of using select_idle_sibling is much reduced because information on idle CPUs is cached and can be reused. Note that this patch actually makes is harder to move to an idle CPU as multiple tasks can race for the same idle CPU due to a race checking numa_migrate_on. This is addressed in the next patch. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Ingo Molnar <mingo@kernel.org> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Valentin Schneider <valentin.schneider@arm.com> Cc: Phil Auld <pauld@redhat.com> Cc: Hillf Danton <hdanton@sina.com> Link: https://lore.kernel.org/r/20200224095223.13361-11-mgorman@techsingularity.net
2020-02-24 09:52:20 +00:00
static inline bool is_core_idle(int cpu)
{
#ifdef CONFIG_SCHED_SMT
int sibling;
for_each_cpu(sibling, cpu_smt_mask(cpu)) {
if (cpu == sibling)
continue;
if (!idle_cpu(cpu))
return false;
}
#endif
return true;
}
struct task_numa_env {
struct task_struct *p;
int src_cpu, src_nid;
int dst_cpu, dst_nid;
struct numa_stats src_stats, dst_stats;
int imbalance_pct;
int dist;
struct task_struct *best_task;
long best_imp;
int best_cpu;
};
static unsigned long cpu_load(struct rq *rq);
static unsigned long cpu_util(int cpu);
static inline long adjust_numa_imbalance(int imbalance, int src_nr_running);
static inline enum
numa_type numa_classify(unsigned int imbalance_pct,
struct numa_stats *ns)
{
if ((ns->nr_running > ns->weight) &&
((ns->compute_capacity * 100) < (ns->util * imbalance_pct)))
return node_overloaded;
if ((ns->nr_running < ns->weight) ||
((ns->compute_capacity * 100) > (ns->util * imbalance_pct)))
return node_has_spare;
return node_fully_busy;
}
#ifdef CONFIG_SCHED_SMT
/* Forward declarations of select_idle_sibling helpers */
static inline bool test_idle_cores(int cpu, bool def);
sched/numa: Prefer using an idle CPU as a migration target instead of comparing tasks task_numa_find_cpu() can scan a node multiple times. Minimally it scans to gather statistics and later to find a suitable target. In some cases, the second scan will simply pick an idle CPU if the load is not imbalanced. This patch caches information on an idle core while gathering statistics and uses it immediately if load is not imbalanced to avoid a second scan of the node runqueues. Preference is given to an idle core rather than an idle SMT sibling to avoid packing HT siblings due to linearly scanning the node cpumask. As a side-effect, even when the second scan is necessary, the importance of using select_idle_sibling is much reduced because information on idle CPUs is cached and can be reused. Note that this patch actually makes is harder to move to an idle CPU as multiple tasks can race for the same idle CPU due to a race checking numa_migrate_on. This is addressed in the next patch. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Ingo Molnar <mingo@kernel.org> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Valentin Schneider <valentin.schneider@arm.com> Cc: Phil Auld <pauld@redhat.com> Cc: Hillf Danton <hdanton@sina.com> Link: https://lore.kernel.org/r/20200224095223.13361-11-mgorman@techsingularity.net
2020-02-24 09:52:20 +00:00
static inline int numa_idle_core(int idle_core, int cpu)
{
if (!static_branch_likely(&sched_smt_present) ||
idle_core >= 0 || !test_idle_cores(cpu, false))
return idle_core;
/*
* Prefer cores instead of packing HT siblings
* and triggering future load balancing.
*/
if (is_core_idle(cpu))
idle_core = cpu;
return idle_core;
}
#else
static inline int numa_idle_core(int idle_core, int cpu)
{
return idle_core;
}
#endif
sched/numa: Prefer using an idle CPU as a migration target instead of comparing tasks task_numa_find_cpu() can scan a node multiple times. Minimally it scans to gather statistics and later to find a suitable target. In some cases, the second scan will simply pick an idle CPU if the load is not imbalanced. This patch caches information on an idle core while gathering statistics and uses it immediately if load is not imbalanced to avoid a second scan of the node runqueues. Preference is given to an idle core rather than an idle SMT sibling to avoid packing HT siblings due to linearly scanning the node cpumask. As a side-effect, even when the second scan is necessary, the importance of using select_idle_sibling is much reduced because information on idle CPUs is cached and can be reused. Note that this patch actually makes is harder to move to an idle CPU as multiple tasks can race for the same idle CPU due to a race checking numa_migrate_on. This is addressed in the next patch. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Ingo Molnar <mingo@kernel.org> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Valentin Schneider <valentin.schneider@arm.com> Cc: Phil Auld <pauld@redhat.com> Cc: Hillf Danton <hdanton@sina.com> Link: https://lore.kernel.org/r/20200224095223.13361-11-mgorman@techsingularity.net
2020-02-24 09:52:20 +00:00
/*
sched/numa: Prefer using an idle CPU as a migration target instead of comparing tasks task_numa_find_cpu() can scan a node multiple times. Minimally it scans to gather statistics and later to find a suitable target. In some cases, the second scan will simply pick an idle CPU if the load is not imbalanced. This patch caches information on an idle core while gathering statistics and uses it immediately if load is not imbalanced to avoid a second scan of the node runqueues. Preference is given to an idle core rather than an idle SMT sibling to avoid packing HT siblings due to linearly scanning the node cpumask. As a side-effect, even when the second scan is necessary, the importance of using select_idle_sibling is much reduced because information on idle CPUs is cached and can be reused. Note that this patch actually makes is harder to move to an idle CPU as multiple tasks can race for the same idle CPU due to a race checking numa_migrate_on. This is addressed in the next patch. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Ingo Molnar <mingo@kernel.org> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Valentin Schneider <valentin.schneider@arm.com> Cc: Phil Auld <pauld@redhat.com> Cc: Hillf Danton <hdanton@sina.com> Link: https://lore.kernel.org/r/20200224095223.13361-11-mgorman@techsingularity.net
2020-02-24 09:52:20 +00:00
* Gather all necessary information to make NUMA balancing placement
* decisions that are compatible with standard load balancer. This
* borrows code and logic from update_sg_lb_stats but sharing a
* common implementation is impractical.
*/
static void update_numa_stats(struct task_numa_env *env,
sched/numa: Prefer using an idle CPU as a migration target instead of comparing tasks task_numa_find_cpu() can scan a node multiple times. Minimally it scans to gather statistics and later to find a suitable target. In some cases, the second scan will simply pick an idle CPU if the load is not imbalanced. This patch caches information on an idle core while gathering statistics and uses it immediately if load is not imbalanced to avoid a second scan of the node runqueues. Preference is given to an idle core rather than an idle SMT sibling to avoid packing HT siblings due to linearly scanning the node cpumask. As a side-effect, even when the second scan is necessary, the importance of using select_idle_sibling is much reduced because information on idle CPUs is cached and can be reused. Note that this patch actually makes is harder to move to an idle CPU as multiple tasks can race for the same idle CPU due to a race checking numa_migrate_on. This is addressed in the next patch. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Ingo Molnar <mingo@kernel.org> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Valentin Schneider <valentin.schneider@arm.com> Cc: Phil Auld <pauld@redhat.com> Cc: Hillf Danton <hdanton@sina.com> Link: https://lore.kernel.org/r/20200224095223.13361-11-mgorman@techsingularity.net
2020-02-24 09:52:20 +00:00
struct numa_stats *ns, int nid,
bool find_idle)
{
sched/numa: Prefer using an idle CPU as a migration target instead of comparing tasks task_numa_find_cpu() can scan a node multiple times. Minimally it scans to gather statistics and later to find a suitable target. In some cases, the second scan will simply pick an idle CPU if the load is not imbalanced. This patch caches information on an idle core while gathering statistics and uses it immediately if load is not imbalanced to avoid a second scan of the node runqueues. Preference is given to an idle core rather than an idle SMT sibling to avoid packing HT siblings due to linearly scanning the node cpumask. As a side-effect, even when the second scan is necessary, the importance of using select_idle_sibling is much reduced because information on idle CPUs is cached and can be reused. Note that this patch actually makes is harder to move to an idle CPU as multiple tasks can race for the same idle CPU due to a race checking numa_migrate_on. This is addressed in the next patch. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Ingo Molnar <mingo@kernel.org> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Valentin Schneider <valentin.schneider@arm.com> Cc: Phil Auld <pauld@redhat.com> Cc: Hillf Danton <hdanton@sina.com> Link: https://lore.kernel.org/r/20200224095223.13361-11-mgorman@techsingularity.net
2020-02-24 09:52:20 +00:00
int cpu, idle_core = -1;
memset(ns, 0, sizeof(*ns));
sched/numa: Prefer using an idle CPU as a migration target instead of comparing tasks task_numa_find_cpu() can scan a node multiple times. Minimally it scans to gather statistics and later to find a suitable target. In some cases, the second scan will simply pick an idle CPU if the load is not imbalanced. This patch caches information on an idle core while gathering statistics and uses it immediately if load is not imbalanced to avoid a second scan of the node runqueues. Preference is given to an idle core rather than an idle SMT sibling to avoid packing HT siblings due to linearly scanning the node cpumask. As a side-effect, even when the second scan is necessary, the importance of using select_idle_sibling is much reduced because information on idle CPUs is cached and can be reused. Note that this patch actually makes is harder to move to an idle CPU as multiple tasks can race for the same idle CPU due to a race checking numa_migrate_on. This is addressed in the next patch. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Ingo Molnar <mingo@kernel.org> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Valentin Schneider <valentin.schneider@arm.com> Cc: Phil Auld <pauld@redhat.com> Cc: Hillf Danton <hdanton@sina.com> Link: https://lore.kernel.org/r/20200224095223.13361-11-mgorman@techsingularity.net
2020-02-24 09:52:20 +00:00
ns->idle_cpu = -1;
sched/numa: Acquire RCU lock for checking idle cores during NUMA balancing Qian Cai reported the following bug: The linux-next commit ff7db0bf24db ("sched/numa: Prefer using an idle CPU as a migration target instead of comparing tasks") introduced a boot warning, [ 86.520534][ T1] WARNING: suspicious RCU usage [ 86.520540][ T1] 5.6.0-rc3-next-20200227 #7 Not tainted [ 86.520545][ T1] ----------------------------- [ 86.520551][ T1] kernel/sched/fair.c:5914 suspicious rcu_dereference_check() usage! [ 86.520555][ T1] [ 86.520555][ T1] other info that might help us debug this: [ 86.520555][ T1] [ 86.520561][ T1] [ 86.520561][ T1] rcu_scheduler_active = 2, debug_locks = 1 [ 86.520567][ T1] 1 lock held by systemd/1: [ 86.520571][ T1] #0: ffff8887f4b14848 (&mm->mmap_sem#2){++++}, at: do_page_fault+0x1d2/0x998 [ 86.520594][ T1] [ 86.520594][ T1] stack backtrace: [ 86.520602][ T1] CPU: 1 PID: 1 Comm: systemd Not tainted 5.6.0-rc3-next-20200227 #7 task_numa_migrate() checks for idle cores when updating NUMA-related statistics. This relies on reading a RCU-protected structure in test_idle_cores() via this call chain task_numa_migrate -> update_numa_stats -> numa_idle_core -> test_idle_cores While the locking could be fine-grained, it is more appropriate to acquire the RCU lock for the entire scan of the domain. This patch removes the warning triggered at boot time. Reported-by: Qian Cai <cai@lca.pw> Reviewed-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Fixes: ff7db0bf24db ("sched/numa: Prefer using an idle CPU as a migration target instead of comparing tasks") Link: https://lkml.kernel.org/r/20200227191804.GJ3818@techsingularity.net
2020-02-27 19:18:04 +00:00
rcu_read_lock();
for_each_cpu(cpu, cpumask_of_node(nid)) {
struct rq *rq = cpu_rq(cpu);
ns->load += cpu_load(rq);
ns->util += cpu_util(cpu);
ns->nr_running += rq->cfs.h_nr_running;
ns->compute_capacity += capacity_of(cpu);
sched/numa: Prefer using an idle CPU as a migration target instead of comparing tasks task_numa_find_cpu() can scan a node multiple times. Minimally it scans to gather statistics and later to find a suitable target. In some cases, the second scan will simply pick an idle CPU if the load is not imbalanced. This patch caches information on an idle core while gathering statistics and uses it immediately if load is not imbalanced to avoid a second scan of the node runqueues. Preference is given to an idle core rather than an idle SMT sibling to avoid packing HT siblings due to linearly scanning the node cpumask. As a side-effect, even when the second scan is necessary, the importance of using select_idle_sibling is much reduced because information on idle CPUs is cached and can be reused. Note that this patch actually makes is harder to move to an idle CPU as multiple tasks can race for the same idle CPU due to a race checking numa_migrate_on. This is addressed in the next patch. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Ingo Molnar <mingo@kernel.org> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Valentin Schneider <valentin.schneider@arm.com> Cc: Phil Auld <pauld@redhat.com> Cc: Hillf Danton <hdanton@sina.com> Link: https://lore.kernel.org/r/20200224095223.13361-11-mgorman@techsingularity.net
2020-02-24 09:52:20 +00:00
if (find_idle && !rq->nr_running && idle_cpu(cpu)) {
if (READ_ONCE(rq->numa_migrate_on) ||
!cpumask_test_cpu(cpu, env->p->cpus_ptr))
continue;
if (ns->idle_cpu == -1)
ns->idle_cpu = cpu;
idle_core = numa_idle_core(idle_core, cpu);
}
}
sched/numa: Acquire RCU lock for checking idle cores during NUMA balancing Qian Cai reported the following bug: The linux-next commit ff7db0bf24db ("sched/numa: Prefer using an idle CPU as a migration target instead of comparing tasks") introduced a boot warning, [ 86.520534][ T1] WARNING: suspicious RCU usage [ 86.520540][ T1] 5.6.0-rc3-next-20200227 #7 Not tainted [ 86.520545][ T1] ----------------------------- [ 86.520551][ T1] kernel/sched/fair.c:5914 suspicious rcu_dereference_check() usage! [ 86.520555][ T1] [ 86.520555][ T1] other info that might help us debug this: [ 86.520555][ T1] [ 86.520561][ T1] [ 86.520561][ T1] rcu_scheduler_active = 2, debug_locks = 1 [ 86.520567][ T1] 1 lock held by systemd/1: [ 86.520571][ T1] #0: ffff8887f4b14848 (&mm->mmap_sem#2){++++}, at: do_page_fault+0x1d2/0x998 [ 86.520594][ T1] [ 86.520594][ T1] stack backtrace: [ 86.520602][ T1] CPU: 1 PID: 1 Comm: systemd Not tainted 5.6.0-rc3-next-20200227 #7 task_numa_migrate() checks for idle cores when updating NUMA-related statistics. This relies on reading a RCU-protected structure in test_idle_cores() via this call chain task_numa_migrate -> update_numa_stats -> numa_idle_core -> test_idle_cores While the locking could be fine-grained, it is more appropriate to acquire the RCU lock for the entire scan of the domain. This patch removes the warning triggered at boot time. Reported-by: Qian Cai <cai@lca.pw> Reviewed-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Fixes: ff7db0bf24db ("sched/numa: Prefer using an idle CPU as a migration target instead of comparing tasks") Link: https://lkml.kernel.org/r/20200227191804.GJ3818@techsingularity.net
2020-02-27 19:18:04 +00:00
rcu_read_unlock();
ns->weight = cpumask_weight(cpumask_of_node(nid));
ns->node_type = numa_classify(env->imbalance_pct, ns);
sched/numa: Prefer using an idle CPU as a migration target instead of comparing tasks task_numa_find_cpu() can scan a node multiple times. Minimally it scans to gather statistics and later to find a suitable target. In some cases, the second scan will simply pick an idle CPU if the load is not imbalanced. This patch caches information on an idle core while gathering statistics and uses it immediately if load is not imbalanced to avoid a second scan of the node runqueues. Preference is given to an idle core rather than an idle SMT sibling to avoid packing HT siblings due to linearly scanning the node cpumask. As a side-effect, even when the second scan is necessary, the importance of using select_idle_sibling is much reduced because information on idle CPUs is cached and can be reused. Note that this patch actually makes is harder to move to an idle CPU as multiple tasks can race for the same idle CPU due to a race checking numa_migrate_on. This is addressed in the next patch. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Ingo Molnar <mingo@kernel.org> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Valentin Schneider <valentin.schneider@arm.com> Cc: Phil Auld <pauld@redhat.com> Cc: Hillf Danton <hdanton@sina.com> Link: https://lore.kernel.org/r/20200224095223.13361-11-mgorman@techsingularity.net
2020-02-24 09:52:20 +00:00
if (idle_core >= 0)
ns->idle_cpu = idle_core;
}
static void task_numa_assign(struct task_numa_env *env,
struct task_struct *p, long imp)
{
sched/numa: Stop multiple tasks from moving to the CPU at the same time Task migration under NUMA balancing can happen in parallel. More than one task might choose to migrate to the same CPU at the same time. This can result in: - During task swap, choosing a task that was not part of the evaluation. - During task swap, task which just got moved into its preferred node, moving to a completely different node. - During task swap, task failing to move to the preferred node, will have to wait an extra interval for the next migrate opportunity. - During task movement, multiple task movements can cause load imbalance. This problem is more likely if there are more cores per node or more nodes in the system. Use a per run-queue variable to check if NUMA-balance is active on the run-queue. Specjbb2005 results (8 warehouses) Higher bops are better 2 Socket - 2 Node Haswell - X86 JVMS Prev Current %Change 4 200194 203353 1.57797 1 311331 328205 5.41995 2 Socket - 4 Node Power8 - PowerNV JVMS Prev Current %Change 1 197654 214384 8.46429 2 Socket - 2 Node Power9 - PowerNV JVMS Prev Current %Change 4 192605 188553 -2.10379 1 213402 196273 -8.02664 4 Socket - 4 Node Power7 - PowerVM JVMS Prev Current %Change 8 52227.1 57581.2 10.2516 1 102529 103468 0.915838 There is a regression on power 9 box. If we look at the details, that box has a sudden jump in cache-misses with this patch. All other parameters seem to be pointing towards NUMA consolidation. perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 13,345,784 13,941,377 migrations 1,127,820 1,157,323 faults 374,736 382,175 cache-misses 55,132,054,603 54,993,823,500 sched:sched_move_numa 1,923 2,005 sched:sched_stick_numa 52 14 sched:sched_swap_numa 595 529 migrate:mm_migrate_pages 1,932 1,573 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 60605 67099 numa_hint_faults_local 51804 58456 numa_hit 239945 240416 numa_huge_pte_updates 14 18 numa_interleave 60 65 numa_local 239865 240339 numa_other 80 77 numa_pages_migrated 1931 1574 numa_pte_updates 67823 77182 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 3,016,467 3,176,453 migrations 37,326 30,238 faults 115,342 87,869 cache-misses 11,692,155,554 12,544,479,391 sched:sched_move_numa 965 23 sched:sched_stick_numa 8 0 sched:sched_swap_numa 35 6 migrate:mm_migrate_pages 1,168 10 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 16286 236 numa_hint_faults_local 11863 201 numa_hit 112482 72293 numa_huge_pte_updates 33 0 numa_interleave 20 26 numa_local 112419 72233 numa_other 63 60 numa_pages_migrated 1144 8 numa_pte_updates 32859 0 perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 8,629,724 8,478,820 migrations 221,052 171,323 faults 308,661 307,499 cache-misses 135,574,913 240,353,599 sched:sched_move_numa 147 214 sched:sched_stick_numa 0 0 sched:sched_swap_numa 2 4 migrate:mm_migrate_pages 64 89 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 11481 5301 numa_hint_faults_local 10968 4745 numa_hit 89773 92943 numa_huge_pte_updates 0 0 numa_interleave 1116 899 numa_local 89220 92345 numa_other 553 598 numa_pages_migrated 62 88 numa_pte_updates 11694 5505 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 2,272,887 2,066,172 migrations 12,206 11,076 faults 163,704 149,544 cache-misses 4,801,186 10,398,067 sched:sched_move_numa 44 43 sched:sched_stick_numa 0 0 sched:sched_swap_numa 0 0 migrate:mm_migrate_pages 17 6 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 2261 3552 numa_hint_faults_local 1993 3347 numa_hit 25726 25611 numa_huge_pte_updates 0 0 numa_interleave 239 213 numa_local 25498 25583 numa_other 228 28 numa_pages_migrated 17 6 numa_pte_updates 2266 3535 perf stats 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 117,980,962 99,358,136 migrations 3,950,220 4,041,607 faults 736,979 749,653 cache-misses 224,976,072,879 225,562,543,251 sched:sched_move_numa 504 771 sched:sched_stick_numa 50 14 sched:sched_swap_numa 239 204 migrate:mm_migrate_pages 1,260 1,180 vmstat 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 18293 27409 numa_hint_faults_local 11969 20677 numa_hit 240854 239988 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 240851 239983 numa_other 3 5 numa_pages_migrated 1190 1016 numa_pte_updates 18106 27916 perf stats 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 61,053,158 60,899,307 migrations 551,586 544,668 faults 244,174 270,834 cache-misses 74,326,766,973 74,543,455,635 sched:sched_move_numa 344 735 sched:sched_stick_numa 24 25 sched:sched_swap_numa 140 174 migrate:mm_migrate_pages 568 816 vmstat 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 6461 11059 numa_hint_faults_local 2283 4733 numa_hit 35661 41384 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 35661 41383 numa_other 0 1 numa_pages_migrated 568 815 numa_pte_updates 6518 11323 Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Rik van Riel <riel@surriel.com> Acked-by: Mel Gorman <mgorman@techsingularity.net> Cc: Jirka Hladky <jhladky@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1537552141-27815-2-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-09-21 17:48:56 +00:00
struct rq *rq = cpu_rq(env->dst_cpu);
/* Check if run-queue part of active NUMA balance. */
if (env->best_cpu != env->dst_cpu && xchg(&rq->numa_migrate_on, 1)) {
int cpu;
int start = env->dst_cpu;
/* Find alternative idle CPU. */
for_each_cpu_wrap(cpu, cpumask_of_node(env->dst_nid), start) {
if (cpu == env->best_cpu || !idle_cpu(cpu) ||
!cpumask_test_cpu(cpu, env->p->cpus_ptr)) {
continue;
}
env->dst_cpu = cpu;
rq = cpu_rq(env->dst_cpu);
if (!xchg(&rq->numa_migrate_on, 1))
goto assign;
}
/* Failed to find an alternative idle CPU */
sched/numa: Stop multiple tasks from moving to the CPU at the same time Task migration under NUMA balancing can happen in parallel. More than one task might choose to migrate to the same CPU at the same time. This can result in: - During task swap, choosing a task that was not part of the evaluation. - During task swap, task which just got moved into its preferred node, moving to a completely different node. - During task swap, task failing to move to the preferred node, will have to wait an extra interval for the next migrate opportunity. - During task movement, multiple task movements can cause load imbalance. This problem is more likely if there are more cores per node or more nodes in the system. Use a per run-queue variable to check if NUMA-balance is active on the run-queue. Specjbb2005 results (8 warehouses) Higher bops are better 2 Socket - 2 Node Haswell - X86 JVMS Prev Current %Change 4 200194 203353 1.57797 1 311331 328205 5.41995 2 Socket - 4 Node Power8 - PowerNV JVMS Prev Current %Change 1 197654 214384 8.46429 2 Socket - 2 Node Power9 - PowerNV JVMS Prev Current %Change 4 192605 188553 -2.10379 1 213402 196273 -8.02664 4 Socket - 4 Node Power7 - PowerVM JVMS Prev Current %Change 8 52227.1 57581.2 10.2516 1 102529 103468 0.915838 There is a regression on power 9 box. If we look at the details, that box has a sudden jump in cache-misses with this patch. All other parameters seem to be pointing towards NUMA consolidation. perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 13,345,784 13,941,377 migrations 1,127,820 1,157,323 faults 374,736 382,175 cache-misses 55,132,054,603 54,993,823,500 sched:sched_move_numa 1,923 2,005 sched:sched_stick_numa 52 14 sched:sched_swap_numa 595 529 migrate:mm_migrate_pages 1,932 1,573 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 60605 67099 numa_hint_faults_local 51804 58456 numa_hit 239945 240416 numa_huge_pte_updates 14 18 numa_interleave 60 65 numa_local 239865 240339 numa_other 80 77 numa_pages_migrated 1931 1574 numa_pte_updates 67823 77182 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 3,016,467 3,176,453 migrations 37,326 30,238 faults 115,342 87,869 cache-misses 11,692,155,554 12,544,479,391 sched:sched_move_numa 965 23 sched:sched_stick_numa 8 0 sched:sched_swap_numa 35 6 migrate:mm_migrate_pages 1,168 10 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 16286 236 numa_hint_faults_local 11863 201 numa_hit 112482 72293 numa_huge_pte_updates 33 0 numa_interleave 20 26 numa_local 112419 72233 numa_other 63 60 numa_pages_migrated 1144 8 numa_pte_updates 32859 0 perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 8,629,724 8,478,820 migrations 221,052 171,323 faults 308,661 307,499 cache-misses 135,574,913 240,353,599 sched:sched_move_numa 147 214 sched:sched_stick_numa 0 0 sched:sched_swap_numa 2 4 migrate:mm_migrate_pages 64 89 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 11481 5301 numa_hint_faults_local 10968 4745 numa_hit 89773 92943 numa_huge_pte_updates 0 0 numa_interleave 1116 899 numa_local 89220 92345 numa_other 553 598 numa_pages_migrated 62 88 numa_pte_updates 11694 5505 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 2,272,887 2,066,172 migrations 12,206 11,076 faults 163,704 149,544 cache-misses 4,801,186 10,398,067 sched:sched_move_numa 44 43 sched:sched_stick_numa 0 0 sched:sched_swap_numa 0 0 migrate:mm_migrate_pages 17 6 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 2261 3552 numa_hint_faults_local 1993 3347 numa_hit 25726 25611 numa_huge_pte_updates 0 0 numa_interleave 239 213 numa_local 25498 25583 numa_other 228 28 numa_pages_migrated 17 6 numa_pte_updates 2266 3535 perf stats 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 117,980,962 99,358,136 migrations 3,950,220 4,041,607 faults 736,979 749,653 cache-misses 224,976,072,879 225,562,543,251 sched:sched_move_numa 504 771 sched:sched_stick_numa 50 14 sched:sched_swap_numa 239 204 migrate:mm_migrate_pages 1,260 1,180 vmstat 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 18293 27409 numa_hint_faults_local 11969 20677 numa_hit 240854 239988 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 240851 239983 numa_other 3 5 numa_pages_migrated 1190 1016 numa_pte_updates 18106 27916 perf stats 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 61,053,158 60,899,307 migrations 551,586 544,668 faults 244,174 270,834 cache-misses 74,326,766,973 74,543,455,635 sched:sched_move_numa 344 735 sched:sched_stick_numa 24 25 sched:sched_swap_numa 140 174 migrate:mm_migrate_pages 568 816 vmstat 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 6461 11059 numa_hint_faults_local 2283 4733 numa_hit 35661 41384 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 35661 41383 numa_other 0 1 numa_pages_migrated 568 815 numa_pte_updates 6518 11323 Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Rik van Riel <riel@surriel.com> Acked-by: Mel Gorman <mgorman@techsingularity.net> Cc: Jirka Hladky <jhladky@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1537552141-27815-2-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-09-21 17:48:56 +00:00
return;
}
sched/numa: Stop multiple tasks from moving to the CPU at the same time Task migration under NUMA balancing can happen in parallel. More than one task might choose to migrate to the same CPU at the same time. This can result in: - During task swap, choosing a task that was not part of the evaluation. - During task swap, task which just got moved into its preferred node, moving to a completely different node. - During task swap, task failing to move to the preferred node, will have to wait an extra interval for the next migrate opportunity. - During task movement, multiple task movements can cause load imbalance. This problem is more likely if there are more cores per node or more nodes in the system. Use a per run-queue variable to check if NUMA-balance is active on the run-queue. Specjbb2005 results (8 warehouses) Higher bops are better 2 Socket - 2 Node Haswell - X86 JVMS Prev Current %Change 4 200194 203353 1.57797 1 311331 328205 5.41995 2 Socket - 4 Node Power8 - PowerNV JVMS Prev Current %Change 1 197654 214384 8.46429 2 Socket - 2 Node Power9 - PowerNV JVMS Prev Current %Change 4 192605 188553 -2.10379 1 213402 196273 -8.02664 4 Socket - 4 Node Power7 - PowerVM JVMS Prev Current %Change 8 52227.1 57581.2 10.2516 1 102529 103468 0.915838 There is a regression on power 9 box. If we look at the details, that box has a sudden jump in cache-misses with this patch. All other parameters seem to be pointing towards NUMA consolidation. perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 13,345,784 13,941,377 migrations 1,127,820 1,157,323 faults 374,736 382,175 cache-misses 55,132,054,603 54,993,823,500 sched:sched_move_numa 1,923 2,005 sched:sched_stick_numa 52 14 sched:sched_swap_numa 595 529 migrate:mm_migrate_pages 1,932 1,573 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 60605 67099 numa_hint_faults_local 51804 58456 numa_hit 239945 240416 numa_huge_pte_updates 14 18 numa_interleave 60 65 numa_local 239865 240339 numa_other 80 77 numa_pages_migrated 1931 1574 numa_pte_updates 67823 77182 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 3,016,467 3,176,453 migrations 37,326 30,238 faults 115,342 87,869 cache-misses 11,692,155,554 12,544,479,391 sched:sched_move_numa 965 23 sched:sched_stick_numa 8 0 sched:sched_swap_numa 35 6 migrate:mm_migrate_pages 1,168 10 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 16286 236 numa_hint_faults_local 11863 201 numa_hit 112482 72293 numa_huge_pte_updates 33 0 numa_interleave 20 26 numa_local 112419 72233 numa_other 63 60 numa_pages_migrated 1144 8 numa_pte_updates 32859 0 perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 8,629,724 8,478,820 migrations 221,052 171,323 faults 308,661 307,499 cache-misses 135,574,913 240,353,599 sched:sched_move_numa 147 214 sched:sched_stick_numa 0 0 sched:sched_swap_numa 2 4 migrate:mm_migrate_pages 64 89 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 11481 5301 numa_hint_faults_local 10968 4745 numa_hit 89773 92943 numa_huge_pte_updates 0 0 numa_interleave 1116 899 numa_local 89220 92345 numa_other 553 598 numa_pages_migrated 62 88 numa_pte_updates 11694 5505 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 2,272,887 2,066,172 migrations 12,206 11,076 faults 163,704 149,544 cache-misses 4,801,186 10,398,067 sched:sched_move_numa 44 43 sched:sched_stick_numa 0 0 sched:sched_swap_numa 0 0 migrate:mm_migrate_pages 17 6 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 2261 3552 numa_hint_faults_local 1993 3347 numa_hit 25726 25611 numa_huge_pte_updates 0 0 numa_interleave 239 213 numa_local 25498 25583 numa_other 228 28 numa_pages_migrated 17 6 numa_pte_updates 2266 3535 perf stats 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 117,980,962 99,358,136 migrations 3,950,220 4,041,607 faults 736,979 749,653 cache-misses 224,976,072,879 225,562,543,251 sched:sched_move_numa 504 771 sched:sched_stick_numa 50 14 sched:sched_swap_numa 239 204 migrate:mm_migrate_pages 1,260 1,180 vmstat 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 18293 27409 numa_hint_faults_local 11969 20677 numa_hit 240854 239988 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 240851 239983 numa_other 3 5 numa_pages_migrated 1190 1016 numa_pte_updates 18106 27916 perf stats 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 61,053,158 60,899,307 migrations 551,586 544,668 faults 244,174 270,834 cache-misses 74,326,766,973 74,543,455,635 sched:sched_move_numa 344 735 sched:sched_stick_numa 24 25 sched:sched_swap_numa 140 174 migrate:mm_migrate_pages 568 816 vmstat 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 6461 11059 numa_hint_faults_local 2283 4733 numa_hit 35661 41384 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 35661 41383 numa_other 0 1 numa_pages_migrated 568 815 numa_pte_updates 6518 11323 Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Rik van Riel <riel@surriel.com> Acked-by: Mel Gorman <mgorman@techsingularity.net> Cc: Jirka Hladky <jhladky@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1537552141-27815-2-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-09-21 17:48:56 +00:00
assign:
sched/numa: Stop multiple tasks from moving to the CPU at the same time Task migration under NUMA balancing can happen in parallel. More than one task might choose to migrate to the same CPU at the same time. This can result in: - During task swap, choosing a task that was not part of the evaluation. - During task swap, task which just got moved into its preferred node, moving to a completely different node. - During task swap, task failing to move to the preferred node, will have to wait an extra interval for the next migrate opportunity. - During task movement, multiple task movements can cause load imbalance. This problem is more likely if there are more cores per node or more nodes in the system. Use a per run-queue variable to check if NUMA-balance is active on the run-queue. Specjbb2005 results (8 warehouses) Higher bops are better 2 Socket - 2 Node Haswell - X86 JVMS Prev Current %Change 4 200194 203353 1.57797 1 311331 328205 5.41995 2 Socket - 4 Node Power8 - PowerNV JVMS Prev Current %Change 1 197654 214384 8.46429 2 Socket - 2 Node Power9 - PowerNV JVMS Prev Current %Change 4 192605 188553 -2.10379 1 213402 196273 -8.02664 4 Socket - 4 Node Power7 - PowerVM JVMS Prev Current %Change 8 52227.1 57581.2 10.2516 1 102529 103468 0.915838 There is a regression on power 9 box. If we look at the details, that box has a sudden jump in cache-misses with this patch. All other parameters seem to be pointing towards NUMA consolidation. perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 13,345,784 13,941,377 migrations 1,127,820 1,157,323 faults 374,736 382,175 cache-misses 55,132,054,603 54,993,823,500 sched:sched_move_numa 1,923 2,005 sched:sched_stick_numa 52 14 sched:sched_swap_numa 595 529 migrate:mm_migrate_pages 1,932 1,573 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 60605 67099 numa_hint_faults_local 51804 58456 numa_hit 239945 240416 numa_huge_pte_updates 14 18 numa_interleave 60 65 numa_local 239865 240339 numa_other 80 77 numa_pages_migrated 1931 1574 numa_pte_updates 67823 77182 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 3,016,467 3,176,453 migrations 37,326 30,238 faults 115,342 87,869 cache-misses 11,692,155,554 12,544,479,391 sched:sched_move_numa 965 23 sched:sched_stick_numa 8 0 sched:sched_swap_numa 35 6 migrate:mm_migrate_pages 1,168 10 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 16286 236 numa_hint_faults_local 11863 201 numa_hit 112482 72293 numa_huge_pte_updates 33 0 numa_interleave 20 26 numa_local 112419 72233 numa_other 63 60 numa_pages_migrated 1144 8 numa_pte_updates 32859 0 perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 8,629,724 8,478,820 migrations 221,052 171,323 faults 308,661 307,499 cache-misses 135,574,913 240,353,599 sched:sched_move_numa 147 214 sched:sched_stick_numa 0 0 sched:sched_swap_numa 2 4 migrate:mm_migrate_pages 64 89 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 11481 5301 numa_hint_faults_local 10968 4745 numa_hit 89773 92943 numa_huge_pte_updates 0 0 numa_interleave 1116 899 numa_local 89220 92345 numa_other 553 598 numa_pages_migrated 62 88 numa_pte_updates 11694 5505 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 2,272,887 2,066,172 migrations 12,206 11,076 faults 163,704 149,544 cache-misses 4,801,186 10,398,067 sched:sched_move_numa 44 43 sched:sched_stick_numa 0 0 sched:sched_swap_numa 0 0 migrate:mm_migrate_pages 17 6 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 2261 3552 numa_hint_faults_local 1993 3347 numa_hit 25726 25611 numa_huge_pte_updates 0 0 numa_interleave 239 213 numa_local 25498 25583 numa_other 228 28 numa_pages_migrated 17 6 numa_pte_updates 2266 3535 perf stats 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 117,980,962 99,358,136 migrations 3,950,220 4,041,607 faults 736,979 749,653 cache-misses 224,976,072,879 225,562,543,251 sched:sched_move_numa 504 771 sched:sched_stick_numa 50 14 sched:sched_swap_numa 239 204 migrate:mm_migrate_pages 1,260 1,180 vmstat 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 18293 27409 numa_hint_faults_local 11969 20677 numa_hit 240854 239988 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 240851 239983 numa_other 3 5 numa_pages_migrated 1190 1016 numa_pte_updates 18106 27916 perf stats 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 61,053,158 60,899,307 migrations 551,586 544,668 faults 244,174 270,834 cache-misses 74,326,766,973 74,543,455,635 sched:sched_move_numa 344 735 sched:sched_stick_numa 24 25 sched:sched_swap_numa 140 174 migrate:mm_migrate_pages 568 816 vmstat 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 6461 11059 numa_hint_faults_local 2283 4733 numa_hit 35661 41384 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 35661 41383 numa_other 0 1 numa_pages_migrated 568 815 numa_pte_updates 6518 11323 Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Rik van Riel <riel@surriel.com> Acked-by: Mel Gorman <mgorman@techsingularity.net> Cc: Jirka Hladky <jhladky@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1537552141-27815-2-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-09-21 17:48:56 +00:00
/*
* Clear previous best_cpu/rq numa-migrate flag, since task now
* found a better CPU to move/swap.
*/
if (env->best_cpu != -1 && env->best_cpu != env->dst_cpu) {
sched/numa: Stop multiple tasks from moving to the CPU at the same time Task migration under NUMA balancing can happen in parallel. More than one task might choose to migrate to the same CPU at the same time. This can result in: - During task swap, choosing a task that was not part of the evaluation. - During task swap, task which just got moved into its preferred node, moving to a completely different node. - During task swap, task failing to move to the preferred node, will have to wait an extra interval for the next migrate opportunity. - During task movement, multiple task movements can cause load imbalance. This problem is more likely if there are more cores per node or more nodes in the system. Use a per run-queue variable to check if NUMA-balance is active on the run-queue. Specjbb2005 results (8 warehouses) Higher bops are better 2 Socket - 2 Node Haswell - X86 JVMS Prev Current %Change 4 200194 203353 1.57797 1 311331 328205 5.41995 2 Socket - 4 Node Power8 - PowerNV JVMS Prev Current %Change 1 197654 214384 8.46429 2 Socket - 2 Node Power9 - PowerNV JVMS Prev Current %Change 4 192605 188553 -2.10379 1 213402 196273 -8.02664 4 Socket - 4 Node Power7 - PowerVM JVMS Prev Current %Change 8 52227.1 57581.2 10.2516 1 102529 103468 0.915838 There is a regression on power 9 box. If we look at the details, that box has a sudden jump in cache-misses with this patch. All other parameters seem to be pointing towards NUMA consolidation. perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 13,345,784 13,941,377 migrations 1,127,820 1,157,323 faults 374,736 382,175 cache-misses 55,132,054,603 54,993,823,500 sched:sched_move_numa 1,923 2,005 sched:sched_stick_numa 52 14 sched:sched_swap_numa 595 529 migrate:mm_migrate_pages 1,932 1,573 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 60605 67099 numa_hint_faults_local 51804 58456 numa_hit 239945 240416 numa_huge_pte_updates 14 18 numa_interleave 60 65 numa_local 239865 240339 numa_other 80 77 numa_pages_migrated 1931 1574 numa_pte_updates 67823 77182 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 3,016,467 3,176,453 migrations 37,326 30,238 faults 115,342 87,869 cache-misses 11,692,155,554 12,544,479,391 sched:sched_move_numa 965 23 sched:sched_stick_numa 8 0 sched:sched_swap_numa 35 6 migrate:mm_migrate_pages 1,168 10 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 16286 236 numa_hint_faults_local 11863 201 numa_hit 112482 72293 numa_huge_pte_updates 33 0 numa_interleave 20 26 numa_local 112419 72233 numa_other 63 60 numa_pages_migrated 1144 8 numa_pte_updates 32859 0 perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 8,629,724 8,478,820 migrations 221,052 171,323 faults 308,661 307,499 cache-misses 135,574,913 240,353,599 sched:sched_move_numa 147 214 sched:sched_stick_numa 0 0 sched:sched_swap_numa 2 4 migrate:mm_migrate_pages 64 89 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 11481 5301 numa_hint_faults_local 10968 4745 numa_hit 89773 92943 numa_huge_pte_updates 0 0 numa_interleave 1116 899 numa_local 89220 92345 numa_other 553 598 numa_pages_migrated 62 88 numa_pte_updates 11694 5505 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 2,272,887 2,066,172 migrations 12,206 11,076 faults 163,704 149,544 cache-misses 4,801,186 10,398,067 sched:sched_move_numa 44 43 sched:sched_stick_numa 0 0 sched:sched_swap_numa 0 0 migrate:mm_migrate_pages 17 6 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 2261 3552 numa_hint_faults_local 1993 3347 numa_hit 25726 25611 numa_huge_pte_updates 0 0 numa_interleave 239 213 numa_local 25498 25583 numa_other 228 28 numa_pages_migrated 17 6 numa_pte_updates 2266 3535 perf stats 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 117,980,962 99,358,136 migrations 3,950,220 4,041,607 faults 736,979 749,653 cache-misses 224,976,072,879 225,562,543,251 sched:sched_move_numa 504 771 sched:sched_stick_numa 50 14 sched:sched_swap_numa 239 204 migrate:mm_migrate_pages 1,260 1,180 vmstat 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 18293 27409 numa_hint_faults_local 11969 20677 numa_hit 240854 239988 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 240851 239983 numa_other 3 5 numa_pages_migrated 1190 1016 numa_pte_updates 18106 27916 perf stats 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 61,053,158 60,899,307 migrations 551,586 544,668 faults 244,174 270,834 cache-misses 74,326,766,973 74,543,455,635 sched:sched_move_numa 344 735 sched:sched_stick_numa 24 25 sched:sched_swap_numa 140 174 migrate:mm_migrate_pages 568 816 vmstat 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 6461 11059 numa_hint_faults_local 2283 4733 numa_hit 35661 41384 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 35661 41383 numa_other 0 1 numa_pages_migrated 568 815 numa_pte_updates 6518 11323 Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Rik van Riel <riel@surriel.com> Acked-by: Mel Gorman <mgorman@techsingularity.net> Cc: Jirka Hladky <jhladky@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1537552141-27815-2-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-09-21 17:48:56 +00:00
rq = cpu_rq(env->best_cpu);
WRITE_ONCE(rq->numa_migrate_on, 0);
}
if (env->best_task)
put_task_struct(env->best_task);
if (p)
get_task_struct(p);
env->best_task = p;
env->best_imp = imp;
env->best_cpu = env->dst_cpu;
}
static bool load_too_imbalanced(long src_load, long dst_load,
struct task_numa_env *env)
{
long imb, old_imb;
long orig_src_load, orig_dst_load;
long src_capacity, dst_capacity;
/*
* The load is corrected for the CPU capacity available on each node.
*
* src_load dst_load
* ------------ vs ---------
* src_capacity dst_capacity
*/
src_capacity = env->src_stats.compute_capacity;
dst_capacity = env->dst_stats.compute_capacity;
sched/numa: Simplify load_too_imbalanced() Currently load_too_imbalance() cares about the slope of imbalance. It doesn't care of the direction of the imbalance. However this may not work if nodes that are being compared have dissimilar capacities. Few nodes might have more cores than other nodes in the system. Also unlike traditional load balance at a NUMA sched domain, multiple requests to migrate from the same source node to same destination node may run in parallel. This can cause huge load imbalance. This is specially true on a larger machines with either large cores per node or more number of nodes in the system. Hence allow move/swap only if the imbalance is going to reduce. Running SPECjbb2005 on a 4 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 16 25058.2 25122.9 0.25 1 72950 73850 1.23 (numbers from v1 based on v4.17-rc5) Testcase Time: Min Max Avg StdDev numa01.sh Real: 516.14 892.41 739.84 151.32 numa01.sh Sys: 153.16 192.99 177.70 14.58 numa01.sh User: 39821.04 69528.92 57193.87 10989.48 numa02.sh Real: 60.91 62.35 61.58 0.63 numa02.sh Sys: 16.47 26.16 21.20 3.85 numa02.sh User: 5227.58 5309.61 5265.17 31.04 numa03.sh Real: 739.07 917.73 795.75 64.45 numa03.sh Sys: 94.46 136.08 109.48 14.58 numa03.sh User: 57478.56 72014.09 61764.48 5343.69 numa04.sh Real: 442.61 715.43 530.31 96.12 numa04.sh Sys: 224.90 348.63 285.61 48.83 numa04.sh User: 35836.84 47522.47 40235.41 3985.26 numa05.sh Real: 386.13 489.17 434.94 43.59 numa05.sh Sys: 144.29 438.56 278.80 105.78 numa05.sh User: 33255.86 36890.82 34879.31 1641.98 Testcase Time: Min Max Avg StdDev %Change numa01.sh Real: 435.78 653.81 534.58 83.20 38.39% numa01.sh Sys: 121.93 187.18 145.90 23.47 21.79% numa01.sh User: 37082.81 51402.80 43647.60 5409.75 31.03% numa02.sh Real: 60.64 61.63 61.19 0.40 0.637% numa02.sh Sys: 14.72 25.68 19.06 4.03 11.22% numa02.sh User: 5210.95 5266.69 5233.30 20.82 0.608% numa03.sh Real: 746.51 808.24 780.36 23.88 1.972% numa03.sh Sys: 97.26 108.48 105.07 4.28 4.197% numa03.sh User: 58956.30 61397.05 60162.95 1050.82 2.661% numa04.sh Real: 465.97 519.27 484.81 19.62 9.385% numa04.sh Sys: 304.43 359.08 334.68 20.64 -14.6% numa04.sh User: 37544.16 41186.15 39262.44 1314.91 2.478% numa05.sh Real: 411.57 457.20 433.29 16.58 0.380% numa05.sh Sys: 230.05 435.48 339.95 67.58 -17.9% numa05.sh User: 33325.54 36896.31 35637.84 1222.64 -2.12% Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Rik van Riel <riel@surriel.com> Acked-by: Mel Gorman <mgorman@techsingularity.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1529514181-9842-4-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-20 17:02:44 +00:00
imb = abs(dst_load * src_capacity - src_load * dst_capacity);
orig_src_load = env->src_stats.load;
orig_dst_load = env->dst_stats.load;
sched/numa: Simplify load_too_imbalanced() Currently load_too_imbalance() cares about the slope of imbalance. It doesn't care of the direction of the imbalance. However this may not work if nodes that are being compared have dissimilar capacities. Few nodes might have more cores than other nodes in the system. Also unlike traditional load balance at a NUMA sched domain, multiple requests to migrate from the same source node to same destination node may run in parallel. This can cause huge load imbalance. This is specially true on a larger machines with either large cores per node or more number of nodes in the system. Hence allow move/swap only if the imbalance is going to reduce. Running SPECjbb2005 on a 4 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 16 25058.2 25122.9 0.25 1 72950 73850 1.23 (numbers from v1 based on v4.17-rc5) Testcase Time: Min Max Avg StdDev numa01.sh Real: 516.14 892.41 739.84 151.32 numa01.sh Sys: 153.16 192.99 177.70 14.58 numa01.sh User: 39821.04 69528.92 57193.87 10989.48 numa02.sh Real: 60.91 62.35 61.58 0.63 numa02.sh Sys: 16.47 26.16 21.20 3.85 numa02.sh User: 5227.58 5309.61 5265.17 31.04 numa03.sh Real: 739.07 917.73 795.75 64.45 numa03.sh Sys: 94.46 136.08 109.48 14.58 numa03.sh User: 57478.56 72014.09 61764.48 5343.69 numa04.sh Real: 442.61 715.43 530.31 96.12 numa04.sh Sys: 224.90 348.63 285.61 48.83 numa04.sh User: 35836.84 47522.47 40235.41 3985.26 numa05.sh Real: 386.13 489.17 434.94 43.59 numa05.sh Sys: 144.29 438.56 278.80 105.78 numa05.sh User: 33255.86 36890.82 34879.31 1641.98 Testcase Time: Min Max Avg StdDev %Change numa01.sh Real: 435.78 653.81 534.58 83.20 38.39% numa01.sh Sys: 121.93 187.18 145.90 23.47 21.79% numa01.sh User: 37082.81 51402.80 43647.60 5409.75 31.03% numa02.sh Real: 60.64 61.63 61.19 0.40 0.637% numa02.sh Sys: 14.72 25.68 19.06 4.03 11.22% numa02.sh User: 5210.95 5266.69 5233.30 20.82 0.608% numa03.sh Real: 746.51 808.24 780.36 23.88 1.972% numa03.sh Sys: 97.26 108.48 105.07 4.28 4.197% numa03.sh User: 58956.30 61397.05 60162.95 1050.82 2.661% numa04.sh Real: 465.97 519.27 484.81 19.62 9.385% numa04.sh Sys: 304.43 359.08 334.68 20.64 -14.6% numa04.sh User: 37544.16 41186.15 39262.44 1314.91 2.478% numa05.sh Real: 411.57 457.20 433.29 16.58 0.380% numa05.sh Sys: 230.05 435.48 339.95 67.58 -17.9% numa05.sh User: 33325.54 36896.31 35637.84 1222.64 -2.12% Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Rik van Riel <riel@surriel.com> Acked-by: Mel Gorman <mgorman@techsingularity.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1529514181-9842-4-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-20 17:02:44 +00:00
old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
/* Would this change make things worse? */
return (imb > old_imb);
}
sched/numa: Avoid task migration for small NUMA improvement If NUMA improvement from the task migration is going to be very minimal, then avoid task migration. Specjbb2005 results (8 warehouses) Higher bops are better 2 Socket - 2 Node Haswell - X86 JVMS Prev Current %Change 4 198512 205910 3.72673 1 313559 318491 1.57291 2 Socket - 4 Node Power8 - PowerNV JVMS Prev Current %Change 8 74761.9 74935.9 0.232739 1 214874 226796 5.54837 2 Socket - 2 Node Power9 - PowerNV JVMS Prev Current %Change 4 180536 189780 5.12031 1 210281 205695 -2.18089 4 Socket - 4 Node Power7 - PowerVM JVMS Prev Current %Change 8 56511.4 60370 6.828 1 104899 108100 3.05151 1/7 cases is regressing, if we look at events migrate_pages seem to vary the most especially in the regressing case. Also some amount of variance is expected between different runs of Specjbb2005. Some events stats before and after applying the patch. perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 13,818,546 13,801,554 migrations 1,149,960 1,151,541 faults 385,583 433,246 cache-misses 55,259,546,768 55,168,691,835 sched:sched_move_numa 2,257 2,551 sched:sched_stick_numa 9 24 sched:sched_swap_numa 512 904 migrate:mm_migrate_pages 2,225 1,571 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 72692 113682 numa_hint_faults_local 62270 102163 numa_hit 238762 240181 numa_huge_pte_updates 48 36 numa_interleave 75 64 numa_local 238676 240103 numa_other 86 78 numa_pages_migrated 2225 1564 numa_pte_updates 98557 134080 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 3,173,490 3,079,150 migrations 36,966 31,455 faults 108,776 99,081 cache-misses 12,200,075,320 11,588,126,740 sched:sched_move_numa 1,264 1 sched:sched_stick_numa 0 0 sched:sched_swap_numa 0 0 migrate:mm_migrate_pages 899 36 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 21109 430 numa_hint_faults_local 17120 77 numa_hit 72934 71277 numa_huge_pte_updates 42 0 numa_interleave 33 22 numa_local 72866 71218 numa_other 68 59 numa_pages_migrated 915 23 numa_pte_updates 42326 0 perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 8,312,022 8,707,565 migrations 231,705 171,342 faults 310,242 310,820 cache-misses 402,324,573 136,115,400 sched:sched_move_numa 193 215 sched:sched_stick_numa 0 6 sched:sched_swap_numa 3 24 migrate:mm_migrate_pages 93 162 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 11838 8985 numa_hint_faults_local 11216 8154 numa_hit 90689 93819 numa_huge_pte_updates 0 0 numa_interleave 1579 882 numa_local 89634 93496 numa_other 1055 323 numa_pages_migrated 92 169 numa_pte_updates 12109 9217 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 2,170,481 2,152,072 migrations 10,126 10,704 faults 160,962 164,376 cache-misses 10,834,845 3,818,437 sched:sched_move_numa 10 16 sched:sched_stick_numa 0 0 sched:sched_swap_numa 0 7 migrate:mm_migrate_pages 2 199 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 403 2248 numa_hint_faults_local 358 1666 numa_hit 25898 25704 numa_huge_pte_updates 0 0 numa_interleave 207 200 numa_local 25860 25679 numa_other 38 25 numa_pages_migrated 2 197 numa_pte_updates 400 2234 perf stats 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 110,339,633 93,330,595 migrations 4,139,812 4,122,061 faults 863,622 865,979 cache-misses 231,838,045,660 225,395,083,479 sched:sched_move_numa 2,196 2,372 sched:sched_stick_numa 33 24 sched:sched_swap_numa 544 769 migrate:mm_migrate_pages 2,469 1,677 vmstat 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 85748 91638 numa_hint_faults_local 66831 78096 numa_hit 242213 242225 numa_huge_pte_updates 0 0 numa_interleave 0 2 numa_local 242211 242219 numa_other 2 6 numa_pages_migrated 2376 1515 numa_pte_updates 86233 92274 perf stats 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 59,331,057 51,487,271 migrations 552,019 537,170 faults 266,586 256,921 cache-misses 73,796,312,990 70,073,831,187 sched:sched_move_numa 981 576 sched:sched_stick_numa 54 24 sched:sched_swap_numa 286 327 migrate:mm_migrate_pages 713 726 vmstat 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 14807 12000 numa_hint_faults_local 5738 5024 numa_hit 36230 36470 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 36228 36465 numa_other 2 5 numa_pages_migrated 703 726 numa_pte_updates 14742 11930 Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Jirka Hladky <jhladky@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1537552141-27815-7-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-09-21 17:49:01 +00:00
/*
* Maximum NUMA importance can be 1998 (2*999);
* SMALLIMP @ 30 would be close to 1998/64.
* Used to deter task migration.
*/
#define SMALLIMP 30
/*
* This checks if the overall compute and NUMA accesses of the system would
* be improved if the source tasks was migrated to the target dst_cpu taking
* into account that it might be best if task running on the dst_cpu should
* be exchanged with the source task
*/
static bool task_numa_compare(struct task_numa_env *env,
sched/numa: Evaluate move once per node task_numa_compare() helps choose the best CPU to move or swap the selected task. To achieve this task_numa_compare() is called for every CPU in the node. Currently it evaluates if the task can be moved/swapped for each of the CPUs. However the move evaluation is mostly independent of the CPU. Evaluating the move logic once per node, provides scope for simplifying task_numa_compare(). Running SPECjbb2005 on a 4 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 16 25705.2 25058.2 -2.51 1 74433 72950 -1.99 Running SPECjbb2005 on a 16 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 8 96589.6 105930 9.670 1 181830 178624 -1.76 (numbers from v1 based on v4.17-rc5) Testcase Time: Min Max Avg StdDev numa01.sh Real: 440.65 941.32 758.98 189.17 numa01.sh Sys: 183.48 320.07 258.42 50.09 numa01.sh User: 37384.65 71818.14 60302.51 13798.96 numa02.sh Real: 61.24 65.35 62.49 1.49 numa02.sh Sys: 16.83 24.18 21.40 2.60 numa02.sh User: 5219.59 5356.34 5264.03 49.07 numa03.sh Real: 822.04 912.40 873.55 37.35 numa03.sh Sys: 118.80 140.94 132.90 7.60 numa03.sh User: 62485.19 70025.01 67208.33 2967.10 numa04.sh Real: 690.66 872.12 778.49 65.44 numa04.sh Sys: 459.26 563.03 494.03 42.39 numa04.sh User: 51116.44 70527.20 58849.44 8461.28 numa05.sh Real: 418.37 562.28 525.77 54.27 numa05.sh Sys: 299.45 481.00 392.49 64.27 numa05.sh User: 34115.09 41324.02 39105.30 2627.68 Testcase Time: Min Max Avg StdDev %Change numa01.sh Real: 516.14 892.41 739.84 151.32 2.587% numa01.sh Sys: 153.16 192.99 177.70 14.58 45.42% numa01.sh User: 39821.04 69528.92 57193.87 10989.48 5.435% numa02.sh Real: 60.91 62.35 61.58 0.63 1.477% numa02.sh Sys: 16.47 26.16 21.20 3.85 0.943% numa02.sh User: 5227.58 5309.61 5265.17 31.04 -0.02% numa03.sh Real: 739.07 917.73 795.75 64.45 9.776% numa03.sh Sys: 94.46 136.08 109.48 14.58 21.39% numa03.sh User: 57478.56 72014.09 61764.48 5343.69 8.813% numa04.sh Real: 442.61 715.43 530.31 96.12 46.79% numa04.sh Sys: 224.90 348.63 285.61 48.83 72.97% numa04.sh User: 35836.84 47522.47 40235.41 3985.26 46.26% numa05.sh Real: 386.13 489.17 434.94 43.59 20.88% numa05.sh Sys: 144.29 438.56 278.80 105.78 40.77% numa05.sh User: 33255.86 36890.82 34879.31 1641.98 12.11% Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mel Gorman <mgorman@techsingularity.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1529514181-9842-3-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-20 17:02:43 +00:00
long taskimp, long groupimp, bool maymove)
{
struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p);
struct rq *dst_rq = cpu_rq(env->dst_cpu);
long imp = p_ng ? groupimp : taskimp;
struct task_struct *cur;
long src_load, dst_load;
int dist = env->dist;
long moveimp = imp;
long load;
bool stopsearch = false;
sched/numa: Stop multiple tasks from moving to the CPU at the same time Task migration under NUMA balancing can happen in parallel. More than one task might choose to migrate to the same CPU at the same time. This can result in: - During task swap, choosing a task that was not part of the evaluation. - During task swap, task which just got moved into its preferred node, moving to a completely different node. - During task swap, task failing to move to the preferred node, will have to wait an extra interval for the next migrate opportunity. - During task movement, multiple task movements can cause load imbalance. This problem is more likely if there are more cores per node or more nodes in the system. Use a per run-queue variable to check if NUMA-balance is active on the run-queue. Specjbb2005 results (8 warehouses) Higher bops are better 2 Socket - 2 Node Haswell - X86 JVMS Prev Current %Change 4 200194 203353 1.57797 1 311331 328205 5.41995 2 Socket - 4 Node Power8 - PowerNV JVMS Prev Current %Change 1 197654 214384 8.46429 2 Socket - 2 Node Power9 - PowerNV JVMS Prev Current %Change 4 192605 188553 -2.10379 1 213402 196273 -8.02664 4 Socket - 4 Node Power7 - PowerVM JVMS Prev Current %Change 8 52227.1 57581.2 10.2516 1 102529 103468 0.915838 There is a regression on power 9 box. If we look at the details, that box has a sudden jump in cache-misses with this patch. All other parameters seem to be pointing towards NUMA consolidation. perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 13,345,784 13,941,377 migrations 1,127,820 1,157,323 faults 374,736 382,175 cache-misses 55,132,054,603 54,993,823,500 sched:sched_move_numa 1,923 2,005 sched:sched_stick_numa 52 14 sched:sched_swap_numa 595 529 migrate:mm_migrate_pages 1,932 1,573 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 60605 67099 numa_hint_faults_local 51804 58456 numa_hit 239945 240416 numa_huge_pte_updates 14 18 numa_interleave 60 65 numa_local 239865 240339 numa_other 80 77 numa_pages_migrated 1931 1574 numa_pte_updates 67823 77182 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 3,016,467 3,176,453 migrations 37,326 30,238 faults 115,342 87,869 cache-misses 11,692,155,554 12,544,479,391 sched:sched_move_numa 965 23 sched:sched_stick_numa 8 0 sched:sched_swap_numa 35 6 migrate:mm_migrate_pages 1,168 10 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 16286 236 numa_hint_faults_local 11863 201 numa_hit 112482 72293 numa_huge_pte_updates 33 0 numa_interleave 20 26 numa_local 112419 72233 numa_other 63 60 numa_pages_migrated 1144 8 numa_pte_updates 32859 0 perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 8,629,724 8,478,820 migrations 221,052 171,323 faults 308,661 307,499 cache-misses 135,574,913 240,353,599 sched:sched_move_numa 147 214 sched:sched_stick_numa 0 0 sched:sched_swap_numa 2 4 migrate:mm_migrate_pages 64 89 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 11481 5301 numa_hint_faults_local 10968 4745 numa_hit 89773 92943 numa_huge_pte_updates 0 0 numa_interleave 1116 899 numa_local 89220 92345 numa_other 553 598 numa_pages_migrated 62 88 numa_pte_updates 11694 5505 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 2,272,887 2,066,172 migrations 12,206 11,076 faults 163,704 149,544 cache-misses 4,801,186 10,398,067 sched:sched_move_numa 44 43 sched:sched_stick_numa 0 0 sched:sched_swap_numa 0 0 migrate:mm_migrate_pages 17 6 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 2261 3552 numa_hint_faults_local 1993 3347 numa_hit 25726 25611 numa_huge_pte_updates 0 0 numa_interleave 239 213 numa_local 25498 25583 numa_other 228 28 numa_pages_migrated 17 6 numa_pte_updates 2266 3535 perf stats 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 117,980,962 99,358,136 migrations 3,950,220 4,041,607 faults 736,979 749,653 cache-misses 224,976,072,879 225,562,543,251 sched:sched_move_numa 504 771 sched:sched_stick_numa 50 14 sched:sched_swap_numa 239 204 migrate:mm_migrate_pages 1,260 1,180 vmstat 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 18293 27409 numa_hint_faults_local 11969 20677 numa_hit 240854 239988 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 240851 239983 numa_other 3 5 numa_pages_migrated 1190 1016 numa_pte_updates 18106 27916 perf stats 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 61,053,158 60,899,307 migrations 551,586 544,668 faults 244,174 270,834 cache-misses 74,326,766,973 74,543,455,635 sched:sched_move_numa 344 735 sched:sched_stick_numa 24 25 sched:sched_swap_numa 140 174 migrate:mm_migrate_pages 568 816 vmstat 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 6461 11059 numa_hint_faults_local 2283 4733 numa_hit 35661 41384 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 35661 41383 numa_other 0 1 numa_pages_migrated 568 815 numa_pte_updates 6518 11323 Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Rik van Riel <riel@surriel.com> Acked-by: Mel Gorman <mgorman@techsingularity.net> Cc: Jirka Hladky <jhladky@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1537552141-27815-2-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-09-21 17:48:56 +00:00
if (READ_ONCE(dst_rq->numa_migrate_on))
return false;
sched/numa: Stop multiple tasks from moving to the CPU at the same time Task migration under NUMA balancing can happen in parallel. More than one task might choose to migrate to the same CPU at the same time. This can result in: - During task swap, choosing a task that was not part of the evaluation. - During task swap, task which just got moved into its preferred node, moving to a completely different node. - During task swap, task failing to move to the preferred node, will have to wait an extra interval for the next migrate opportunity. - During task movement, multiple task movements can cause load imbalance. This problem is more likely if there are more cores per node or more nodes in the system. Use a per run-queue variable to check if NUMA-balance is active on the run-queue. Specjbb2005 results (8 warehouses) Higher bops are better 2 Socket - 2 Node Haswell - X86 JVMS Prev Current %Change 4 200194 203353 1.57797 1 311331 328205 5.41995 2 Socket - 4 Node Power8 - PowerNV JVMS Prev Current %Change 1 197654 214384 8.46429 2 Socket - 2 Node Power9 - PowerNV JVMS Prev Current %Change 4 192605 188553 -2.10379 1 213402 196273 -8.02664 4 Socket - 4 Node Power7 - PowerVM JVMS Prev Current %Change 8 52227.1 57581.2 10.2516 1 102529 103468 0.915838 There is a regression on power 9 box. If we look at the details, that box has a sudden jump in cache-misses with this patch. All other parameters seem to be pointing towards NUMA consolidation. perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 13,345,784 13,941,377 migrations 1,127,820 1,157,323 faults 374,736 382,175 cache-misses 55,132,054,603 54,993,823,500 sched:sched_move_numa 1,923 2,005 sched:sched_stick_numa 52 14 sched:sched_swap_numa 595 529 migrate:mm_migrate_pages 1,932 1,573 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 60605 67099 numa_hint_faults_local 51804 58456 numa_hit 239945 240416 numa_huge_pte_updates 14 18 numa_interleave 60 65 numa_local 239865 240339 numa_other 80 77 numa_pages_migrated 1931 1574 numa_pte_updates 67823 77182 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 3,016,467 3,176,453 migrations 37,326 30,238 faults 115,342 87,869 cache-misses 11,692,155,554 12,544,479,391 sched:sched_move_numa 965 23 sched:sched_stick_numa 8 0 sched:sched_swap_numa 35 6 migrate:mm_migrate_pages 1,168 10 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 16286 236 numa_hint_faults_local 11863 201 numa_hit 112482 72293 numa_huge_pte_updates 33 0 numa_interleave 20 26 numa_local 112419 72233 numa_other 63 60 numa_pages_migrated 1144 8 numa_pte_updates 32859 0 perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 8,629,724 8,478,820 migrations 221,052 171,323 faults 308,661 307,499 cache-misses 135,574,913 240,353,599 sched:sched_move_numa 147 214 sched:sched_stick_numa 0 0 sched:sched_swap_numa 2 4 migrate:mm_migrate_pages 64 89 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 11481 5301 numa_hint_faults_local 10968 4745 numa_hit 89773 92943 numa_huge_pte_updates 0 0 numa_interleave 1116 899 numa_local 89220 92345 numa_other 553 598 numa_pages_migrated 62 88 numa_pte_updates 11694 5505 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 2,272,887 2,066,172 migrations 12,206 11,076 faults 163,704 149,544 cache-misses 4,801,186 10,398,067 sched:sched_move_numa 44 43 sched:sched_stick_numa 0 0 sched:sched_swap_numa 0 0 migrate:mm_migrate_pages 17 6 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 2261 3552 numa_hint_faults_local 1993 3347 numa_hit 25726 25611 numa_huge_pte_updates 0 0 numa_interleave 239 213 numa_local 25498 25583 numa_other 228 28 numa_pages_migrated 17 6 numa_pte_updates 2266 3535 perf stats 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 117,980,962 99,358,136 migrations 3,950,220 4,041,607 faults 736,979 749,653 cache-misses 224,976,072,879 225,562,543,251 sched:sched_move_numa 504 771 sched:sched_stick_numa 50 14 sched:sched_swap_numa 239 204 migrate:mm_migrate_pages 1,260 1,180 vmstat 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 18293 27409 numa_hint_faults_local 11969 20677 numa_hit 240854 239988 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 240851 239983 numa_other 3 5 numa_pages_migrated 1190 1016 numa_pte_updates 18106 27916 perf stats 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 61,053,158 60,899,307 migrations 551,586 544,668 faults 244,174 270,834 cache-misses 74,326,766,973 74,543,455,635 sched:sched_move_numa 344 735 sched:sched_stick_numa 24 25 sched:sched_swap_numa 140 174 migrate:mm_migrate_pages 568 816 vmstat 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 6461 11059 numa_hint_faults_local 2283 4733 numa_hit 35661 41384 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 35661 41383 numa_other 0 1 numa_pages_migrated 568 815 numa_pte_updates 6518 11323 Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Rik van Riel <riel@surriel.com> Acked-by: Mel Gorman <mgorman@techsingularity.net> Cc: Jirka Hladky <jhladky@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1537552141-27815-2-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-09-21 17:48:56 +00:00
rcu_read_lock();
tasks, sched/core: With a grace period after finish_task_switch(), remove unnecessary code Remove work arounds that were written before there was a grace period after tasks left the runqueue in finish_task_switch(). In particular now that there tasks exiting the runqueue exprience a RCU grace period none of the work performed by task_rcu_dereference() excpet the rcu_dereference() is necessary so replace task_rcu_dereference() with rcu_dereference(). Remove the code in rcuwait_wait_event() that checks to ensure the current task has not exited. It is no longer necessary as it is guaranteed that any running task will experience a RCU grace period after it leaves the run queueue. Remove the comment in rcuwait_wake_up() as it is no longer relevant. Ref: 8f95c90ceb54 ("sched/wait, RCU: Introduce rcuwait machinery") Ref: 150593bf8693 ("sched/api: Introduce task_rcu_dereference() and try_get_task_struct()") Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Chris Metcalf <cmetcalf@ezchip.com> Cc: Christoph Lameter <cl@linux.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Kirill Tkhai <tkhai@yandex.ru> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Paul E. McKenney <paulmck@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Russell King - ARM Linux admin <linux@armlinux.org.uk> Cc: Thomas Gleixner <tglx@linutronix.de> Link: https://lkml.kernel.org/r/87lfurdpk9.fsf_-_@x220.int.ebiederm.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-09-14 12:34:30 +00:00
cur = rcu_dereference(dst_rq->curr);
if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
cur = NULL;
/*
* Because we have preemption enabled we can get migrated around and
* end try selecting ourselves (current == env->p) as a swap candidate.
*/
if (cur == env->p) {
stopsearch = true;
goto unlock;
}
sched/numa: Evaluate move once per node task_numa_compare() helps choose the best CPU to move or swap the selected task. To achieve this task_numa_compare() is called for every CPU in the node. Currently it evaluates if the task can be moved/swapped for each of the CPUs. However the move evaluation is mostly independent of the CPU. Evaluating the move logic once per node, provides scope for simplifying task_numa_compare(). Running SPECjbb2005 on a 4 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 16 25705.2 25058.2 -2.51 1 74433 72950 -1.99 Running SPECjbb2005 on a 16 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 8 96589.6 105930 9.670 1 181830 178624 -1.76 (numbers from v1 based on v4.17-rc5) Testcase Time: Min Max Avg StdDev numa01.sh Real: 440.65 941.32 758.98 189.17 numa01.sh Sys: 183.48 320.07 258.42 50.09 numa01.sh User: 37384.65 71818.14 60302.51 13798.96 numa02.sh Real: 61.24 65.35 62.49 1.49 numa02.sh Sys: 16.83 24.18 21.40 2.60 numa02.sh User: 5219.59 5356.34 5264.03 49.07 numa03.sh Real: 822.04 912.40 873.55 37.35 numa03.sh Sys: 118.80 140.94 132.90 7.60 numa03.sh User: 62485.19 70025.01 67208.33 2967.10 numa04.sh Real: 690.66 872.12 778.49 65.44 numa04.sh Sys: 459.26 563.03 494.03 42.39 numa04.sh User: 51116.44 70527.20 58849.44 8461.28 numa05.sh Real: 418.37 562.28 525.77 54.27 numa05.sh Sys: 299.45 481.00 392.49 64.27 numa05.sh User: 34115.09 41324.02 39105.30 2627.68 Testcase Time: Min Max Avg StdDev %Change numa01.sh Real: 516.14 892.41 739.84 151.32 2.587% numa01.sh Sys: 153.16 192.99 177.70 14.58 45.42% numa01.sh User: 39821.04 69528.92 57193.87 10989.48 5.435% numa02.sh Real: 60.91 62.35 61.58 0.63 1.477% numa02.sh Sys: 16.47 26.16 21.20 3.85 0.943% numa02.sh User: 5227.58 5309.61 5265.17 31.04 -0.02% numa03.sh Real: 739.07 917.73 795.75 64.45 9.776% numa03.sh Sys: 94.46 136.08 109.48 14.58 21.39% numa03.sh User: 57478.56 72014.09 61764.48 5343.69 8.813% numa04.sh Real: 442.61 715.43 530.31 96.12 46.79% numa04.sh Sys: 224.90 348.63 285.61 48.83 72.97% numa04.sh User: 35836.84 47522.47 40235.41 3985.26 46.26% numa05.sh Real: 386.13 489.17 434.94 43.59 20.88% numa05.sh Sys: 144.29 438.56 278.80 105.78 40.77% numa05.sh User: 33255.86 36890.82 34879.31 1641.98 12.11% Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mel Gorman <mgorman@techsingularity.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1529514181-9842-3-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-20 17:02:43 +00:00
if (!cur) {
sched/numa: Avoid task migration for small NUMA improvement If NUMA improvement from the task migration is going to be very minimal, then avoid task migration. Specjbb2005 results (8 warehouses) Higher bops are better 2 Socket - 2 Node Haswell - X86 JVMS Prev Current %Change 4 198512 205910 3.72673 1 313559 318491 1.57291 2 Socket - 4 Node Power8 - PowerNV JVMS Prev Current %Change 8 74761.9 74935.9 0.232739 1 214874 226796 5.54837 2 Socket - 2 Node Power9 - PowerNV JVMS Prev Current %Change 4 180536 189780 5.12031 1 210281 205695 -2.18089 4 Socket - 4 Node Power7 - PowerVM JVMS Prev Current %Change 8 56511.4 60370 6.828 1 104899 108100 3.05151 1/7 cases is regressing, if we look at events migrate_pages seem to vary the most especially in the regressing case. Also some amount of variance is expected between different runs of Specjbb2005. Some events stats before and after applying the patch. perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 13,818,546 13,801,554 migrations 1,149,960 1,151,541 faults 385,583 433,246 cache-misses 55,259,546,768 55,168,691,835 sched:sched_move_numa 2,257 2,551 sched:sched_stick_numa 9 24 sched:sched_swap_numa 512 904 migrate:mm_migrate_pages 2,225 1,571 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 72692 113682 numa_hint_faults_local 62270 102163 numa_hit 238762 240181 numa_huge_pte_updates 48 36 numa_interleave 75 64 numa_local 238676 240103 numa_other 86 78 numa_pages_migrated 2225 1564 numa_pte_updates 98557 134080 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 3,173,490 3,079,150 migrations 36,966 31,455 faults 108,776 99,081 cache-misses 12,200,075,320 11,588,126,740 sched:sched_move_numa 1,264 1 sched:sched_stick_numa 0 0 sched:sched_swap_numa 0 0 migrate:mm_migrate_pages 899 36 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 21109 430 numa_hint_faults_local 17120 77 numa_hit 72934 71277 numa_huge_pte_updates 42 0 numa_interleave 33 22 numa_local 72866 71218 numa_other 68 59 numa_pages_migrated 915 23 numa_pte_updates 42326 0 perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 8,312,022 8,707,565 migrations 231,705 171,342 faults 310,242 310,820 cache-misses 402,324,573 136,115,400 sched:sched_move_numa 193 215 sched:sched_stick_numa 0 6 sched:sched_swap_numa 3 24 migrate:mm_migrate_pages 93 162 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 11838 8985 numa_hint_faults_local 11216 8154 numa_hit 90689 93819 numa_huge_pte_updates 0 0 numa_interleave 1579 882 numa_local 89634 93496 numa_other 1055 323 numa_pages_migrated 92 169 numa_pte_updates 12109 9217 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 2,170,481 2,152,072 migrations 10,126 10,704 faults 160,962 164,376 cache-misses 10,834,845 3,818,437 sched:sched_move_numa 10 16 sched:sched_stick_numa 0 0 sched:sched_swap_numa 0 7 migrate:mm_migrate_pages 2 199 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 403 2248 numa_hint_faults_local 358 1666 numa_hit 25898 25704 numa_huge_pte_updates 0 0 numa_interleave 207 200 numa_local 25860 25679 numa_other 38 25 numa_pages_migrated 2 197 numa_pte_updates 400 2234 perf stats 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 110,339,633 93,330,595 migrations 4,139,812 4,122,061 faults 863,622 865,979 cache-misses 231,838,045,660 225,395,083,479 sched:sched_move_numa 2,196 2,372 sched:sched_stick_numa 33 24 sched:sched_swap_numa 544 769 migrate:mm_migrate_pages 2,469 1,677 vmstat 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 85748 91638 numa_hint_faults_local 66831 78096 numa_hit 242213 242225 numa_huge_pte_updates 0 0 numa_interleave 0 2 numa_local 242211 242219 numa_other 2 6 numa_pages_migrated 2376 1515 numa_pte_updates 86233 92274 perf stats 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 59,331,057 51,487,271 migrations 552,019 537,170 faults 266,586 256,921 cache-misses 73,796,312,990 70,073,831,187 sched:sched_move_numa 981 576 sched:sched_stick_numa 54 24 sched:sched_swap_numa 286 327 migrate:mm_migrate_pages 713 726 vmstat 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 14807 12000 numa_hint_faults_local 5738 5024 numa_hit 36230 36470 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 36228 36465 numa_other 2 5 numa_pages_migrated 703 726 numa_pte_updates 14742 11930 Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Jirka Hladky <jhladky@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1537552141-27815-7-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-09-21 17:49:01 +00:00
if (maymove && moveimp >= env->best_imp)
sched/numa: Evaluate move once per node task_numa_compare() helps choose the best CPU to move or swap the selected task. To achieve this task_numa_compare() is called for every CPU in the node. Currently it evaluates if the task can be moved/swapped for each of the CPUs. However the move evaluation is mostly independent of the CPU. Evaluating the move logic once per node, provides scope for simplifying task_numa_compare(). Running SPECjbb2005 on a 4 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 16 25705.2 25058.2 -2.51 1 74433 72950 -1.99 Running SPECjbb2005 on a 16 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 8 96589.6 105930 9.670 1 181830 178624 -1.76 (numbers from v1 based on v4.17-rc5) Testcase Time: Min Max Avg StdDev numa01.sh Real: 440.65 941.32 758.98 189.17 numa01.sh Sys: 183.48 320.07 258.42 50.09 numa01.sh User: 37384.65 71818.14 60302.51 13798.96 numa02.sh Real: 61.24 65.35 62.49 1.49 numa02.sh Sys: 16.83 24.18 21.40 2.60 numa02.sh User: 5219.59 5356.34 5264.03 49.07 numa03.sh Real: 822.04 912.40 873.55 37.35 numa03.sh Sys: 118.80 140.94 132.90 7.60 numa03.sh User: 62485.19 70025.01 67208.33 2967.10 numa04.sh Real: 690.66 872.12 778.49 65.44 numa04.sh Sys: 459.26 563.03 494.03 42.39 numa04.sh User: 51116.44 70527.20 58849.44 8461.28 numa05.sh Real: 418.37 562.28 525.77 54.27 numa05.sh Sys: 299.45 481.00 392.49 64.27 numa05.sh User: 34115.09 41324.02 39105.30 2627.68 Testcase Time: Min Max Avg StdDev %Change numa01.sh Real: 516.14 892.41 739.84 151.32 2.587% numa01.sh Sys: 153.16 192.99 177.70 14.58 45.42% numa01.sh User: 39821.04 69528.92 57193.87 10989.48 5.435% numa02.sh Real: 60.91 62.35 61.58 0.63 1.477% numa02.sh Sys: 16.47 26.16 21.20 3.85 0.943% numa02.sh User: 5227.58 5309.61 5265.17 31.04 -0.02% numa03.sh Real: 739.07 917.73 795.75 64.45 9.776% numa03.sh Sys: 94.46 136.08 109.48 14.58 21.39% numa03.sh User: 57478.56 72014.09 61764.48 5343.69 8.813% numa04.sh Real: 442.61 715.43 530.31 96.12 46.79% numa04.sh Sys: 224.90 348.63 285.61 48.83 72.97% numa04.sh User: 35836.84 47522.47 40235.41 3985.26 46.26% numa05.sh Real: 386.13 489.17 434.94 43.59 20.88% numa05.sh Sys: 144.29 438.56 278.80 105.78 40.77% numa05.sh User: 33255.86 36890.82 34879.31 1641.98 12.11% Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mel Gorman <mgorman@techsingularity.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1529514181-9842-3-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-20 17:02:43 +00:00
goto assign;
else
goto unlock;
}
/* Skip this swap candidate if cannot move to the source cpu. */
if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr))
goto unlock;
/*
* Skip this swap candidate if it is not moving to its preferred
* node and the best task is.
*/
if (env->best_task &&
env->best_task->numa_preferred_nid == env->src_nid &&
cur->numa_preferred_nid != env->src_nid) {
goto unlock;
}
/*
* "imp" is the fault differential for the source task between the
* source and destination node. Calculate the total differential for
* the source task and potential destination task. The more negative
sched/numa: Evaluate move once per node task_numa_compare() helps choose the best CPU to move or swap the selected task. To achieve this task_numa_compare() is called for every CPU in the node. Currently it evaluates if the task can be moved/swapped for each of the CPUs. However the move evaluation is mostly independent of the CPU. Evaluating the move logic once per node, provides scope for simplifying task_numa_compare(). Running SPECjbb2005 on a 4 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 16 25705.2 25058.2 -2.51 1 74433 72950 -1.99 Running SPECjbb2005 on a 16 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 8 96589.6 105930 9.670 1 181830 178624 -1.76 (numbers from v1 based on v4.17-rc5) Testcase Time: Min Max Avg StdDev numa01.sh Real: 440.65 941.32 758.98 189.17 numa01.sh Sys: 183.48 320.07 258.42 50.09 numa01.sh User: 37384.65 71818.14 60302.51 13798.96 numa02.sh Real: 61.24 65.35 62.49 1.49 numa02.sh Sys: 16.83 24.18 21.40 2.60 numa02.sh User: 5219.59 5356.34 5264.03 49.07 numa03.sh Real: 822.04 912.40 873.55 37.35 numa03.sh Sys: 118.80 140.94 132.90 7.60 numa03.sh User: 62485.19 70025.01 67208.33 2967.10 numa04.sh Real: 690.66 872.12 778.49 65.44 numa04.sh Sys: 459.26 563.03 494.03 42.39 numa04.sh User: 51116.44 70527.20 58849.44 8461.28 numa05.sh Real: 418.37 562.28 525.77 54.27 numa05.sh Sys: 299.45 481.00 392.49 64.27 numa05.sh User: 34115.09 41324.02 39105.30 2627.68 Testcase Time: Min Max Avg StdDev %Change numa01.sh Real: 516.14 892.41 739.84 151.32 2.587% numa01.sh Sys: 153.16 192.99 177.70 14.58 45.42% numa01.sh User: 39821.04 69528.92 57193.87 10989.48 5.435% numa02.sh Real: 60.91 62.35 61.58 0.63 1.477% numa02.sh Sys: 16.47 26.16 21.20 3.85 0.943% numa02.sh User: 5227.58 5309.61 5265.17 31.04 -0.02% numa03.sh Real: 739.07 917.73 795.75 64.45 9.776% numa03.sh Sys: 94.46 136.08 109.48 14.58 21.39% numa03.sh User: 57478.56 72014.09 61764.48 5343.69 8.813% numa04.sh Real: 442.61 715.43 530.31 96.12 46.79% numa04.sh Sys: 224.90 348.63 285.61 48.83 72.97% numa04.sh User: 35836.84 47522.47 40235.41 3985.26 46.26% numa05.sh Real: 386.13 489.17 434.94 43.59 20.88% numa05.sh Sys: 144.29 438.56 278.80 105.78 40.77% numa05.sh User: 33255.86 36890.82 34879.31 1641.98 12.11% Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mel Gorman <mgorman@techsingularity.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1529514181-9842-3-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-20 17:02:43 +00:00
* the value is, the more remote accesses that would be expected to
* be incurred if the tasks were swapped.
*
sched/numa: Evaluate move once per node task_numa_compare() helps choose the best CPU to move or swap the selected task. To achieve this task_numa_compare() is called for every CPU in the node. Currently it evaluates if the task can be moved/swapped for each of the CPUs. However the move evaluation is mostly independent of the CPU. Evaluating the move logic once per node, provides scope for simplifying task_numa_compare(). Running SPECjbb2005 on a 4 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 16 25705.2 25058.2 -2.51 1 74433 72950 -1.99 Running SPECjbb2005 on a 16 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 8 96589.6 105930 9.670 1 181830 178624 -1.76 (numbers from v1 based on v4.17-rc5) Testcase Time: Min Max Avg StdDev numa01.sh Real: 440.65 941.32 758.98 189.17 numa01.sh Sys: 183.48 320.07 258.42 50.09 numa01.sh User: 37384.65 71818.14 60302.51 13798.96 numa02.sh Real: 61.24 65.35 62.49 1.49 numa02.sh Sys: 16.83 24.18 21.40 2.60 numa02.sh User: 5219.59 5356.34 5264.03 49.07 numa03.sh Real: 822.04 912.40 873.55 37.35 numa03.sh Sys: 118.80 140.94 132.90 7.60 numa03.sh User: 62485.19 70025.01 67208.33 2967.10 numa04.sh Real: 690.66 872.12 778.49 65.44 numa04.sh Sys: 459.26 563.03 494.03 42.39 numa04.sh User: 51116.44 70527.20 58849.44 8461.28 numa05.sh Real: 418.37 562.28 525.77 54.27 numa05.sh Sys: 299.45 481.00 392.49 64.27 numa05.sh User: 34115.09 41324.02 39105.30 2627.68 Testcase Time: Min Max Avg StdDev %Change numa01.sh Real: 516.14 892.41 739.84 151.32 2.587% numa01.sh Sys: 153.16 192.99 177.70 14.58 45.42% numa01.sh User: 39821.04 69528.92 57193.87 10989.48 5.435% numa02.sh Real: 60.91 62.35 61.58 0.63 1.477% numa02.sh Sys: 16.47 26.16 21.20 3.85 0.943% numa02.sh User: 5227.58 5309.61 5265.17 31.04 -0.02% numa03.sh Real: 739.07 917.73 795.75 64.45 9.776% numa03.sh Sys: 94.46 136.08 109.48 14.58 21.39% numa03.sh User: 57478.56 72014.09 61764.48 5343.69 8.813% numa04.sh Real: 442.61 715.43 530.31 96.12 46.79% numa04.sh Sys: 224.90 348.63 285.61 48.83 72.97% numa04.sh User: 35836.84 47522.47 40235.41 3985.26 46.26% numa05.sh Real: 386.13 489.17 434.94 43.59 20.88% numa05.sh Sys: 144.29 438.56 278.80 105.78 40.77% numa05.sh User: 33255.86 36890.82 34879.31 1641.98 12.11% Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mel Gorman <mgorman@techsingularity.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1529514181-9842-3-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-20 17:02:43 +00:00
* If dst and source tasks are in the same NUMA group, or not
* in any group then look only at task weights.
*/
cur_ng = rcu_dereference(cur->numa_group);
if (cur_ng == p_ng) {
sched/numa: Evaluate move once per node task_numa_compare() helps choose the best CPU to move or swap the selected task. To achieve this task_numa_compare() is called for every CPU in the node. Currently it evaluates if the task can be moved/swapped for each of the CPUs. However the move evaluation is mostly independent of the CPU. Evaluating the move logic once per node, provides scope for simplifying task_numa_compare(). Running SPECjbb2005 on a 4 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 16 25705.2 25058.2 -2.51 1 74433 72950 -1.99 Running SPECjbb2005 on a 16 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 8 96589.6 105930 9.670 1 181830 178624 -1.76 (numbers from v1 based on v4.17-rc5) Testcase Time: Min Max Avg StdDev numa01.sh Real: 440.65 941.32 758.98 189.17 numa01.sh Sys: 183.48 320.07 258.42 50.09 numa01.sh User: 37384.65 71818.14 60302.51 13798.96 numa02.sh Real: 61.24 65.35 62.49 1.49 numa02.sh Sys: 16.83 24.18 21.40 2.60 numa02.sh User: 5219.59 5356.34 5264.03 49.07 numa03.sh Real: 822.04 912.40 873.55 37.35 numa03.sh Sys: 118.80 140.94 132.90 7.60 numa03.sh User: 62485.19 70025.01 67208.33 2967.10 numa04.sh Real: 690.66 872.12 778.49 65.44 numa04.sh Sys: 459.26 563.03 494.03 42.39 numa04.sh User: 51116.44 70527.20 58849.44 8461.28 numa05.sh Real: 418.37 562.28 525.77 54.27 numa05.sh Sys: 299.45 481.00 392.49 64.27 numa05.sh User: 34115.09 41324.02 39105.30 2627.68 Testcase Time: Min Max Avg StdDev %Change numa01.sh Real: 516.14 892.41 739.84 151.32 2.587% numa01.sh Sys: 153.16 192.99 177.70 14.58 45.42% numa01.sh User: 39821.04 69528.92 57193.87 10989.48 5.435% numa02.sh Real: 60.91 62.35 61.58 0.63 1.477% numa02.sh Sys: 16.47 26.16 21.20 3.85 0.943% numa02.sh User: 5227.58 5309.61 5265.17 31.04 -0.02% numa03.sh Real: 739.07 917.73 795.75 64.45 9.776% numa03.sh Sys: 94.46 136.08 109.48 14.58 21.39% numa03.sh User: 57478.56 72014.09 61764.48 5343.69 8.813% numa04.sh Real: 442.61 715.43 530.31 96.12 46.79% numa04.sh Sys: 224.90 348.63 285.61 48.83 72.97% numa04.sh User: 35836.84 47522.47 40235.41 3985.26 46.26% numa05.sh Real: 386.13 489.17 434.94 43.59 20.88% numa05.sh Sys: 144.29 438.56 278.80 105.78 40.77% numa05.sh User: 33255.86 36890.82 34879.31 1641.98 12.11% Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mel Gorman <mgorman@techsingularity.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1529514181-9842-3-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-20 17:02:43 +00:00
imp = taskimp + task_weight(cur, env->src_nid, dist) -
task_weight(cur, env->dst_nid, dist);
/*
sched/numa: Evaluate move once per node task_numa_compare() helps choose the best CPU to move or swap the selected task. To achieve this task_numa_compare() is called for every CPU in the node. Currently it evaluates if the task can be moved/swapped for each of the CPUs. However the move evaluation is mostly independent of the CPU. Evaluating the move logic once per node, provides scope for simplifying task_numa_compare(). Running SPECjbb2005 on a 4 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 16 25705.2 25058.2 -2.51 1 74433 72950 -1.99 Running SPECjbb2005 on a 16 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 8 96589.6 105930 9.670 1 181830 178624 -1.76 (numbers from v1 based on v4.17-rc5) Testcase Time: Min Max Avg StdDev numa01.sh Real: 440.65 941.32 758.98 189.17 numa01.sh Sys: 183.48 320.07 258.42 50.09 numa01.sh User: 37384.65 71818.14 60302.51 13798.96 numa02.sh Real: 61.24 65.35 62.49 1.49 numa02.sh Sys: 16.83 24.18 21.40 2.60 numa02.sh User: 5219.59 5356.34 5264.03 49.07 numa03.sh Real: 822.04 912.40 873.55 37.35 numa03.sh Sys: 118.80 140.94 132.90 7.60 numa03.sh User: 62485.19 70025.01 67208.33 2967.10 numa04.sh Real: 690.66 872.12 778.49 65.44 numa04.sh Sys: 459.26 563.03 494.03 42.39 numa04.sh User: 51116.44 70527.20 58849.44 8461.28 numa05.sh Real: 418.37 562.28 525.77 54.27 numa05.sh Sys: 299.45 481.00 392.49 64.27 numa05.sh User: 34115.09 41324.02 39105.30 2627.68 Testcase Time: Min Max Avg StdDev %Change numa01.sh Real: 516.14 892.41 739.84 151.32 2.587% numa01.sh Sys: 153.16 192.99 177.70 14.58 45.42% numa01.sh User: 39821.04 69528.92 57193.87 10989.48 5.435% numa02.sh Real: 60.91 62.35 61.58 0.63 1.477% numa02.sh Sys: 16.47 26.16 21.20 3.85 0.943% numa02.sh User: 5227.58 5309.61 5265.17 31.04 -0.02% numa03.sh Real: 739.07 917.73 795.75 64.45 9.776% numa03.sh Sys: 94.46 136.08 109.48 14.58 21.39% numa03.sh User: 57478.56 72014.09 61764.48 5343.69 8.813% numa04.sh Real: 442.61 715.43 530.31 96.12 46.79% numa04.sh Sys: 224.90 348.63 285.61 48.83 72.97% numa04.sh User: 35836.84 47522.47 40235.41 3985.26 46.26% numa05.sh Real: 386.13 489.17 434.94 43.59 20.88% numa05.sh Sys: 144.29 438.56 278.80 105.78 40.77% numa05.sh User: 33255.86 36890.82 34879.31 1641.98 12.11% Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mel Gorman <mgorman@techsingularity.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1529514181-9842-3-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-20 17:02:43 +00:00
* Add some hysteresis to prevent swapping the
* tasks within a group over tiny differences.
*/
if (cur_ng)
sched/numa: Evaluate move once per node task_numa_compare() helps choose the best CPU to move or swap the selected task. To achieve this task_numa_compare() is called for every CPU in the node. Currently it evaluates if the task can be moved/swapped for each of the CPUs. However the move evaluation is mostly independent of the CPU. Evaluating the move logic once per node, provides scope for simplifying task_numa_compare(). Running SPECjbb2005 on a 4 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 16 25705.2 25058.2 -2.51 1 74433 72950 -1.99 Running SPECjbb2005 on a 16 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 8 96589.6 105930 9.670 1 181830 178624 -1.76 (numbers from v1 based on v4.17-rc5) Testcase Time: Min Max Avg StdDev numa01.sh Real: 440.65 941.32 758.98 189.17 numa01.sh Sys: 183.48 320.07 258.42 50.09 numa01.sh User: 37384.65 71818.14 60302.51 13798.96 numa02.sh Real: 61.24 65.35 62.49 1.49 numa02.sh Sys: 16.83 24.18 21.40 2.60 numa02.sh User: 5219.59 5356.34 5264.03 49.07 numa03.sh Real: 822.04 912.40 873.55 37.35 numa03.sh Sys: 118.80 140.94 132.90 7.60 numa03.sh User: 62485.19 70025.01 67208.33 2967.10 numa04.sh Real: 690.66 872.12 778.49 65.44 numa04.sh Sys: 459.26 563.03 494.03 42.39 numa04.sh User: 51116.44 70527.20 58849.44 8461.28 numa05.sh Real: 418.37 562.28 525.77 54.27 numa05.sh Sys: 299.45 481.00 392.49 64.27 numa05.sh User: 34115.09 41324.02 39105.30 2627.68 Testcase Time: Min Max Avg StdDev %Change numa01.sh Real: 516.14 892.41 739.84 151.32 2.587% numa01.sh Sys: 153.16 192.99 177.70 14.58 45.42% numa01.sh User: 39821.04 69528.92 57193.87 10989.48 5.435% numa02.sh Real: 60.91 62.35 61.58 0.63 1.477% numa02.sh Sys: 16.47 26.16 21.20 3.85 0.943% numa02.sh User: 5227.58 5309.61 5265.17 31.04 -0.02% numa03.sh Real: 739.07 917.73 795.75 64.45 9.776% numa03.sh Sys: 94.46 136.08 109.48 14.58 21.39% numa03.sh User: 57478.56 72014.09 61764.48 5343.69 8.813% numa04.sh Real: 442.61 715.43 530.31 96.12 46.79% numa04.sh Sys: 224.90 348.63 285.61 48.83 72.97% numa04.sh User: 35836.84 47522.47 40235.41 3985.26 46.26% numa05.sh Real: 386.13 489.17 434.94 43.59 20.88% numa05.sh Sys: 144.29 438.56 278.80 105.78 40.77% numa05.sh User: 33255.86 36890.82 34879.31 1641.98 12.11% Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mel Gorman <mgorman@techsingularity.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1529514181-9842-3-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-20 17:02:43 +00:00
imp -= imp / 16;
} else {
/*
* Compare the group weights. If a task is all by itself
* (not part of a group), use the task weight instead.
*/
if (cur_ng && p_ng)
sched/numa: Evaluate move once per node task_numa_compare() helps choose the best CPU to move or swap the selected task. To achieve this task_numa_compare() is called for every CPU in the node. Currently it evaluates if the task can be moved/swapped for each of the CPUs. However the move evaluation is mostly independent of the CPU. Evaluating the move logic once per node, provides scope for simplifying task_numa_compare(). Running SPECjbb2005 on a 4 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 16 25705.2 25058.2 -2.51 1 74433 72950 -1.99 Running SPECjbb2005 on a 16 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 8 96589.6 105930 9.670 1 181830 178624 -1.76 (numbers from v1 based on v4.17-rc5) Testcase Time: Min Max Avg StdDev numa01.sh Real: 440.65 941.32 758.98 189.17 numa01.sh Sys: 183.48 320.07 258.42 50.09 numa01.sh User: 37384.65 71818.14 60302.51 13798.96 numa02.sh Real: 61.24 65.35 62.49 1.49 numa02.sh Sys: 16.83 24.18 21.40 2.60 numa02.sh User: 5219.59 5356.34 5264.03 49.07 numa03.sh Real: 822.04 912.40 873.55 37.35 numa03.sh Sys: 118.80 140.94 132.90 7.60 numa03.sh User: 62485.19 70025.01 67208.33 2967.10 numa04.sh Real: 690.66 872.12 778.49 65.44 numa04.sh Sys: 459.26 563.03 494.03 42.39 numa04.sh User: 51116.44 70527.20 58849.44 8461.28 numa05.sh Real: 418.37 562.28 525.77 54.27 numa05.sh Sys: 299.45 481.00 392.49 64.27 numa05.sh User: 34115.09 41324.02 39105.30 2627.68 Testcase Time: Min Max Avg StdDev %Change numa01.sh Real: 516.14 892.41 739.84 151.32 2.587% numa01.sh Sys: 153.16 192.99 177.70 14.58 45.42% numa01.sh User: 39821.04 69528.92 57193.87 10989.48 5.435% numa02.sh Real: 60.91 62.35 61.58 0.63 1.477% numa02.sh Sys: 16.47 26.16 21.20 3.85 0.943% numa02.sh User: 5227.58 5309.61 5265.17 31.04 -0.02% numa03.sh Real: 739.07 917.73 795.75 64.45 9.776% numa03.sh Sys: 94.46 136.08 109.48 14.58 21.39% numa03.sh User: 57478.56 72014.09 61764.48 5343.69 8.813% numa04.sh Real: 442.61 715.43 530.31 96.12 46.79% numa04.sh Sys: 224.90 348.63 285.61 48.83 72.97% numa04.sh User: 35836.84 47522.47 40235.41 3985.26 46.26% numa05.sh Real: 386.13 489.17 434.94 43.59 20.88% numa05.sh Sys: 144.29 438.56 278.80 105.78 40.77% numa05.sh User: 33255.86 36890.82 34879.31 1641.98 12.11% Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mel Gorman <mgorman@techsingularity.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1529514181-9842-3-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-20 17:02:43 +00:00
imp += group_weight(cur, env->src_nid, dist) -
group_weight(cur, env->dst_nid, dist);
else
imp += task_weight(cur, env->src_nid, dist) -
task_weight(cur, env->dst_nid, dist);
}
/* Discourage picking a task already on its preferred node */
if (cur->numa_preferred_nid == env->dst_nid)
imp -= imp / 16;
/*
* Encourage picking a task that moves to its preferred node.
* This potentially makes imp larger than it's maximum of
* 1998 (see SMALLIMP and task_weight for why) but in this
* case, it does not matter.
*/
if (cur->numa_preferred_nid == env->src_nid)
imp += imp / 8;
sched/numa: Evaluate move once per node task_numa_compare() helps choose the best CPU to move or swap the selected task. To achieve this task_numa_compare() is called for every CPU in the node. Currently it evaluates if the task can be moved/swapped for each of the CPUs. However the move evaluation is mostly independent of the CPU. Evaluating the move logic once per node, provides scope for simplifying task_numa_compare(). Running SPECjbb2005 on a 4 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 16 25705.2 25058.2 -2.51 1 74433 72950 -1.99 Running SPECjbb2005 on a 16 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 8 96589.6 105930 9.670 1 181830 178624 -1.76 (numbers from v1 based on v4.17-rc5) Testcase Time: Min Max Avg StdDev numa01.sh Real: 440.65 941.32 758.98 189.17 numa01.sh Sys: 183.48 320.07 258.42 50.09 numa01.sh User: 37384.65 71818.14 60302.51 13798.96 numa02.sh Real: 61.24 65.35 62.49 1.49 numa02.sh Sys: 16.83 24.18 21.40 2.60 numa02.sh User: 5219.59 5356.34 5264.03 49.07 numa03.sh Real: 822.04 912.40 873.55 37.35 numa03.sh Sys: 118.80 140.94 132.90 7.60 numa03.sh User: 62485.19 70025.01 67208.33 2967.10 numa04.sh Real: 690.66 872.12 778.49 65.44 numa04.sh Sys: 459.26 563.03 494.03 42.39 numa04.sh User: 51116.44 70527.20 58849.44 8461.28 numa05.sh Real: 418.37 562.28 525.77 54.27 numa05.sh Sys: 299.45 481.00 392.49 64.27 numa05.sh User: 34115.09 41324.02 39105.30 2627.68 Testcase Time: Min Max Avg StdDev %Change numa01.sh Real: 516.14 892.41 739.84 151.32 2.587% numa01.sh Sys: 153.16 192.99 177.70 14.58 45.42% numa01.sh User: 39821.04 69528.92 57193.87 10989.48 5.435% numa02.sh Real: 60.91 62.35 61.58 0.63 1.477% numa02.sh Sys: 16.47 26.16 21.20 3.85 0.943% numa02.sh User: 5227.58 5309.61 5265.17 31.04 -0.02% numa03.sh Real: 739.07 917.73 795.75 64.45 9.776% numa03.sh Sys: 94.46 136.08 109.48 14.58 21.39% numa03.sh User: 57478.56 72014.09 61764.48 5343.69 8.813% numa04.sh Real: 442.61 715.43 530.31 96.12 46.79% numa04.sh Sys: 224.90 348.63 285.61 48.83 72.97% numa04.sh User: 35836.84 47522.47 40235.41 3985.26 46.26% numa05.sh Real: 386.13 489.17 434.94 43.59 20.88% numa05.sh Sys: 144.29 438.56 278.80 105.78 40.77% numa05.sh User: 33255.86 36890.82 34879.31 1641.98 12.11% Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mel Gorman <mgorman@techsingularity.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1529514181-9842-3-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-20 17:02:43 +00:00
if (maymove && moveimp > imp && moveimp > env->best_imp) {
sched/numa: Avoid task migration for small NUMA improvement If NUMA improvement from the task migration is going to be very minimal, then avoid task migration. Specjbb2005 results (8 warehouses) Higher bops are better 2 Socket - 2 Node Haswell - X86 JVMS Prev Current %Change 4 198512 205910 3.72673 1 313559 318491 1.57291 2 Socket - 4 Node Power8 - PowerNV JVMS Prev Current %Change 8 74761.9 74935.9 0.232739 1 214874 226796 5.54837 2 Socket - 2 Node Power9 - PowerNV JVMS Prev Current %Change 4 180536 189780 5.12031 1 210281 205695 -2.18089 4 Socket - 4 Node Power7 - PowerVM JVMS Prev Current %Change 8 56511.4 60370 6.828 1 104899 108100 3.05151 1/7 cases is regressing, if we look at events migrate_pages seem to vary the most especially in the regressing case. Also some amount of variance is expected between different runs of Specjbb2005. Some events stats before and after applying the patch. perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 13,818,546 13,801,554 migrations 1,149,960 1,151,541 faults 385,583 433,246 cache-misses 55,259,546,768 55,168,691,835 sched:sched_move_numa 2,257 2,551 sched:sched_stick_numa 9 24 sched:sched_swap_numa 512 904 migrate:mm_migrate_pages 2,225 1,571 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 72692 113682 numa_hint_faults_local 62270 102163 numa_hit 238762 240181 numa_huge_pte_updates 48 36 numa_interleave 75 64 numa_local 238676 240103 numa_other 86 78 numa_pages_migrated 2225 1564 numa_pte_updates 98557 134080 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 3,173,490 3,079,150 migrations 36,966 31,455 faults 108,776 99,081 cache-misses 12,200,075,320 11,588,126,740 sched:sched_move_numa 1,264 1 sched:sched_stick_numa 0 0 sched:sched_swap_numa 0 0 migrate:mm_migrate_pages 899 36 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 21109 430 numa_hint_faults_local 17120 77 numa_hit 72934 71277 numa_huge_pte_updates 42 0 numa_interleave 33 22 numa_local 72866 71218 numa_other 68 59 numa_pages_migrated 915 23 numa_pte_updates 42326 0 perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 8,312,022 8,707,565 migrations 231,705 171,342 faults 310,242 310,820 cache-misses 402,324,573 136,115,400 sched:sched_move_numa 193 215 sched:sched_stick_numa 0 6 sched:sched_swap_numa 3 24 migrate:mm_migrate_pages 93 162 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 11838 8985 numa_hint_faults_local 11216 8154 numa_hit 90689 93819 numa_huge_pte_updates 0 0 numa_interleave 1579 882 numa_local 89634 93496 numa_other 1055 323 numa_pages_migrated 92 169 numa_pte_updates 12109 9217 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 2,170,481 2,152,072 migrations 10,126 10,704 faults 160,962 164,376 cache-misses 10,834,845 3,818,437 sched:sched_move_numa 10 16 sched:sched_stick_numa 0 0 sched:sched_swap_numa 0 7 migrate:mm_migrate_pages 2 199 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 403 2248 numa_hint_faults_local 358 1666 numa_hit 25898 25704 numa_huge_pte_updates 0 0 numa_interleave 207 200 numa_local 25860 25679 numa_other 38 25 numa_pages_migrated 2 197 numa_pte_updates 400 2234 perf stats 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 110,339,633 93,330,595 migrations 4,139,812 4,122,061 faults 863,622 865,979 cache-misses 231,838,045,660 225,395,083,479 sched:sched_move_numa 2,196 2,372 sched:sched_stick_numa 33 24 sched:sched_swap_numa 544 769 migrate:mm_migrate_pages 2,469 1,677 vmstat 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 85748 91638 numa_hint_faults_local 66831 78096 numa_hit 242213 242225 numa_huge_pte_updates 0 0 numa_interleave 0 2 numa_local 242211 242219 numa_other 2 6 numa_pages_migrated 2376 1515 numa_pte_updates 86233 92274 perf stats 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 59,331,057 51,487,271 migrations 552,019 537,170 faults 266,586 256,921 cache-misses 73,796,312,990 70,073,831,187 sched:sched_move_numa 981 576 sched:sched_stick_numa 54 24 sched:sched_swap_numa 286 327 migrate:mm_migrate_pages 713 726 vmstat 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 14807 12000 numa_hint_faults_local 5738 5024 numa_hit 36230 36470 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 36228 36465 numa_other 2 5 numa_pages_migrated 703 726 numa_pte_updates 14742 11930 Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Jirka Hladky <jhladky@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1537552141-27815-7-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-09-21 17:49:01 +00:00
imp = moveimp;
sched/numa: Evaluate move once per node task_numa_compare() helps choose the best CPU to move or swap the selected task. To achieve this task_numa_compare() is called for every CPU in the node. Currently it evaluates if the task can be moved/swapped for each of the CPUs. However the move evaluation is mostly independent of the CPU. Evaluating the move logic once per node, provides scope for simplifying task_numa_compare(). Running SPECjbb2005 on a 4 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 16 25705.2 25058.2 -2.51 1 74433 72950 -1.99 Running SPECjbb2005 on a 16 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 8 96589.6 105930 9.670 1 181830 178624 -1.76 (numbers from v1 based on v4.17-rc5) Testcase Time: Min Max Avg StdDev numa01.sh Real: 440.65 941.32 758.98 189.17 numa01.sh Sys: 183.48 320.07 258.42 50.09 numa01.sh User: 37384.65 71818.14 60302.51 13798.96 numa02.sh Real: 61.24 65.35 62.49 1.49 numa02.sh Sys: 16.83 24.18 21.40 2.60 numa02.sh User: 5219.59 5356.34 5264.03 49.07 numa03.sh Real: 822.04 912.40 873.55 37.35 numa03.sh Sys: 118.80 140.94 132.90 7.60 numa03.sh User: 62485.19 70025.01 67208.33 2967.10 numa04.sh Real: 690.66 872.12 778.49 65.44 numa04.sh Sys: 459.26 563.03 494.03 42.39 numa04.sh User: 51116.44 70527.20 58849.44 8461.28 numa05.sh Real: 418.37 562.28 525.77 54.27 numa05.sh Sys: 299.45 481.00 392.49 64.27 numa05.sh User: 34115.09 41324.02 39105.30 2627.68 Testcase Time: Min Max Avg StdDev %Change numa01.sh Real: 516.14 892.41 739.84 151.32 2.587% numa01.sh Sys: 153.16 192.99 177.70 14.58 45.42% numa01.sh User: 39821.04 69528.92 57193.87 10989.48 5.435% numa02.sh Real: 60.91 62.35 61.58 0.63 1.477% numa02.sh Sys: 16.47 26.16 21.20 3.85 0.943% numa02.sh User: 5227.58 5309.61 5265.17 31.04 -0.02% numa03.sh Real: 739.07 917.73 795.75 64.45 9.776% numa03.sh Sys: 94.46 136.08 109.48 14.58 21.39% numa03.sh User: 57478.56 72014.09 61764.48 5343.69 8.813% numa04.sh Real: 442.61 715.43 530.31 96.12 46.79% numa04.sh Sys: 224.90 348.63 285.61 48.83 72.97% numa04.sh User: 35836.84 47522.47 40235.41 3985.26 46.26% numa05.sh Real: 386.13 489.17 434.94 43.59 20.88% numa05.sh Sys: 144.29 438.56 278.80 105.78 40.77% numa05.sh User: 33255.86 36890.82 34879.31 1641.98 12.11% Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mel Gorman <mgorman@techsingularity.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1529514181-9842-3-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-20 17:02:43 +00:00
cur = NULL;
goto assign;
sched/numa: Evaluate move once per node task_numa_compare() helps choose the best CPU to move or swap the selected task. To achieve this task_numa_compare() is called for every CPU in the node. Currently it evaluates if the task can be moved/swapped for each of the CPUs. However the move evaluation is mostly independent of the CPU. Evaluating the move logic once per node, provides scope for simplifying task_numa_compare(). Running SPECjbb2005 on a 4 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 16 25705.2 25058.2 -2.51 1 74433 72950 -1.99 Running SPECjbb2005 on a 16 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 8 96589.6 105930 9.670 1 181830 178624 -1.76 (numbers from v1 based on v4.17-rc5) Testcase Time: Min Max Avg StdDev numa01.sh Real: 440.65 941.32 758.98 189.17 numa01.sh Sys: 183.48 320.07 258.42 50.09 numa01.sh User: 37384.65 71818.14 60302.51 13798.96 numa02.sh Real: 61.24 65.35 62.49 1.49 numa02.sh Sys: 16.83 24.18 21.40 2.60 numa02.sh User: 5219.59 5356.34 5264.03 49.07 numa03.sh Real: 822.04 912.40 873.55 37.35 numa03.sh Sys: 118.80 140.94 132.90 7.60 numa03.sh User: 62485.19 70025.01 67208.33 2967.10 numa04.sh Real: 690.66 872.12 778.49 65.44 numa04.sh Sys: 459.26 563.03 494.03 42.39 numa04.sh User: 51116.44 70527.20 58849.44 8461.28 numa05.sh Real: 418.37 562.28 525.77 54.27 numa05.sh Sys: 299.45 481.00 392.49 64.27 numa05.sh User: 34115.09 41324.02 39105.30 2627.68 Testcase Time: Min Max Avg StdDev %Change numa01.sh Real: 516.14 892.41 739.84 151.32 2.587% numa01.sh Sys: 153.16 192.99 177.70 14.58 45.42% numa01.sh User: 39821.04 69528.92 57193.87 10989.48 5.435% numa02.sh Real: 60.91 62.35 61.58 0.63 1.477% numa02.sh Sys: 16.47 26.16 21.20 3.85 0.943% numa02.sh User: 5227.58 5309.61 5265.17 31.04 -0.02% numa03.sh Real: 739.07 917.73 795.75 64.45 9.776% numa03.sh Sys: 94.46 136.08 109.48 14.58 21.39% numa03.sh User: 57478.56 72014.09 61764.48 5343.69 8.813% numa04.sh Real: 442.61 715.43 530.31 96.12 46.79% numa04.sh Sys: 224.90 348.63 285.61 48.83 72.97% numa04.sh User: 35836.84 47522.47 40235.41 3985.26 46.26% numa05.sh Real: 386.13 489.17 434.94 43.59 20.88% numa05.sh Sys: 144.29 438.56 278.80 105.78 40.77% numa05.sh User: 33255.86 36890.82 34879.31 1641.98 12.11% Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mel Gorman <mgorman@techsingularity.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1529514181-9842-3-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-20 17:02:43 +00:00
}
/*
* Prefer swapping with a task moving to its preferred node over a
* task that is not.
*/
if (env->best_task && cur->numa_preferred_nid == env->src_nid &&
env->best_task->numa_preferred_nid != env->src_nid) {
goto assign;
}
sched/numa: Avoid task migration for small NUMA improvement If NUMA improvement from the task migration is going to be very minimal, then avoid task migration. Specjbb2005 results (8 warehouses) Higher bops are better 2 Socket - 2 Node Haswell - X86 JVMS Prev Current %Change 4 198512 205910 3.72673 1 313559 318491 1.57291 2 Socket - 4 Node Power8 - PowerNV JVMS Prev Current %Change 8 74761.9 74935.9 0.232739 1 214874 226796 5.54837 2 Socket - 2 Node Power9 - PowerNV JVMS Prev Current %Change 4 180536 189780 5.12031 1 210281 205695 -2.18089 4 Socket - 4 Node Power7 - PowerVM JVMS Prev Current %Change 8 56511.4 60370 6.828 1 104899 108100 3.05151 1/7 cases is regressing, if we look at events migrate_pages seem to vary the most especially in the regressing case. Also some amount of variance is expected between different runs of Specjbb2005. Some events stats before and after applying the patch. perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 13,818,546 13,801,554 migrations 1,149,960 1,151,541 faults 385,583 433,246 cache-misses 55,259,546,768 55,168,691,835 sched:sched_move_numa 2,257 2,551 sched:sched_stick_numa 9 24 sched:sched_swap_numa 512 904 migrate:mm_migrate_pages 2,225 1,571 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 72692 113682 numa_hint_faults_local 62270 102163 numa_hit 238762 240181 numa_huge_pte_updates 48 36 numa_interleave 75 64 numa_local 238676 240103 numa_other 86 78 numa_pages_migrated 2225 1564 numa_pte_updates 98557 134080 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 3,173,490 3,079,150 migrations 36,966 31,455 faults 108,776 99,081 cache-misses 12,200,075,320 11,588,126,740 sched:sched_move_numa 1,264 1 sched:sched_stick_numa 0 0 sched:sched_swap_numa 0 0 migrate:mm_migrate_pages 899 36 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 21109 430 numa_hint_faults_local 17120 77 numa_hit 72934 71277 numa_huge_pte_updates 42 0 numa_interleave 33 22 numa_local 72866 71218 numa_other 68 59 numa_pages_migrated 915 23 numa_pte_updates 42326 0 perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 8,312,022 8,707,565 migrations 231,705 171,342 faults 310,242 310,820 cache-misses 402,324,573 136,115,400 sched:sched_move_numa 193 215 sched:sched_stick_numa 0 6 sched:sched_swap_numa 3 24 migrate:mm_migrate_pages 93 162 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 11838 8985 numa_hint_faults_local 11216 8154 numa_hit 90689 93819 numa_huge_pte_updates 0 0 numa_interleave 1579 882 numa_local 89634 93496 numa_other 1055 323 numa_pages_migrated 92 169 numa_pte_updates 12109 9217 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 2,170,481 2,152,072 migrations 10,126 10,704 faults 160,962 164,376 cache-misses 10,834,845 3,818,437 sched:sched_move_numa 10 16 sched:sched_stick_numa 0 0 sched:sched_swap_numa 0 7 migrate:mm_migrate_pages 2 199 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 403 2248 numa_hint_faults_local 358 1666 numa_hit 25898 25704 numa_huge_pte_updates 0 0 numa_interleave 207 200 numa_local 25860 25679 numa_other 38 25 numa_pages_migrated 2 197 numa_pte_updates 400 2234 perf stats 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 110,339,633 93,330,595 migrations 4,139,812 4,122,061 faults 863,622 865,979 cache-misses 231,838,045,660 225,395,083,479 sched:sched_move_numa 2,196 2,372 sched:sched_stick_numa 33 24 sched:sched_swap_numa 544 769 migrate:mm_migrate_pages 2,469 1,677 vmstat 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 85748 91638 numa_hint_faults_local 66831 78096 numa_hit 242213 242225 numa_huge_pte_updates 0 0 numa_interleave 0 2 numa_local 242211 242219 numa_other 2 6 numa_pages_migrated 2376 1515 numa_pte_updates 86233 92274 perf stats 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 59,331,057 51,487,271 migrations 552,019 537,170 faults 266,586 256,921 cache-misses 73,796,312,990 70,073,831,187 sched:sched_move_numa 981 576 sched:sched_stick_numa 54 24 sched:sched_swap_numa 286 327 migrate:mm_migrate_pages 713 726 vmstat 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 14807 12000 numa_hint_faults_local 5738 5024 numa_hit 36230 36470 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 36228 36465 numa_other 2 5 numa_pages_migrated 703 726 numa_pte_updates 14742 11930 Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Jirka Hladky <jhladky@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1537552141-27815-7-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-09-21 17:49:01 +00:00
/*
* If the NUMA importance is less than SMALLIMP,
* task migration might only result in ping pong
* of tasks and also hurt performance due to cache
* misses.
*/
if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
goto unlock;
/*
* In the overloaded case, try and keep the load balanced.
*/
sched/numa: Evaluate move once per node task_numa_compare() helps choose the best CPU to move or swap the selected task. To achieve this task_numa_compare() is called for every CPU in the node. Currently it evaluates if the task can be moved/swapped for each of the CPUs. However the move evaluation is mostly independent of the CPU. Evaluating the move logic once per node, provides scope for simplifying task_numa_compare(). Running SPECjbb2005 on a 4 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 16 25705.2 25058.2 -2.51 1 74433 72950 -1.99 Running SPECjbb2005 on a 16 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 8 96589.6 105930 9.670 1 181830 178624 -1.76 (numbers from v1 based on v4.17-rc5) Testcase Time: Min Max Avg StdDev numa01.sh Real: 440.65 941.32 758.98 189.17 numa01.sh Sys: 183.48 320.07 258.42 50.09 numa01.sh User: 37384.65 71818.14 60302.51 13798.96 numa02.sh Real: 61.24 65.35 62.49 1.49 numa02.sh Sys: 16.83 24.18 21.40 2.60 numa02.sh User: 5219.59 5356.34 5264.03 49.07 numa03.sh Real: 822.04 912.40 873.55 37.35 numa03.sh Sys: 118.80 140.94 132.90 7.60 numa03.sh User: 62485.19 70025.01 67208.33 2967.10 numa04.sh Real: 690.66 872.12 778.49 65.44 numa04.sh Sys: 459.26 563.03 494.03 42.39 numa04.sh User: 51116.44 70527.20 58849.44 8461.28 numa05.sh Real: 418.37 562.28 525.77 54.27 numa05.sh Sys: 299.45 481.00 392.49 64.27 numa05.sh User: 34115.09 41324.02 39105.30 2627.68 Testcase Time: Min Max Avg StdDev %Change numa01.sh Real: 516.14 892.41 739.84 151.32 2.587% numa01.sh Sys: 153.16 192.99 177.70 14.58 45.42% numa01.sh User: 39821.04 69528.92 57193.87 10989.48 5.435% numa02.sh Real: 60.91 62.35 61.58 0.63 1.477% numa02.sh Sys: 16.47 26.16 21.20 3.85 0.943% numa02.sh User: 5227.58 5309.61 5265.17 31.04 -0.02% numa03.sh Real: 739.07 917.73 795.75 64.45 9.776% numa03.sh Sys: 94.46 136.08 109.48 14.58 21.39% numa03.sh User: 57478.56 72014.09 61764.48 5343.69 8.813% numa04.sh Real: 442.61 715.43 530.31 96.12 46.79% numa04.sh Sys: 224.90 348.63 285.61 48.83 72.97% numa04.sh User: 35836.84 47522.47 40235.41 3985.26 46.26% numa05.sh Real: 386.13 489.17 434.94 43.59 20.88% numa05.sh Sys: 144.29 438.56 278.80 105.78 40.77% numa05.sh User: 33255.86 36890.82 34879.31 1641.98 12.11% Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mel Gorman <mgorman@techsingularity.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1529514181-9842-3-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-20 17:02:43 +00:00
load = task_h_load(env->p) - task_h_load(cur);
if (!load)
goto assign;
dst_load = env->dst_stats.load + load;
src_load = env->src_stats.load - load;
if (load_too_imbalanced(src_load, dst_load, env))
goto unlock;
sched/numa: Evaluate move once per node task_numa_compare() helps choose the best CPU to move or swap the selected task. To achieve this task_numa_compare() is called for every CPU in the node. Currently it evaluates if the task can be moved/swapped for each of the CPUs. However the move evaluation is mostly independent of the CPU. Evaluating the move logic once per node, provides scope for simplifying task_numa_compare(). Running SPECjbb2005 on a 4 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 16 25705.2 25058.2 -2.51 1 74433 72950 -1.99 Running SPECjbb2005 on a 16 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 8 96589.6 105930 9.670 1 181830 178624 -1.76 (numbers from v1 based on v4.17-rc5) Testcase Time: Min Max Avg StdDev numa01.sh Real: 440.65 941.32 758.98 189.17 numa01.sh Sys: 183.48 320.07 258.42 50.09 numa01.sh User: 37384.65 71818.14 60302.51 13798.96 numa02.sh Real: 61.24 65.35 62.49 1.49 numa02.sh Sys: 16.83 24.18 21.40 2.60 numa02.sh User: 5219.59 5356.34 5264.03 49.07 numa03.sh Real: 822.04 912.40 873.55 37.35 numa03.sh Sys: 118.80 140.94 132.90 7.60 numa03.sh User: 62485.19 70025.01 67208.33 2967.10 numa04.sh Real: 690.66 872.12 778.49 65.44 numa04.sh Sys: 459.26 563.03 494.03 42.39 numa04.sh User: 51116.44 70527.20 58849.44 8461.28 numa05.sh Real: 418.37 562.28 525.77 54.27 numa05.sh Sys: 299.45 481.00 392.49 64.27 numa05.sh User: 34115.09 41324.02 39105.30 2627.68 Testcase Time: Min Max Avg StdDev %Change numa01.sh Real: 516.14 892.41 739.84 151.32 2.587% numa01.sh Sys: 153.16 192.99 177.70 14.58 45.42% numa01.sh User: 39821.04 69528.92 57193.87 10989.48 5.435% numa02.sh Real: 60.91 62.35 61.58 0.63 1.477% numa02.sh Sys: 16.47 26.16 21.20 3.85 0.943% numa02.sh User: 5227.58 5309.61 5265.17 31.04 -0.02% numa03.sh Real: 739.07 917.73 795.75 64.45 9.776% numa03.sh Sys: 94.46 136.08 109.48 14.58 21.39% numa03.sh User: 57478.56 72014.09 61764.48 5343.69 8.813% numa04.sh Real: 442.61 715.43 530.31 96.12 46.79% numa04.sh Sys: 224.90 348.63 285.61 48.83 72.97% numa04.sh User: 35836.84 47522.47 40235.41 3985.26 46.26% numa05.sh Real: 386.13 489.17 434.94 43.59 20.88% numa05.sh Sys: 144.29 438.56 278.80 105.78 40.77% numa05.sh User: 33255.86 36890.82 34879.31 1641.98 12.11% Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mel Gorman <mgorman@techsingularity.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1529514181-9842-3-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-20 17:02:43 +00:00
assign:
sched/numa: Prefer using an idle CPU as a migration target instead of comparing tasks task_numa_find_cpu() can scan a node multiple times. Minimally it scans to gather statistics and later to find a suitable target. In some cases, the second scan will simply pick an idle CPU if the load is not imbalanced. This patch caches information on an idle core while gathering statistics and uses it immediately if load is not imbalanced to avoid a second scan of the node runqueues. Preference is given to an idle core rather than an idle SMT sibling to avoid packing HT siblings due to linearly scanning the node cpumask. As a side-effect, even when the second scan is necessary, the importance of using select_idle_sibling is much reduced because information on idle CPUs is cached and can be reused. Note that this patch actually makes is harder to move to an idle CPU as multiple tasks can race for the same idle CPU due to a race checking numa_migrate_on. This is addressed in the next patch. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Ingo Molnar <mingo@kernel.org> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Valentin Schneider <valentin.schneider@arm.com> Cc: Phil Auld <pauld@redhat.com> Cc: Hillf Danton <hdanton@sina.com> Link: https://lore.kernel.org/r/20200224095223.13361-11-mgorman@techsingularity.net
2020-02-24 09:52:20 +00:00
/* Evaluate an idle CPU for a task numa move. */
sched/core: Rewrite and improve select_idle_siblings() select_idle_siblings() is a known pain point for a number of workloads; it either does too much or not enough and sometimes just does plain wrong. This rewrite attempts to address a number of issues (but sadly not all). The current code does an unconditional sched_domain iteration; with the intent of finding an idle core (on SMT hardware). The problems which this patch tries to address are: - its pointless to look for idle cores if the machine is real busy; at which point you're just wasting cycles. - it's behaviour is inconsistent between SMT and !SMT hardware in that !SMT hardware ends up doing a scan for any idle CPU in the LLC domain, while SMT hardware does a scan for idle cores and if that fails, falls back to a scan for idle threads on the 'target' core. The new code replaces the sched_domain scan with 3 explicit scans: 1) search for an idle core in the LLC 2) search for an idle CPU in the LLC 3) search for an idle thread in the 'target' core where 1 and 3 are conditional on SMT support and 1 and 2 have runtime heuristics to skip the step. Step 1) is conditional on sd_llc_shared->has_idle_cores; when a cpu goes idle and sd_llc_shared->has_idle_cores is false, we scan all SMT siblings of the CPU going idle. Similarly, we clear sd_llc_shared->has_idle_cores when we fail to find an idle core. Step 2) tracks the average cost of the scan and compares this to the average idle time guestimate for the CPU doing the wakeup. There is a significant fudge factor involved to deal with the variability of the averages. Esp. hackbench was sensitive to this. Step 3) is unconditional; we assume (also per step 1) that scanning all SMT siblings in a core is 'cheap'. With this; SMT systems gain step 2, which cures a few benchmarks -- notably one from Facebook. One 'feature' of the sched_domain iteration, which we preserve in the new code, is that it would start scanning from the 'target' CPU, instead of scanning the cpumask in cpu id order. This avoids multiple CPUs in the LLC scanning for idle to gang up and find the same CPU quite as much. The down side is that tasks can end up hopping across the LLC for no apparent reason. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-05-09 08:38:05 +00:00
if (!cur) {
sched/numa: Prefer using an idle CPU as a migration target instead of comparing tasks task_numa_find_cpu() can scan a node multiple times. Minimally it scans to gather statistics and later to find a suitable target. In some cases, the second scan will simply pick an idle CPU if the load is not imbalanced. This patch caches information on an idle core while gathering statistics and uses it immediately if load is not imbalanced to avoid a second scan of the node runqueues. Preference is given to an idle core rather than an idle SMT sibling to avoid packing HT siblings due to linearly scanning the node cpumask. As a side-effect, even when the second scan is necessary, the importance of using select_idle_sibling is much reduced because information on idle CPUs is cached and can be reused. Note that this patch actually makes is harder to move to an idle CPU as multiple tasks can race for the same idle CPU due to a race checking numa_migrate_on. This is addressed in the next patch. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Ingo Molnar <mingo@kernel.org> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Valentin Schneider <valentin.schneider@arm.com> Cc: Phil Auld <pauld@redhat.com> Cc: Hillf Danton <hdanton@sina.com> Link: https://lore.kernel.org/r/20200224095223.13361-11-mgorman@techsingularity.net
2020-02-24 09:52:20 +00:00
int cpu = env->dst_stats.idle_cpu;
/* Nothing cached so current CPU went idle since the search. */
if (cpu < 0)
cpu = env->dst_cpu;
sched/core: Rewrite and improve select_idle_siblings() select_idle_siblings() is a known pain point for a number of workloads; it either does too much or not enough and sometimes just does plain wrong. This rewrite attempts to address a number of issues (but sadly not all). The current code does an unconditional sched_domain iteration; with the intent of finding an idle core (on SMT hardware). The problems which this patch tries to address are: - its pointless to look for idle cores if the machine is real busy; at which point you're just wasting cycles. - it's behaviour is inconsistent between SMT and !SMT hardware in that !SMT hardware ends up doing a scan for any idle CPU in the LLC domain, while SMT hardware does a scan for idle cores and if that fails, falls back to a scan for idle threads on the 'target' core. The new code replaces the sched_domain scan with 3 explicit scans: 1) search for an idle core in the LLC 2) search for an idle CPU in the LLC 3) search for an idle thread in the 'target' core where 1 and 3 are conditional on SMT support and 1 and 2 have runtime heuristics to skip the step. Step 1) is conditional on sd_llc_shared->has_idle_cores; when a cpu goes idle and sd_llc_shared->has_idle_cores is false, we scan all SMT siblings of the CPU going idle. Similarly, we clear sd_llc_shared->has_idle_cores when we fail to find an idle core. Step 2) tracks the average cost of the scan and compares this to the average idle time guestimate for the CPU doing the wakeup. There is a significant fudge factor involved to deal with the variability of the averages. Esp. hackbench was sensitive to this. Step 3) is unconditional; we assume (also per step 1) that scanning all SMT siblings in a core is 'cheap'. With this; SMT systems gain step 2, which cures a few benchmarks -- notably one from Facebook. One 'feature' of the sched_domain iteration, which we preserve in the new code, is that it would start scanning from the 'target' CPU, instead of scanning the cpumask in cpu id order. This avoids multiple CPUs in the LLC scanning for idle to gang up and find the same CPU quite as much. The down side is that tasks can end up hopping across the LLC for no apparent reason. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-05-09 08:38:05 +00:00
/*
sched/numa: Prefer using an idle CPU as a migration target instead of comparing tasks task_numa_find_cpu() can scan a node multiple times. Minimally it scans to gather statistics and later to find a suitable target. In some cases, the second scan will simply pick an idle CPU if the load is not imbalanced. This patch caches information on an idle core while gathering statistics and uses it immediately if load is not imbalanced to avoid a second scan of the node runqueues. Preference is given to an idle core rather than an idle SMT sibling to avoid packing HT siblings due to linearly scanning the node cpumask. As a side-effect, even when the second scan is necessary, the importance of using select_idle_sibling is much reduced because information on idle CPUs is cached and can be reused. Note that this patch actually makes is harder to move to an idle CPU as multiple tasks can race for the same idle CPU due to a race checking numa_migrate_on. This is addressed in the next patch. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Ingo Molnar <mingo@kernel.org> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Valentin Schneider <valentin.schneider@arm.com> Cc: Phil Auld <pauld@redhat.com> Cc: Hillf Danton <hdanton@sina.com> Link: https://lore.kernel.org/r/20200224095223.13361-11-mgorman@techsingularity.net
2020-02-24 09:52:20 +00:00
* If the CPU is no longer truly idle and the previous best CPU
* is, keep using it.
sched/core: Rewrite and improve select_idle_siblings() select_idle_siblings() is a known pain point for a number of workloads; it either does too much or not enough and sometimes just does plain wrong. This rewrite attempts to address a number of issues (but sadly not all). The current code does an unconditional sched_domain iteration; with the intent of finding an idle core (on SMT hardware). The problems which this patch tries to address are: - its pointless to look for idle cores if the machine is real busy; at which point you're just wasting cycles. - it's behaviour is inconsistent between SMT and !SMT hardware in that !SMT hardware ends up doing a scan for any idle CPU in the LLC domain, while SMT hardware does a scan for idle cores and if that fails, falls back to a scan for idle threads on the 'target' core. The new code replaces the sched_domain scan with 3 explicit scans: 1) search for an idle core in the LLC 2) search for an idle CPU in the LLC 3) search for an idle thread in the 'target' core where 1 and 3 are conditional on SMT support and 1 and 2 have runtime heuristics to skip the step. Step 1) is conditional on sd_llc_shared->has_idle_cores; when a cpu goes idle and sd_llc_shared->has_idle_cores is false, we scan all SMT siblings of the CPU going idle. Similarly, we clear sd_llc_shared->has_idle_cores when we fail to find an idle core. Step 2) tracks the average cost of the scan and compares this to the average idle time guestimate for the CPU doing the wakeup. There is a significant fudge factor involved to deal with the variability of the averages. Esp. hackbench was sensitive to this. Step 3) is unconditional; we assume (also per step 1) that scanning all SMT siblings in a core is 'cheap'. With this; SMT systems gain step 2, which cures a few benchmarks -- notably one from Facebook. One 'feature' of the sched_domain iteration, which we preserve in the new code, is that it would start scanning from the 'target' CPU, instead of scanning the cpumask in cpu id order. This avoids multiple CPUs in the LLC scanning for idle to gang up and find the same CPU quite as much. The down side is that tasks can end up hopping across the LLC for no apparent reason. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-05-09 08:38:05 +00:00
*/
sched/numa: Prefer using an idle CPU as a migration target instead of comparing tasks task_numa_find_cpu() can scan a node multiple times. Minimally it scans to gather statistics and later to find a suitable target. In some cases, the second scan will simply pick an idle CPU if the load is not imbalanced. This patch caches information on an idle core while gathering statistics and uses it immediately if load is not imbalanced to avoid a second scan of the node runqueues. Preference is given to an idle core rather than an idle SMT sibling to avoid packing HT siblings due to linearly scanning the node cpumask. As a side-effect, even when the second scan is necessary, the importance of using select_idle_sibling is much reduced because information on idle CPUs is cached and can be reused. Note that this patch actually makes is harder to move to an idle CPU as multiple tasks can race for the same idle CPU due to a race checking numa_migrate_on. This is addressed in the next patch. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Ingo Molnar <mingo@kernel.org> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Valentin Schneider <valentin.schneider@arm.com> Cc: Phil Auld <pauld@redhat.com> Cc: Hillf Danton <hdanton@sina.com> Link: https://lore.kernel.org/r/20200224095223.13361-11-mgorman@techsingularity.net
2020-02-24 09:52:20 +00:00
if (!idle_cpu(cpu) && env->best_cpu >= 0 &&
idle_cpu(env->best_cpu)) {
cpu = env->best_cpu;
}
env->dst_cpu = cpu;
sched/core: Rewrite and improve select_idle_siblings() select_idle_siblings() is a known pain point for a number of workloads; it either does too much or not enough and sometimes just does plain wrong. This rewrite attempts to address a number of issues (but sadly not all). The current code does an unconditional sched_domain iteration; with the intent of finding an idle core (on SMT hardware). The problems which this patch tries to address are: - its pointless to look for idle cores if the machine is real busy; at which point you're just wasting cycles. - it's behaviour is inconsistent between SMT and !SMT hardware in that !SMT hardware ends up doing a scan for any idle CPU in the LLC domain, while SMT hardware does a scan for idle cores and if that fails, falls back to a scan for idle threads on the 'target' core. The new code replaces the sched_domain scan with 3 explicit scans: 1) search for an idle core in the LLC 2) search for an idle CPU in the LLC 3) search for an idle thread in the 'target' core where 1 and 3 are conditional on SMT support and 1 and 2 have runtime heuristics to skip the step. Step 1) is conditional on sd_llc_shared->has_idle_cores; when a cpu goes idle and sd_llc_shared->has_idle_cores is false, we scan all SMT siblings of the CPU going idle. Similarly, we clear sd_llc_shared->has_idle_cores when we fail to find an idle core. Step 2) tracks the average cost of the scan and compares this to the average idle time guestimate for the CPU doing the wakeup. There is a significant fudge factor involved to deal with the variability of the averages. Esp. hackbench was sensitive to this. Step 3) is unconditional; we assume (also per step 1) that scanning all SMT siblings in a core is 'cheap'. With this; SMT systems gain step 2, which cures a few benchmarks -- notably one from Facebook. One 'feature' of the sched_domain iteration, which we preserve in the new code, is that it would start scanning from the 'target' CPU, instead of scanning the cpumask in cpu id order. This avoids multiple CPUs in the LLC scanning for idle to gang up and find the same CPU quite as much. The down side is that tasks can end up hopping across the LLC for no apparent reason. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-05-09 08:38:05 +00:00
}
task_numa_assign(env, cur, imp);
/*
* If a move to idle is allowed because there is capacity or load
* balance improves then stop the search. While a better swap
* candidate may exist, a search is not free.
*/
if (maymove && !cur && env->best_cpu >= 0 && idle_cpu(env->best_cpu))
stopsearch = true;
/*
* If a swap candidate must be identified and the current best task
* moves its preferred node then stop the search.
*/
if (!maymove && env->best_task &&
env->best_task->numa_preferred_nid == env->src_nid) {
stopsearch = true;
}
unlock:
rcu_read_unlock();
return stopsearch;
}
static void task_numa_find_cpu(struct task_numa_env *env,
long taskimp, long groupimp)
{
sched/numa: Evaluate move once per node task_numa_compare() helps choose the best CPU to move or swap the selected task. To achieve this task_numa_compare() is called for every CPU in the node. Currently it evaluates if the task can be moved/swapped for each of the CPUs. However the move evaluation is mostly independent of the CPU. Evaluating the move logic once per node, provides scope for simplifying task_numa_compare(). Running SPECjbb2005 on a 4 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 16 25705.2 25058.2 -2.51 1 74433 72950 -1.99 Running SPECjbb2005 on a 16 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 8 96589.6 105930 9.670 1 181830 178624 -1.76 (numbers from v1 based on v4.17-rc5) Testcase Time: Min Max Avg StdDev numa01.sh Real: 440.65 941.32 758.98 189.17 numa01.sh Sys: 183.48 320.07 258.42 50.09 numa01.sh User: 37384.65 71818.14 60302.51 13798.96 numa02.sh Real: 61.24 65.35 62.49 1.49 numa02.sh Sys: 16.83 24.18 21.40 2.60 numa02.sh User: 5219.59 5356.34 5264.03 49.07 numa03.sh Real: 822.04 912.40 873.55 37.35 numa03.sh Sys: 118.80 140.94 132.90 7.60 numa03.sh User: 62485.19 70025.01 67208.33 2967.10 numa04.sh Real: 690.66 872.12 778.49 65.44 numa04.sh Sys: 459.26 563.03 494.03 42.39 numa04.sh User: 51116.44 70527.20 58849.44 8461.28 numa05.sh Real: 418.37 562.28 525.77 54.27 numa05.sh Sys: 299.45 481.00 392.49 64.27 numa05.sh User: 34115.09 41324.02 39105.30 2627.68 Testcase Time: Min Max Avg StdDev %Change numa01.sh Real: 516.14 892.41 739.84 151.32 2.587% numa01.sh Sys: 153.16 192.99 177.70 14.58 45.42% numa01.sh User: 39821.04 69528.92 57193.87 10989.48 5.435% numa02.sh Real: 60.91 62.35 61.58 0.63 1.477% numa02.sh Sys: 16.47 26.16 21.20 3.85 0.943% numa02.sh User: 5227.58 5309.61 5265.17 31.04 -0.02% numa03.sh Real: 739.07 917.73 795.75 64.45 9.776% numa03.sh Sys: 94.46 136.08 109.48 14.58 21.39% numa03.sh User: 57478.56 72014.09 61764.48 5343.69 8.813% numa04.sh Real: 442.61 715.43 530.31 96.12 46.79% numa04.sh Sys: 224.90 348.63 285.61 48.83 72.97% numa04.sh User: 35836.84 47522.47 40235.41 3985.26 46.26% numa05.sh Real: 386.13 489.17 434.94 43.59 20.88% numa05.sh Sys: 144.29 438.56 278.80 105.78 40.77% numa05.sh User: 33255.86 36890.82 34879.31 1641.98 12.11% Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mel Gorman <mgorman@techsingularity.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1529514181-9842-3-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-20 17:02:43 +00:00
bool maymove = false;
int cpu;
sched/numa: Evaluate move once per node task_numa_compare() helps choose the best CPU to move or swap the selected task. To achieve this task_numa_compare() is called for every CPU in the node. Currently it evaluates if the task can be moved/swapped for each of the CPUs. However the move evaluation is mostly independent of the CPU. Evaluating the move logic once per node, provides scope for simplifying task_numa_compare(). Running SPECjbb2005 on a 4 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 16 25705.2 25058.2 -2.51 1 74433 72950 -1.99 Running SPECjbb2005 on a 16 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 8 96589.6 105930 9.670 1 181830 178624 -1.76 (numbers from v1 based on v4.17-rc5) Testcase Time: Min Max Avg StdDev numa01.sh Real: 440.65 941.32 758.98 189.17 numa01.sh Sys: 183.48 320.07 258.42 50.09 numa01.sh User: 37384.65 71818.14 60302.51 13798.96 numa02.sh Real: 61.24 65.35 62.49 1.49 numa02.sh Sys: 16.83 24.18 21.40 2.60 numa02.sh User: 5219.59 5356.34 5264.03 49.07 numa03.sh Real: 822.04 912.40 873.55 37.35 numa03.sh Sys: 118.80 140.94 132.90 7.60 numa03.sh User: 62485.19 70025.01 67208.33 2967.10 numa04.sh Real: 690.66 872.12 778.49 65.44 numa04.sh Sys: 459.26 563.03 494.03 42.39 numa04.sh User: 51116.44 70527.20 58849.44 8461.28 numa05.sh Real: 418.37 562.28 525.77 54.27 numa05.sh Sys: 299.45 481.00 392.49 64.27 numa05.sh User: 34115.09 41324.02 39105.30 2627.68 Testcase Time: Min Max Avg StdDev %Change numa01.sh Real: 516.14 892.41 739.84 151.32 2.587% numa01.sh Sys: 153.16 192.99 177.70 14.58 45.42% numa01.sh User: 39821.04 69528.92 57193.87 10989.48 5.435% numa02.sh Real: 60.91 62.35 61.58 0.63 1.477% numa02.sh Sys: 16.47 26.16 21.20 3.85 0.943% numa02.sh User: 5227.58 5309.61 5265.17 31.04 -0.02% numa03.sh Real: 739.07 917.73 795.75 64.45 9.776% numa03.sh Sys: 94.46 136.08 109.48 14.58 21.39% numa03.sh User: 57478.56 72014.09 61764.48 5343.69 8.813% numa04.sh Real: 442.61 715.43 530.31 96.12 46.79% numa04.sh Sys: 224.90 348.63 285.61 48.83 72.97% numa04.sh User: 35836.84 47522.47 40235.41 3985.26 46.26% numa05.sh Real: 386.13 489.17 434.94 43.59 20.88% numa05.sh Sys: 144.29 438.56 278.80 105.78 40.77% numa05.sh User: 33255.86 36890.82 34879.31 1641.98 12.11% Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mel Gorman <mgorman@techsingularity.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1529514181-9842-3-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-20 17:02:43 +00:00
/*
* If dst node has spare capacity, then check if there is an
* imbalance that would be overruled by the load balancer.
sched/numa: Evaluate move once per node task_numa_compare() helps choose the best CPU to move or swap the selected task. To achieve this task_numa_compare() is called for every CPU in the node. Currently it evaluates if the task can be moved/swapped for each of the CPUs. However the move evaluation is mostly independent of the CPU. Evaluating the move logic once per node, provides scope for simplifying task_numa_compare(). Running SPECjbb2005 on a 4 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 16 25705.2 25058.2 -2.51 1 74433 72950 -1.99 Running SPECjbb2005 on a 16 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 8 96589.6 105930 9.670 1 181830 178624 -1.76 (numbers from v1 based on v4.17-rc5) Testcase Time: Min Max Avg StdDev numa01.sh Real: 440.65 941.32 758.98 189.17 numa01.sh Sys: 183.48 320.07 258.42 50.09 numa01.sh User: 37384.65 71818.14 60302.51 13798.96 numa02.sh Real: 61.24 65.35 62.49 1.49 numa02.sh Sys: 16.83 24.18 21.40 2.60 numa02.sh User: 5219.59 5356.34 5264.03 49.07 numa03.sh Real: 822.04 912.40 873.55 37.35 numa03.sh Sys: 118.80 140.94 132.90 7.60 numa03.sh User: 62485.19 70025.01 67208.33 2967.10 numa04.sh Real: 690.66 872.12 778.49 65.44 numa04.sh Sys: 459.26 563.03 494.03 42.39 numa04.sh User: 51116.44 70527.20 58849.44 8461.28 numa05.sh Real: 418.37 562.28 525.77 54.27 numa05.sh Sys: 299.45 481.00 392.49 64.27 numa05.sh User: 34115.09 41324.02 39105.30 2627.68 Testcase Time: Min Max Avg StdDev %Change numa01.sh Real: 516.14 892.41 739.84 151.32 2.587% numa01.sh Sys: 153.16 192.99 177.70 14.58 45.42% numa01.sh User: 39821.04 69528.92 57193.87 10989.48 5.435% numa02.sh Real: 60.91 62.35 61.58 0.63 1.477% numa02.sh Sys: 16.47 26.16 21.20 3.85 0.943% numa02.sh User: 5227.58 5309.61 5265.17 31.04 -0.02% numa03.sh Real: 739.07 917.73 795.75 64.45 9.776% numa03.sh Sys: 94.46 136.08 109.48 14.58 21.39% numa03.sh User: 57478.56 72014.09 61764.48 5343.69 8.813% numa04.sh Real: 442.61 715.43 530.31 96.12 46.79% numa04.sh Sys: 224.90 348.63 285.61 48.83 72.97% numa04.sh User: 35836.84 47522.47 40235.41 3985.26 46.26% numa05.sh Real: 386.13 489.17 434.94 43.59 20.88% numa05.sh Sys: 144.29 438.56 278.80 105.78 40.77% numa05.sh User: 33255.86 36890.82 34879.31 1641.98 12.11% Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mel Gorman <mgorman@techsingularity.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1529514181-9842-3-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-20 17:02:43 +00:00
*/
if (env->dst_stats.node_type == node_has_spare) {
unsigned int imbalance;
int src_running, dst_running;
/*
* Would movement cause an imbalance? Note that if src has
* more running tasks that the imbalance is ignored as the
* move improves the imbalance from the perspective of the
* CPU load balancer.
* */
src_running = env->src_stats.nr_running - 1;
dst_running = env->dst_stats.nr_running + 1;
imbalance = max(0, dst_running - src_running);
imbalance = adjust_numa_imbalance(imbalance, src_running);
/* Use idle CPU if there is no imbalance */
sched/numa: Prefer using an idle CPU as a migration target instead of comparing tasks task_numa_find_cpu() can scan a node multiple times. Minimally it scans to gather statistics and later to find a suitable target. In some cases, the second scan will simply pick an idle CPU if the load is not imbalanced. This patch caches information on an idle core while gathering statistics and uses it immediately if load is not imbalanced to avoid a second scan of the node runqueues. Preference is given to an idle core rather than an idle SMT sibling to avoid packing HT siblings due to linearly scanning the node cpumask. As a side-effect, even when the second scan is necessary, the importance of using select_idle_sibling is much reduced because information on idle CPUs is cached and can be reused. Note that this patch actually makes is harder to move to an idle CPU as multiple tasks can race for the same idle CPU due to a race checking numa_migrate_on. This is addressed in the next patch. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Ingo Molnar <mingo@kernel.org> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Valentin Schneider <valentin.schneider@arm.com> Cc: Phil Auld <pauld@redhat.com> Cc: Hillf Danton <hdanton@sina.com> Link: https://lore.kernel.org/r/20200224095223.13361-11-mgorman@techsingularity.net
2020-02-24 09:52:20 +00:00
if (!imbalance) {
maymove = true;
sched/numa: Prefer using an idle CPU as a migration target instead of comparing tasks task_numa_find_cpu() can scan a node multiple times. Minimally it scans to gather statistics and later to find a suitable target. In some cases, the second scan will simply pick an idle CPU if the load is not imbalanced. This patch caches information on an idle core while gathering statistics and uses it immediately if load is not imbalanced to avoid a second scan of the node runqueues. Preference is given to an idle core rather than an idle SMT sibling to avoid packing HT siblings due to linearly scanning the node cpumask. As a side-effect, even when the second scan is necessary, the importance of using select_idle_sibling is much reduced because information on idle CPUs is cached and can be reused. Note that this patch actually makes is harder to move to an idle CPU as multiple tasks can race for the same idle CPU due to a race checking numa_migrate_on. This is addressed in the next patch. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Ingo Molnar <mingo@kernel.org> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Valentin Schneider <valentin.schneider@arm.com> Cc: Phil Auld <pauld@redhat.com> Cc: Hillf Danton <hdanton@sina.com> Link: https://lore.kernel.org/r/20200224095223.13361-11-mgorman@techsingularity.net
2020-02-24 09:52:20 +00:00
if (env->dst_stats.idle_cpu >= 0) {
env->dst_cpu = env->dst_stats.idle_cpu;
task_numa_assign(env, NULL, 0);
return;
}
}
} else {
long src_load, dst_load, load;
/*
* If the improvement from just moving env->p direction is better
* than swapping tasks around, check if a move is possible.
*/
load = task_h_load(env->p);
dst_load = env->dst_stats.load + load;
src_load = env->src_stats.load - load;
maymove = !load_too_imbalanced(src_load, dst_load, env);
}
sched/numa: Evaluate move once per node task_numa_compare() helps choose the best CPU to move or swap the selected task. To achieve this task_numa_compare() is called for every CPU in the node. Currently it evaluates if the task can be moved/swapped for each of the CPUs. However the move evaluation is mostly independent of the CPU. Evaluating the move logic once per node, provides scope for simplifying task_numa_compare(). Running SPECjbb2005 on a 4 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 16 25705.2 25058.2 -2.51 1 74433 72950 -1.99 Running SPECjbb2005 on a 16 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 8 96589.6 105930 9.670 1 181830 178624 -1.76 (numbers from v1 based on v4.17-rc5) Testcase Time: Min Max Avg StdDev numa01.sh Real: 440.65 941.32 758.98 189.17 numa01.sh Sys: 183.48 320.07 258.42 50.09 numa01.sh User: 37384.65 71818.14 60302.51 13798.96 numa02.sh Real: 61.24 65.35 62.49 1.49 numa02.sh Sys: 16.83 24.18 21.40 2.60 numa02.sh User: 5219.59 5356.34 5264.03 49.07 numa03.sh Real: 822.04 912.40 873.55 37.35 numa03.sh Sys: 118.80 140.94 132.90 7.60 numa03.sh User: 62485.19 70025.01 67208.33 2967.10 numa04.sh Real: 690.66 872.12 778.49 65.44 numa04.sh Sys: 459.26 563.03 494.03 42.39 numa04.sh User: 51116.44 70527.20 58849.44 8461.28 numa05.sh Real: 418.37 562.28 525.77 54.27 numa05.sh Sys: 299.45 481.00 392.49 64.27 numa05.sh User: 34115.09 41324.02 39105.30 2627.68 Testcase Time: Min Max Avg StdDev %Change numa01.sh Real: 516.14 892.41 739.84 151.32 2.587% numa01.sh Sys: 153.16 192.99 177.70 14.58 45.42% numa01.sh User: 39821.04 69528.92 57193.87 10989.48 5.435% numa02.sh Real: 60.91 62.35 61.58 0.63 1.477% numa02.sh Sys: 16.47 26.16 21.20 3.85 0.943% numa02.sh User: 5227.58 5309.61 5265.17 31.04 -0.02% numa03.sh Real: 739.07 917.73 795.75 64.45 9.776% numa03.sh Sys: 94.46 136.08 109.48 14.58 21.39% numa03.sh User: 57478.56 72014.09 61764.48 5343.69 8.813% numa04.sh Real: 442.61 715.43 530.31 96.12 46.79% numa04.sh Sys: 224.90 348.63 285.61 48.83 72.97% numa04.sh User: 35836.84 47522.47 40235.41 3985.26 46.26% numa05.sh Real: 386.13 489.17 434.94 43.59 20.88% numa05.sh Sys: 144.29 438.56 278.80 105.78 40.77% numa05.sh User: 33255.86 36890.82 34879.31 1641.98 12.11% Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mel Gorman <mgorman@techsingularity.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1529514181-9842-3-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-20 17:02:43 +00:00
for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
/* Skip this CPU if the source task cannot migrate */
if (!cpumask_test_cpu(cpu, env->p->cpus_ptr))
continue;
env->dst_cpu = cpu;
if (task_numa_compare(env, taskimp, groupimp, maymove))
break;
}
}
static int task_numa_migrate(struct task_struct *p)
{
struct task_numa_env env = {
.p = p,
.src_cpu = task_cpu(p),
.src_nid = task_node(p),
.imbalance_pct = 112,
.best_task = NULL,
.best_imp = 0,
sched/numa: Spread memory according to CPU and memory use The pseudo-interleaving in NUMA placement has a fundamental problem: using hard usage thresholds to spread memory equally between nodes can prevent workloads from converging, or keep memory "trapped" on nodes where the workload is barely running any more. In order for workloads to properly converge, the memory migration should not be stopped when nodes reach parity, but instead be distributed according to how heavily memory is used from each node. This way memory migration and task migration reinforce each other, instead of one putting the brakes on the other. Remove the hard thresholds from the pseudo-interleaving code, and instead use a more gradual policy on memory placement. This also seems to improve convergence of workloads that do not run flat out, but sleep in between bursts of activity. We still want to slow down NUMA scanning and migration once a workload has settled on a few actively used nodes, so keep the 3/4 hysteresis in place. Keep track of whether a workload is actively running on multiple nodes, so task_numa_migrate does a full scan of the system for better task placement. In the case of running 3 SPECjbb2005 instances on a 4 node system, this code seems to result in fairer distribution of memory between nodes, with more memory bandwidth for each instance. Signed-off-by: Rik van Riel <riel@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: mgorman@suse.de Link: http://lkml.kernel.org/r/20160125170739.2fc9a641@annuminas.surriel.com [ Minor readability tweaks. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-01-25 22:07:39 +00:00
.best_cpu = -1,
};
unsigned long taskweight, groupweight;
struct sched_domain *sd;
long taskimp, groupimp;
struct numa_group *ng;
sched/numa: Stop multiple tasks from moving to the CPU at the same time Task migration under NUMA balancing can happen in parallel. More than one task might choose to migrate to the same CPU at the same time. This can result in: - During task swap, choosing a task that was not part of the evaluation. - During task swap, task which just got moved into its preferred node, moving to a completely different node. - During task swap, task failing to move to the preferred node, will have to wait an extra interval for the next migrate opportunity. - During task movement, multiple task movements can cause load imbalance. This problem is more likely if there are more cores per node or more nodes in the system. Use a per run-queue variable to check if NUMA-balance is active on the run-queue. Specjbb2005 results (8 warehouses) Higher bops are better 2 Socket - 2 Node Haswell - X86 JVMS Prev Current %Change 4 200194 203353 1.57797 1 311331 328205 5.41995 2 Socket - 4 Node Power8 - PowerNV JVMS Prev Current %Change 1 197654 214384 8.46429 2 Socket - 2 Node Power9 - PowerNV JVMS Prev Current %Change 4 192605 188553 -2.10379 1 213402 196273 -8.02664 4 Socket - 4 Node Power7 - PowerVM JVMS Prev Current %Change 8 52227.1 57581.2 10.2516 1 102529 103468 0.915838 There is a regression on power 9 box. If we look at the details, that box has a sudden jump in cache-misses with this patch. All other parameters seem to be pointing towards NUMA consolidation. perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 13,345,784 13,941,377 migrations 1,127,820 1,157,323 faults 374,736 382,175 cache-misses 55,132,054,603 54,993,823,500 sched:sched_move_numa 1,923 2,005 sched:sched_stick_numa 52 14 sched:sched_swap_numa 595 529 migrate:mm_migrate_pages 1,932 1,573 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 60605 67099 numa_hint_faults_local 51804 58456 numa_hit 239945 240416 numa_huge_pte_updates 14 18 numa_interleave 60 65 numa_local 239865 240339 numa_other 80 77 numa_pages_migrated 1931 1574 numa_pte_updates 67823 77182 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 3,016,467 3,176,453 migrations 37,326 30,238 faults 115,342 87,869 cache-misses 11,692,155,554 12,544,479,391 sched:sched_move_numa 965 23 sched:sched_stick_numa 8 0 sched:sched_swap_numa 35 6 migrate:mm_migrate_pages 1,168 10 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 16286 236 numa_hint_faults_local 11863 201 numa_hit 112482 72293 numa_huge_pte_updates 33 0 numa_interleave 20 26 numa_local 112419 72233 numa_other 63 60 numa_pages_migrated 1144 8 numa_pte_updates 32859 0 perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 8,629,724 8,478,820 migrations 221,052 171,323 faults 308,661 307,499 cache-misses 135,574,913 240,353,599 sched:sched_move_numa 147 214 sched:sched_stick_numa 0 0 sched:sched_swap_numa 2 4 migrate:mm_migrate_pages 64 89 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 11481 5301 numa_hint_faults_local 10968 4745 numa_hit 89773 92943 numa_huge_pte_updates 0 0 numa_interleave 1116 899 numa_local 89220 92345 numa_other 553 598 numa_pages_migrated 62 88 numa_pte_updates 11694 5505 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 2,272,887 2,066,172 migrations 12,206 11,076 faults 163,704 149,544 cache-misses 4,801,186 10,398,067 sched:sched_move_numa 44 43 sched:sched_stick_numa 0 0 sched:sched_swap_numa 0 0 migrate:mm_migrate_pages 17 6 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 2261 3552 numa_hint_faults_local 1993 3347 numa_hit 25726 25611 numa_huge_pte_updates 0 0 numa_interleave 239 213 numa_local 25498 25583 numa_other 228 28 numa_pages_migrated 17 6 numa_pte_updates 2266 3535 perf stats 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 117,980,962 99,358,136 migrations 3,950,220 4,041,607 faults 736,979 749,653 cache-misses 224,976,072,879 225,562,543,251 sched:sched_move_numa 504 771 sched:sched_stick_numa 50 14 sched:sched_swap_numa 239 204 migrate:mm_migrate_pages 1,260 1,180 vmstat 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 18293 27409 numa_hint_faults_local 11969 20677 numa_hit 240854 239988 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 240851 239983 numa_other 3 5 numa_pages_migrated 1190 1016 numa_pte_updates 18106 27916 perf stats 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 61,053,158 60,899,307 migrations 551,586 544,668 faults 244,174 270,834 cache-misses 74,326,766,973 74,543,455,635 sched:sched_move_numa 344 735 sched:sched_stick_numa 24 25 sched:sched_swap_numa 140 174 migrate:mm_migrate_pages 568 816 vmstat 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 6461 11059 numa_hint_faults_local 2283 4733 numa_hit 35661 41384 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 35661 41383 numa_other 0 1 numa_pages_migrated 568 815 numa_pte_updates 6518 11323 Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Rik van Riel <riel@surriel.com> Acked-by: Mel Gorman <mgorman@techsingularity.net> Cc: Jirka Hladky <jhladky@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1537552141-27815-2-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-09-21 17:48:56 +00:00
struct rq *best_rq;
int nid, ret, dist;
/*
* Pick the lowest SD_NUMA domain, as that would have the smallest
* imbalance and would be the first to start moving tasks about.
*
* And we want to avoid any moving of tasks about, as that would create
* random movement of tasks -- counter the numa conditions we're trying
* to satisfy here.
*/
rcu_read_lock();
sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
if (sd)
env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
rcu_read_unlock();
/*
* Cpusets can break the scheduler domain tree into smaller
* balance domains, some of which do not cross NUMA boundaries.
* Tasks that are "trapped" in such domains cannot be migrated
* elsewhere, so there is no point in (re)trying.
*/
if (unlikely(!sd)) {
sched/numa: Set preferred_node based on best_cpu Currently preferred node is set to dst_nid which is the last node in the iteration whose group weight or task weight is greater than the current node. However it doesn't guarantee that dst_nid has the numa capacity to move. It also doesn't guarantee that dst_nid has the best_cpu which is the CPU/node ideal for node migration. Lets consider faults on a 4 node system with group weight numbers in different nodes being in 0 < 1 < 2 < 3 proportion. Consider the task is running on 3 and 0 is its preferred node but its capacity is full. Consider nodes 1, 2 and 3 have capacity. Then the task should be migrated to node 1. Currently the task gets moved to node 2. env.dst_nid points to the last node whose faults were greater than current node. Modify to set the preferred node based of best_cpu. Earlier setting preferred node was skipped if nr_active_nodes is 1. This could result in the task being moved out of the preferred node to a random node during regular load balancing. Also while modifying task_numa_migrate(), use sched_setnuma to set preferred node. This ensures out numa accounting is correct. Running SPECjbb2005 on a 4 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 16 25122.9 25549.6 1.698 1 73850 73190 -0.89 Running SPECjbb2005 on a 16 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 8 105930 113437 7.08676 1 178624 196130 9.80047 (numbers from v1 based on v4.17-rc5) Testcase Time: Min Max Avg StdDev numa01.sh Real: 435.78 653.81 534.58 83.20 numa01.sh Sys: 121.93 187.18 145.90 23.47 numa01.sh User: 37082.81 51402.80 43647.60 5409.75 numa02.sh Real: 60.64 61.63 61.19 0.40 numa02.sh Sys: 14.72 25.68 19.06 4.03 numa02.sh User: 5210.95 5266.69 5233.30 20.82 numa03.sh Real: 746.51 808.24 780.36 23.88 numa03.sh Sys: 97.26 108.48 105.07 4.28 numa03.sh User: 58956.30 61397.05 60162.95 1050.82 numa04.sh Real: 465.97 519.27 484.81 19.62 numa04.sh Sys: 304.43 359.08 334.68 20.64 numa04.sh User: 37544.16 41186.15 39262.44 1314.91 numa05.sh Real: 411.57 457.20 433.29 16.58 numa05.sh Sys: 230.05 435.48 339.95 67.58 numa05.sh User: 33325.54 36896.31 35637.84 1222.64 Testcase Time: Min Max Avg StdDev %Change numa01.sh Real: 506.35 794.46 599.06 104.26 -10.76% numa01.sh Sys: 150.37 223.56 195.99 24.94 -25.55% numa01.sh User: 43450.69 61752.04 49281.50 6635.33 -11.43% numa02.sh Real: 60.33 62.40 61.31 0.90 -0.195% numa02.sh Sys: 18.12 31.66 24.28 5.89 -21.49% numa02.sh User: 5203.91 5325.32 5260.29 49.98 -0.513% numa03.sh Real: 696.47 853.62 745.80 57.28 4.6339% numa03.sh Sys: 85.68 123.71 97.89 13.48 7.3347% numa03.sh User: 55978.45 66418.63 59254.94 3737.97 1.5323% numa04.sh Real: 444.05 514.83 497.06 26.85 -2.464% numa04.sh Sys: 230.39 375.79 316.23 48.58 5.8343% numa04.sh User: 35403.12 41004.10 39720.80 2163.08 -1.153% numa05.sh Real: 423.09 460.41 439.57 13.92 -1.428% numa05.sh Sys: 287.38 480.15 369.37 68.52 -7.964% numa05.sh User: 34732.12 38016.80 36255.85 1070.51 -1.704% Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mel Gorman <mgorman@techsingularity.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1529514181-9842-5-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-20 17:02:45 +00:00
sched_setnuma(p, task_node(p));
return -EINVAL;
}
env.dst_nid = p->numa_preferred_nid;
dist = env.dist = node_distance(env.src_nid, env.dst_nid);
taskweight = task_weight(p, env.src_nid, dist);
groupweight = group_weight(p, env.src_nid, dist);
sched/numa: Prefer using an idle CPU as a migration target instead of comparing tasks task_numa_find_cpu() can scan a node multiple times. Minimally it scans to gather statistics and later to find a suitable target. In some cases, the second scan will simply pick an idle CPU if the load is not imbalanced. This patch caches information on an idle core while gathering statistics and uses it immediately if load is not imbalanced to avoid a second scan of the node runqueues. Preference is given to an idle core rather than an idle SMT sibling to avoid packing HT siblings due to linearly scanning the node cpumask. As a side-effect, even when the second scan is necessary, the importance of using select_idle_sibling is much reduced because information on idle CPUs is cached and can be reused. Note that this patch actually makes is harder to move to an idle CPU as multiple tasks can race for the same idle CPU due to a race checking numa_migrate_on. This is addressed in the next patch. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Ingo Molnar <mingo@kernel.org> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Valentin Schneider <valentin.schneider@arm.com> Cc: Phil Auld <pauld@redhat.com> Cc: Hillf Danton <hdanton@sina.com> Link: https://lore.kernel.org/r/20200224095223.13361-11-mgorman@techsingularity.net
2020-02-24 09:52:20 +00:00
update_numa_stats(&env, &env.src_stats, env.src_nid, false);
taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
sched/numa: Prefer using an idle CPU as a migration target instead of comparing tasks task_numa_find_cpu() can scan a node multiple times. Minimally it scans to gather statistics and later to find a suitable target. In some cases, the second scan will simply pick an idle CPU if the load is not imbalanced. This patch caches information on an idle core while gathering statistics and uses it immediately if load is not imbalanced to avoid a second scan of the node runqueues. Preference is given to an idle core rather than an idle SMT sibling to avoid packing HT siblings due to linearly scanning the node cpumask. As a side-effect, even when the second scan is necessary, the importance of using select_idle_sibling is much reduced because information on idle CPUs is cached and can be reused. Note that this patch actually makes is harder to move to an idle CPU as multiple tasks can race for the same idle CPU due to a race checking numa_migrate_on. This is addressed in the next patch. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Ingo Molnar <mingo@kernel.org> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Valentin Schneider <valentin.schneider@arm.com> Cc: Phil Auld <pauld@redhat.com> Cc: Hillf Danton <hdanton@sina.com> Link: https://lore.kernel.org/r/20200224095223.13361-11-mgorman@techsingularity.net
2020-02-24 09:52:20 +00:00
update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
/* Try to find a spot on the preferred nid. */
task_numa_find_cpu(&env, taskimp, groupimp);
/*
* Look at other nodes in these cases:
* - there is no space available on the preferred_nid
* - the task is part of a numa_group that is interleaved across
* multiple NUMA nodes; in order to better consolidate the group,
* we need to check other locations.
*/
ng = deref_curr_numa_group(p);
if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) {
for_each_online_node(nid) {
if (nid == env.src_nid || nid == p->numa_preferred_nid)
continue;
dist = node_distance(env.src_nid, env.dst_nid);
sched/numa: Calculate node scores in complex NUMA topologies In order to do task placement on systems with complex NUMA topologies, it is necessary to count the faults on nodes nearby the node that is being examined for a potential move. In case of a system with a backplane interconnect, we are dealing with groups of NUMA nodes; each of the nodes within a group is the same number of hops away from nodes in other groups in the system. Optimal placement on this topology is achieved by counting all nearby nodes equally. When comparing nodes A and B at distance N, nearby nodes are those at distances smaller than N from nodes A or B. Placement strategy on a system with a glueless mesh NUMA topology needs to be different, because there are no natural groups of nodes determined by the hardware. Instead, when dealing with two nodes A and B at distance N, N >= 2, there will be intermediate nodes at distance < N from both nodes A and B. Good placement can be achieved by right shifting the faults on nearby nodes by the number of hops from the node being scored. In this context, a nearby node is any node less than the maximum distance in the system away from the node. Those nodes are skipped for efficiency reasons, there is no real policy reason to do so. Placement policy on directly connected NUMA systems is not affected. Signed-off-by: Rik van Riel <riel@redhat.com> Tested-by: Chegu Vinod <chegu_vinod@hp.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: mgorman@suse.de Cc: chegu_vinod@hp.com Link: http://lkml.kernel.org/r/1413530994-9732-5-git-send-email-riel@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-10-17 07:29:52 +00:00
if (sched_numa_topology_type == NUMA_BACKPLANE &&
dist != env.dist) {
taskweight = task_weight(p, env.src_nid, dist);
groupweight = group_weight(p, env.src_nid, dist);
}
/* Only consider nodes where both task and groups benefit */
taskimp = task_weight(p, nid, dist) - taskweight;
groupimp = group_weight(p, nid, dist) - groupweight;
if (taskimp < 0 && groupimp < 0)
continue;
env.dist = dist;
env.dst_nid = nid;
sched/numa: Prefer using an idle CPU as a migration target instead of comparing tasks task_numa_find_cpu() can scan a node multiple times. Minimally it scans to gather statistics and later to find a suitable target. In some cases, the second scan will simply pick an idle CPU if the load is not imbalanced. This patch caches information on an idle core while gathering statistics and uses it immediately if load is not imbalanced to avoid a second scan of the node runqueues. Preference is given to an idle core rather than an idle SMT sibling to avoid packing HT siblings due to linearly scanning the node cpumask. As a side-effect, even when the second scan is necessary, the importance of using select_idle_sibling is much reduced because information on idle CPUs is cached and can be reused. Note that this patch actually makes is harder to move to an idle CPU as multiple tasks can race for the same idle CPU due to a race checking numa_migrate_on. This is addressed in the next patch. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Ingo Molnar <mingo@kernel.org> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Valentin Schneider <valentin.schneider@arm.com> Cc: Phil Auld <pauld@redhat.com> Cc: Hillf Danton <hdanton@sina.com> Link: https://lore.kernel.org/r/20200224095223.13361-11-mgorman@techsingularity.net
2020-02-24 09:52:20 +00:00
update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
task_numa_find_cpu(&env, taskimp, groupimp);
}
}
/*
* If the task is part of a workload that spans multiple NUMA nodes,
* and is migrating into one of the workload's active nodes, remember
* this node as the task's preferred numa node, so the workload can
* settle down.
* A task that migrated to a second choice node will be better off
* trying for a better one later. Do not set the preferred node here.
*/
if (ng) {
if (env.best_cpu == -1)
nid = env.src_nid;
else
sched/numa: Set preferred_node based on best_cpu Currently preferred node is set to dst_nid which is the last node in the iteration whose group weight or task weight is greater than the current node. However it doesn't guarantee that dst_nid has the numa capacity to move. It also doesn't guarantee that dst_nid has the best_cpu which is the CPU/node ideal for node migration. Lets consider faults on a 4 node system with group weight numbers in different nodes being in 0 < 1 < 2 < 3 proportion. Consider the task is running on 3 and 0 is its preferred node but its capacity is full. Consider nodes 1, 2 and 3 have capacity. Then the task should be migrated to node 1. Currently the task gets moved to node 2. env.dst_nid points to the last node whose faults were greater than current node. Modify to set the preferred node based of best_cpu. Earlier setting preferred node was skipped if nr_active_nodes is 1. This could result in the task being moved out of the preferred node to a random node during regular load balancing. Also while modifying task_numa_migrate(), use sched_setnuma to set preferred node. This ensures out numa accounting is correct. Running SPECjbb2005 on a 4 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 16 25122.9 25549.6 1.698 1 73850 73190 -0.89 Running SPECjbb2005 on a 16 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 8 105930 113437 7.08676 1 178624 196130 9.80047 (numbers from v1 based on v4.17-rc5) Testcase Time: Min Max Avg StdDev numa01.sh Real: 435.78 653.81 534.58 83.20 numa01.sh Sys: 121.93 187.18 145.90 23.47 numa01.sh User: 37082.81 51402.80 43647.60 5409.75 numa02.sh Real: 60.64 61.63 61.19 0.40 numa02.sh Sys: 14.72 25.68 19.06 4.03 numa02.sh User: 5210.95 5266.69 5233.30 20.82 numa03.sh Real: 746.51 808.24 780.36 23.88 numa03.sh Sys: 97.26 108.48 105.07 4.28 numa03.sh User: 58956.30 61397.05 60162.95 1050.82 numa04.sh Real: 465.97 519.27 484.81 19.62 numa04.sh Sys: 304.43 359.08 334.68 20.64 numa04.sh User: 37544.16 41186.15 39262.44 1314.91 numa05.sh Real: 411.57 457.20 433.29 16.58 numa05.sh Sys: 230.05 435.48 339.95 67.58 numa05.sh User: 33325.54 36896.31 35637.84 1222.64 Testcase Time: Min Max Avg StdDev %Change numa01.sh Real: 506.35 794.46 599.06 104.26 -10.76% numa01.sh Sys: 150.37 223.56 195.99 24.94 -25.55% numa01.sh User: 43450.69 61752.04 49281.50 6635.33 -11.43% numa02.sh Real: 60.33 62.40 61.31 0.90 -0.195% numa02.sh Sys: 18.12 31.66 24.28 5.89 -21.49% numa02.sh User: 5203.91 5325.32 5260.29 49.98 -0.513% numa03.sh Real: 696.47 853.62 745.80 57.28 4.6339% numa03.sh Sys: 85.68 123.71 97.89 13.48 7.3347% numa03.sh User: 55978.45 66418.63 59254.94 3737.97 1.5323% numa04.sh Real: 444.05 514.83 497.06 26.85 -2.464% numa04.sh Sys: 230.39 375.79 316.23 48.58 5.8343% numa04.sh User: 35403.12 41004.10 39720.80 2163.08 -1.153% numa05.sh Real: 423.09 460.41 439.57 13.92 -1.428% numa05.sh Sys: 287.38 480.15 369.37 68.52 -7.964% numa05.sh User: 34732.12 38016.80 36255.85 1070.51 -1.704% Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mel Gorman <mgorman@techsingularity.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1529514181-9842-5-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-20 17:02:45 +00:00
nid = cpu_to_node(env.best_cpu);
sched/numa: Set preferred_node based on best_cpu Currently preferred node is set to dst_nid which is the last node in the iteration whose group weight or task weight is greater than the current node. However it doesn't guarantee that dst_nid has the numa capacity to move. It also doesn't guarantee that dst_nid has the best_cpu which is the CPU/node ideal for node migration. Lets consider faults on a 4 node system with group weight numbers in different nodes being in 0 < 1 < 2 < 3 proportion. Consider the task is running on 3 and 0 is its preferred node but its capacity is full. Consider nodes 1, 2 and 3 have capacity. Then the task should be migrated to node 1. Currently the task gets moved to node 2. env.dst_nid points to the last node whose faults were greater than current node. Modify to set the preferred node based of best_cpu. Earlier setting preferred node was skipped if nr_active_nodes is 1. This could result in the task being moved out of the preferred node to a random node during regular load balancing. Also while modifying task_numa_migrate(), use sched_setnuma to set preferred node. This ensures out numa accounting is correct. Running SPECjbb2005 on a 4 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 16 25122.9 25549.6 1.698 1 73850 73190 -0.89 Running SPECjbb2005 on a 16 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 8 105930 113437 7.08676 1 178624 196130 9.80047 (numbers from v1 based on v4.17-rc5) Testcase Time: Min Max Avg StdDev numa01.sh Real: 435.78 653.81 534.58 83.20 numa01.sh Sys: 121.93 187.18 145.90 23.47 numa01.sh User: 37082.81 51402.80 43647.60 5409.75 numa02.sh Real: 60.64 61.63 61.19 0.40 numa02.sh Sys: 14.72 25.68 19.06 4.03 numa02.sh User: 5210.95 5266.69 5233.30 20.82 numa03.sh Real: 746.51 808.24 780.36 23.88 numa03.sh Sys: 97.26 108.48 105.07 4.28 numa03.sh User: 58956.30 61397.05 60162.95 1050.82 numa04.sh Real: 465.97 519.27 484.81 19.62 numa04.sh Sys: 304.43 359.08 334.68 20.64 numa04.sh User: 37544.16 41186.15 39262.44 1314.91 numa05.sh Real: 411.57 457.20 433.29 16.58 numa05.sh Sys: 230.05 435.48 339.95 67.58 numa05.sh User: 33325.54 36896.31 35637.84 1222.64 Testcase Time: Min Max Avg StdDev %Change numa01.sh Real: 506.35 794.46 599.06 104.26 -10.76% numa01.sh Sys: 150.37 223.56 195.99 24.94 -25.55% numa01.sh User: 43450.69 61752.04 49281.50 6635.33 -11.43% numa02.sh Real: 60.33 62.40 61.31 0.90 -0.195% numa02.sh Sys: 18.12 31.66 24.28 5.89 -21.49% numa02.sh User: 5203.91 5325.32 5260.29 49.98 -0.513% numa03.sh Real: 696.47 853.62 745.80 57.28 4.6339% numa03.sh Sys: 85.68 123.71 97.89 13.48 7.3347% numa03.sh User: 55978.45 66418.63 59254.94 3737.97 1.5323% numa04.sh Real: 444.05 514.83 497.06 26.85 -2.464% numa04.sh Sys: 230.39 375.79 316.23 48.58 5.8343% numa04.sh User: 35403.12 41004.10 39720.80 2163.08 -1.153% numa05.sh Real: 423.09 460.41 439.57 13.92 -1.428% numa05.sh Sys: 287.38 480.15 369.37 68.52 -7.964% numa05.sh User: 34732.12 38016.80 36255.85 1070.51 -1.704% Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mel Gorman <mgorman@techsingularity.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1529514181-9842-5-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-20 17:02:45 +00:00
if (nid != p->numa_preferred_nid)
sched_setnuma(p, nid);
}
/* No better CPU than the current one was found. */
if (env.best_cpu == -1) {
trace_sched_stick_numa(p, env.src_cpu, NULL, -1);
return -EAGAIN;
}
sched/numa: Stop multiple tasks from moving to the CPU at the same time Task migration under NUMA balancing can happen in parallel. More than one task might choose to migrate to the same CPU at the same time. This can result in: - During task swap, choosing a task that was not part of the evaluation. - During task swap, task which just got moved into its preferred node, moving to a completely different node. - During task swap, task failing to move to the preferred node, will have to wait an extra interval for the next migrate opportunity. - During task movement, multiple task movements can cause load imbalance. This problem is more likely if there are more cores per node or more nodes in the system. Use a per run-queue variable to check if NUMA-balance is active on the run-queue. Specjbb2005 results (8 warehouses) Higher bops are better 2 Socket - 2 Node Haswell - X86 JVMS Prev Current %Change 4 200194 203353 1.57797 1 311331 328205 5.41995 2 Socket - 4 Node Power8 - PowerNV JVMS Prev Current %Change 1 197654 214384 8.46429 2 Socket - 2 Node Power9 - PowerNV JVMS Prev Current %Change 4 192605 188553 -2.10379 1 213402 196273 -8.02664 4 Socket - 4 Node Power7 - PowerVM JVMS Prev Current %Change 8 52227.1 57581.2 10.2516 1 102529 103468 0.915838 There is a regression on power 9 box. If we look at the details, that box has a sudden jump in cache-misses with this patch. All other parameters seem to be pointing towards NUMA consolidation. perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 13,345,784 13,941,377 migrations 1,127,820 1,157,323 faults 374,736 382,175 cache-misses 55,132,054,603 54,993,823,500 sched:sched_move_numa 1,923 2,005 sched:sched_stick_numa 52 14 sched:sched_swap_numa 595 529 migrate:mm_migrate_pages 1,932 1,573 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 60605 67099 numa_hint_faults_local 51804 58456 numa_hit 239945 240416 numa_huge_pte_updates 14 18 numa_interleave 60 65 numa_local 239865 240339 numa_other 80 77 numa_pages_migrated 1931 1574 numa_pte_updates 67823 77182 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 3,016,467 3,176,453 migrations 37,326 30,238 faults 115,342 87,869 cache-misses 11,692,155,554 12,544,479,391 sched:sched_move_numa 965 23 sched:sched_stick_numa 8 0 sched:sched_swap_numa 35 6 migrate:mm_migrate_pages 1,168 10 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 16286 236 numa_hint_faults_local 11863 201 numa_hit 112482 72293 numa_huge_pte_updates 33 0 numa_interleave 20 26 numa_local 112419 72233 numa_other 63 60 numa_pages_migrated 1144 8 numa_pte_updates 32859 0 perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 8,629,724 8,478,820 migrations 221,052 171,323 faults 308,661 307,499 cache-misses 135,574,913 240,353,599 sched:sched_move_numa 147 214 sched:sched_stick_numa 0 0 sched:sched_swap_numa 2 4 migrate:mm_migrate_pages 64 89 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 11481 5301 numa_hint_faults_local 10968 4745 numa_hit 89773 92943 numa_huge_pte_updates 0 0 numa_interleave 1116 899 numa_local 89220 92345 numa_other 553 598 numa_pages_migrated 62 88 numa_pte_updates 11694 5505 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 2,272,887 2,066,172 migrations 12,206 11,076 faults 163,704 149,544 cache-misses 4,801,186 10,398,067 sched:sched_move_numa 44 43 sched:sched_stick_numa 0 0 sched:sched_swap_numa 0 0 migrate:mm_migrate_pages 17 6 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 2261 3552 numa_hint_faults_local 1993 3347 numa_hit 25726 25611 numa_huge_pte_updates 0 0 numa_interleave 239 213 numa_local 25498 25583 numa_other 228 28 numa_pages_migrated 17 6 numa_pte_updates 2266 3535 perf stats 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 117,980,962 99,358,136 migrations 3,950,220 4,041,607 faults 736,979 749,653 cache-misses 224,976,072,879 225,562,543,251 sched:sched_move_numa 504 771 sched:sched_stick_numa 50 14 sched:sched_swap_numa 239 204 migrate:mm_migrate_pages 1,260 1,180 vmstat 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 18293 27409 numa_hint_faults_local 11969 20677 numa_hit 240854 239988 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 240851 239983 numa_other 3 5 numa_pages_migrated 1190 1016 numa_pte_updates 18106 27916 perf stats 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 61,053,158 60,899,307 migrations 551,586 544,668 faults 244,174 270,834 cache-misses 74,326,766,973 74,543,455,635 sched:sched_move_numa 344 735 sched:sched_stick_numa 24 25 sched:sched_swap_numa 140 174 migrate:mm_migrate_pages 568 816 vmstat 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 6461 11059 numa_hint_faults_local 2283 4733 numa_hit 35661 41384 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 35661 41383 numa_other 0 1 numa_pages_migrated 568 815 numa_pte_updates 6518 11323 Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Rik van Riel <riel@surriel.com> Acked-by: Mel Gorman <mgorman@techsingularity.net> Cc: Jirka Hladky <jhladky@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1537552141-27815-2-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-09-21 17:48:56 +00:00
best_rq = cpu_rq(env.best_cpu);
if (env.best_task == NULL) {
sched: add tracepoints related to NUMA task migration This patch adds three tracepoints o trace_sched_move_numa when a task is moved to a node o trace_sched_swap_numa when a task is swapped with another task o trace_sched_stick_numa when a numa-related migration fails The tracepoints allow the NUMA scheduler activity to be monitored and the following high-level metrics can be calculated o NUMA migrated stuck nr trace_sched_stick_numa o NUMA migrated idle nr trace_sched_move_numa o NUMA migrated swapped nr trace_sched_swap_numa o NUMA local swapped trace_sched_swap_numa src_nid == dst_nid (should never happen) o NUMA remote swapped trace_sched_swap_numa src_nid != dst_nid (should == NUMA migrated swapped) o NUMA group swapped trace_sched_swap_numa src_ngid == dst_ngid Maybe a small number of these are acceptable but a high number would be a major surprise. It would be even worse if bounces are frequent. o NUMA avg task migs. Average number of migrations for tasks o NUMA stddev task mig Self-explanatory o NUMA max task migs. Maximum number of migrations for a single task In general the intent of the tracepoints is to help diagnose problems where automatic NUMA balancing appears to be doing an excessive amount of useless work. [akpm@linux-foundation.org: remove semicolon-after-if, repair coding-style] Signed-off-by: Mel Gorman <mgorman@suse.de> Reviewed-by: Rik van Riel <riel@redhat.com> Cc: Alex Thorlton <athorlton@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-01-21 23:51:03 +00:00
ret = migrate_task_to(p, env.best_cpu);
sched/numa: Stop multiple tasks from moving to the CPU at the same time Task migration under NUMA balancing can happen in parallel. More than one task might choose to migrate to the same CPU at the same time. This can result in: - During task swap, choosing a task that was not part of the evaluation. - During task swap, task which just got moved into its preferred node, moving to a completely different node. - During task swap, task failing to move to the preferred node, will have to wait an extra interval for the next migrate opportunity. - During task movement, multiple task movements can cause load imbalance. This problem is more likely if there are more cores per node or more nodes in the system. Use a per run-queue variable to check if NUMA-balance is active on the run-queue. Specjbb2005 results (8 warehouses) Higher bops are better 2 Socket - 2 Node Haswell - X86 JVMS Prev Current %Change 4 200194 203353 1.57797 1 311331 328205 5.41995 2 Socket - 4 Node Power8 - PowerNV JVMS Prev Current %Change 1 197654 214384 8.46429 2 Socket - 2 Node Power9 - PowerNV JVMS Prev Current %Change 4 192605 188553 -2.10379 1 213402 196273 -8.02664 4 Socket - 4 Node Power7 - PowerVM JVMS Prev Current %Change 8 52227.1 57581.2 10.2516 1 102529 103468 0.915838 There is a regression on power 9 box. If we look at the details, that box has a sudden jump in cache-misses with this patch. All other parameters seem to be pointing towards NUMA consolidation. perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 13,345,784 13,941,377 migrations 1,127,820 1,157,323 faults 374,736 382,175 cache-misses 55,132,054,603 54,993,823,500 sched:sched_move_numa 1,923 2,005 sched:sched_stick_numa 52 14 sched:sched_swap_numa 595 529 migrate:mm_migrate_pages 1,932 1,573 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 60605 67099 numa_hint_faults_local 51804 58456 numa_hit 239945 240416 numa_huge_pte_updates 14 18 numa_interleave 60 65 numa_local 239865 240339 numa_other 80 77 numa_pages_migrated 1931 1574 numa_pte_updates 67823 77182 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 3,016,467 3,176,453 migrations 37,326 30,238 faults 115,342 87,869 cache-misses 11,692,155,554 12,544,479,391 sched:sched_move_numa 965 23 sched:sched_stick_numa 8 0 sched:sched_swap_numa 35 6 migrate:mm_migrate_pages 1,168 10 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 16286 236 numa_hint_faults_local 11863 201 numa_hit 112482 72293 numa_huge_pte_updates 33 0 numa_interleave 20 26 numa_local 112419 72233 numa_other 63 60 numa_pages_migrated 1144 8 numa_pte_updates 32859 0 perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 8,629,724 8,478,820 migrations 221,052 171,323 faults 308,661 307,499 cache-misses 135,574,913 240,353,599 sched:sched_move_numa 147 214 sched:sched_stick_numa 0 0 sched:sched_swap_numa 2 4 migrate:mm_migrate_pages 64 89 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 11481 5301 numa_hint_faults_local 10968 4745 numa_hit 89773 92943 numa_huge_pte_updates 0 0 numa_interleave 1116 899 numa_local 89220 92345 numa_other 553 598 numa_pages_migrated 62 88 numa_pte_updates 11694 5505 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 2,272,887 2,066,172 migrations 12,206 11,076 faults 163,704 149,544 cache-misses 4,801,186 10,398,067 sched:sched_move_numa 44 43 sched:sched_stick_numa 0 0 sched:sched_swap_numa 0 0 migrate:mm_migrate_pages 17 6 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 2261 3552 numa_hint_faults_local 1993 3347 numa_hit 25726 25611 numa_huge_pte_updates 0 0 numa_interleave 239 213 numa_local 25498 25583 numa_other 228 28 numa_pages_migrated 17 6 numa_pte_updates 2266 3535 perf stats 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 117,980,962 99,358,136 migrations 3,950,220 4,041,607 faults 736,979 749,653 cache-misses 224,976,072,879 225,562,543,251 sched:sched_move_numa 504 771 sched:sched_stick_numa 50 14 sched:sched_swap_numa 239 204 migrate:mm_migrate_pages 1,260 1,180 vmstat 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 18293 27409 numa_hint_faults_local 11969 20677 numa_hit 240854 239988 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 240851 239983 numa_other 3 5 numa_pages_migrated 1190 1016 numa_pte_updates 18106 27916 perf stats 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 61,053,158 60,899,307 migrations 551,586 544,668 faults 244,174 270,834 cache-misses 74,326,766,973 74,543,455,635 sched:sched_move_numa 344 735 sched:sched_stick_numa 24 25 sched:sched_swap_numa 140 174 migrate:mm_migrate_pages 568 816 vmstat 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 6461 11059 numa_hint_faults_local 2283 4733 numa_hit 35661 41384 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 35661 41383 numa_other 0 1 numa_pages_migrated 568 815 numa_pte_updates 6518 11323 Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Rik van Riel <riel@surriel.com> Acked-by: Mel Gorman <mgorman@techsingularity.net> Cc: Jirka Hladky <jhladky@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1537552141-27815-2-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-09-21 17:48:56 +00:00
WRITE_ONCE(best_rq->numa_migrate_on, 0);
sched: add tracepoints related to NUMA task migration This patch adds three tracepoints o trace_sched_move_numa when a task is moved to a node o trace_sched_swap_numa when a task is swapped with another task o trace_sched_stick_numa when a numa-related migration fails The tracepoints allow the NUMA scheduler activity to be monitored and the following high-level metrics can be calculated o NUMA migrated stuck nr trace_sched_stick_numa o NUMA migrated idle nr trace_sched_move_numa o NUMA migrated swapped nr trace_sched_swap_numa o NUMA local swapped trace_sched_swap_numa src_nid == dst_nid (should never happen) o NUMA remote swapped trace_sched_swap_numa src_nid != dst_nid (should == NUMA migrated swapped) o NUMA group swapped trace_sched_swap_numa src_ngid == dst_ngid Maybe a small number of these are acceptable but a high number would be a major surprise. It would be even worse if bounces are frequent. o NUMA avg task migs. Average number of migrations for tasks o NUMA stddev task mig Self-explanatory o NUMA max task migs. Maximum number of migrations for a single task In general the intent of the tracepoints is to help diagnose problems where automatic NUMA balancing appears to be doing an excessive amount of useless work. [akpm@linux-foundation.org: remove semicolon-after-if, repair coding-style] Signed-off-by: Mel Gorman <mgorman@suse.de> Reviewed-by: Rik van Riel <riel@redhat.com> Cc: Alex Thorlton <athorlton@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-01-21 23:51:03 +00:00
if (ret != 0)
trace_sched_stick_numa(p, env.src_cpu, NULL, env.best_cpu);
return ret;
}
ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
sched/numa: Stop multiple tasks from moving to the CPU at the same time Task migration under NUMA balancing can happen in parallel. More than one task might choose to migrate to the same CPU at the same time. This can result in: - During task swap, choosing a task that was not part of the evaluation. - During task swap, task which just got moved into its preferred node, moving to a completely different node. - During task swap, task failing to move to the preferred node, will have to wait an extra interval for the next migrate opportunity. - During task movement, multiple task movements can cause load imbalance. This problem is more likely if there are more cores per node or more nodes in the system. Use a per run-queue variable to check if NUMA-balance is active on the run-queue. Specjbb2005 results (8 warehouses) Higher bops are better 2 Socket - 2 Node Haswell - X86 JVMS Prev Current %Change 4 200194 203353 1.57797 1 311331 328205 5.41995 2 Socket - 4 Node Power8 - PowerNV JVMS Prev Current %Change 1 197654 214384 8.46429 2 Socket - 2 Node Power9 - PowerNV JVMS Prev Current %Change 4 192605 188553 -2.10379 1 213402 196273 -8.02664 4 Socket - 4 Node Power7 - PowerVM JVMS Prev Current %Change 8 52227.1 57581.2 10.2516 1 102529 103468 0.915838 There is a regression on power 9 box. If we look at the details, that box has a sudden jump in cache-misses with this patch. All other parameters seem to be pointing towards NUMA consolidation. perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 13,345,784 13,941,377 migrations 1,127,820 1,157,323 faults 374,736 382,175 cache-misses 55,132,054,603 54,993,823,500 sched:sched_move_numa 1,923 2,005 sched:sched_stick_numa 52 14 sched:sched_swap_numa 595 529 migrate:mm_migrate_pages 1,932 1,573 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 60605 67099 numa_hint_faults_local 51804 58456 numa_hit 239945 240416 numa_huge_pte_updates 14 18 numa_interleave 60 65 numa_local 239865 240339 numa_other 80 77 numa_pages_migrated 1931 1574 numa_pte_updates 67823 77182 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 3,016,467 3,176,453 migrations 37,326 30,238 faults 115,342 87,869 cache-misses 11,692,155,554 12,544,479,391 sched:sched_move_numa 965 23 sched:sched_stick_numa 8 0 sched:sched_swap_numa 35 6 migrate:mm_migrate_pages 1,168 10 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 16286 236 numa_hint_faults_local 11863 201 numa_hit 112482 72293 numa_huge_pte_updates 33 0 numa_interleave 20 26 numa_local 112419 72233 numa_other 63 60 numa_pages_migrated 1144 8 numa_pte_updates 32859 0 perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 8,629,724 8,478,820 migrations 221,052 171,323 faults 308,661 307,499 cache-misses 135,574,913 240,353,599 sched:sched_move_numa 147 214 sched:sched_stick_numa 0 0 sched:sched_swap_numa 2 4 migrate:mm_migrate_pages 64 89 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 11481 5301 numa_hint_faults_local 10968 4745 numa_hit 89773 92943 numa_huge_pte_updates 0 0 numa_interleave 1116 899 numa_local 89220 92345 numa_other 553 598 numa_pages_migrated 62 88 numa_pte_updates 11694 5505 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 2,272,887 2,066,172 migrations 12,206 11,076 faults 163,704 149,544 cache-misses 4,801,186 10,398,067 sched:sched_move_numa 44 43 sched:sched_stick_numa 0 0 sched:sched_swap_numa 0 0 migrate:mm_migrate_pages 17 6 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 2261 3552 numa_hint_faults_local 1993 3347 numa_hit 25726 25611 numa_huge_pte_updates 0 0 numa_interleave 239 213 numa_local 25498 25583 numa_other 228 28 numa_pages_migrated 17 6 numa_pte_updates 2266 3535 perf stats 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 117,980,962 99,358,136 migrations 3,950,220 4,041,607 faults 736,979 749,653 cache-misses 224,976,072,879 225,562,543,251 sched:sched_move_numa 504 771 sched:sched_stick_numa 50 14 sched:sched_swap_numa 239 204 migrate:mm_migrate_pages 1,260 1,180 vmstat 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 18293 27409 numa_hint_faults_local 11969 20677 numa_hit 240854 239988 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 240851 239983 numa_other 3 5 numa_pages_migrated 1190 1016 numa_pte_updates 18106 27916 perf stats 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 61,053,158 60,899,307 migrations 551,586 544,668 faults 244,174 270,834 cache-misses 74,326,766,973 74,543,455,635 sched:sched_move_numa 344 735 sched:sched_stick_numa 24 25 sched:sched_swap_numa 140 174 migrate:mm_migrate_pages 568 816 vmstat 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 6461 11059 numa_hint_faults_local 2283 4733 numa_hit 35661 41384 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 35661 41383 numa_other 0 1 numa_pages_migrated 568 815 numa_pte_updates 6518 11323 Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Rik van Riel <riel@surriel.com> Acked-by: Mel Gorman <mgorman@techsingularity.net> Cc: Jirka Hladky <jhladky@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1537552141-27815-2-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-09-21 17:48:56 +00:00
WRITE_ONCE(best_rq->numa_migrate_on, 0);
sched: add tracepoints related to NUMA task migration This patch adds three tracepoints o trace_sched_move_numa when a task is moved to a node o trace_sched_swap_numa when a task is swapped with another task o trace_sched_stick_numa when a numa-related migration fails The tracepoints allow the NUMA scheduler activity to be monitored and the following high-level metrics can be calculated o NUMA migrated stuck nr trace_sched_stick_numa o NUMA migrated idle nr trace_sched_move_numa o NUMA migrated swapped nr trace_sched_swap_numa o NUMA local swapped trace_sched_swap_numa src_nid == dst_nid (should never happen) o NUMA remote swapped trace_sched_swap_numa src_nid != dst_nid (should == NUMA migrated swapped) o NUMA group swapped trace_sched_swap_numa src_ngid == dst_ngid Maybe a small number of these are acceptable but a high number would be a major surprise. It would be even worse if bounces are frequent. o NUMA avg task migs. Average number of migrations for tasks o NUMA stddev task mig Self-explanatory o NUMA max task migs. Maximum number of migrations for a single task In general the intent of the tracepoints is to help diagnose problems where automatic NUMA balancing appears to be doing an excessive amount of useless work. [akpm@linux-foundation.org: remove semicolon-after-if, repair coding-style] Signed-off-by: Mel Gorman <mgorman@suse.de> Reviewed-by: Rik van Riel <riel@redhat.com> Cc: Alex Thorlton <athorlton@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-01-21 23:51:03 +00:00
if (ret != 0)
trace_sched_stick_numa(p, env.src_cpu, env.best_task, env.best_cpu);
put_task_struct(env.best_task);
return ret;
}
/* Attempt to migrate a task to a CPU on the preferred node. */
static void numa_migrate_preferred(struct task_struct *p)
{
unsigned long interval = HZ;
/* This task has no NUMA fault statistics yet */
mm: replace all open encodings for NUMA_NO_NODE Patch series "Replace all open encodings for NUMA_NO_NODE", v3. All these places for replacement were found by running the following grep patterns on the entire kernel code. Please let me know if this might have missed some instances. This might also have replaced some false positives. I will appreciate suggestions, inputs and review. 1. git grep "nid == -1" 2. git grep "node == -1" 3. git grep "nid = -1" 4. git grep "node = -1" This patch (of 2): At present there are multiple places where invalid node number is encoded as -1. Even though implicitly understood it is always better to have macros in there. Replace these open encodings for an invalid node number with the global macro NUMA_NO_NODE. This helps remove NUMA related assumptions like 'invalid node' from various places redirecting them to a common definition. Link: http://lkml.kernel.org/r/1545127933-10711-2-git-send-email-anshuman.khandual@arm.com Signed-off-by: Anshuman Khandual <anshuman.khandual@arm.com> Reviewed-by: David Hildenbrand <david@redhat.com> Acked-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com> [ixgbe] Acked-by: Jens Axboe <axboe@kernel.dk> [mtip32xx] Acked-by: Vinod Koul <vkoul@kernel.org> [dmaengine.c] Acked-by: Michael Ellerman <mpe@ellerman.id.au> [powerpc] Acked-by: Doug Ledford <dledford@redhat.com> [drivers/infiniband] Cc: Joseph Qi <jiangqi903@gmail.com> Cc: Hans Verkuil <hverkuil@xs4all.nl> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:42:58 +00:00
if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults))
return;
/* Periodically retry migrating the task to the preferred node */
interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
Revert "sched/numa: Delay retrying placement for automatic NUMA balance after wake_affine()" This reverts commit 7347fc87dfe6b7315e74310ee1243dc222c68086. Srikar Dronamra pointed out that while the commit in question did show a performance improvement on ppc64, it did so at the cost of disabling active CPU migration by automatic NUMA balancing which was not the intent. The issue was that a serious flaw in the logic failed to ever active balance if SD_WAKE_AFFINE was disabled on scheduler domains. Even when it's enabled, the logic is still bizarre and against the original intent. Investigation showed that fixing the patch in either the way he suggested, using the correct comparison for jiffies values or introducing a new numa_migrate_deferred variable in task_struct all perform similarly to a revert with a mix of gains and losses depending on the workload, machine and socket count. The original intent of the commit was to handle a problem whereby wake_affine, idle balancing and automatic NUMA balancing disagree on the appropriate placement for a task. This was particularly true for cases where a single task was a massive waker of tasks but where wake_wide logic did not apply. This was particularly noticeable when a futex (a barrier) woke all worker threads and tried pulling the wakees to the waker nodes. In that specific case, it could be handled by tuning MPI or openMP appropriately, but the behavior is not illogical and was worth attempting to fix. However, the approach was wrong. Given that we're at rc4 and a fix is not obvious, it's better to play safe, revert this commit and retry later. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: efault@gmx.de Cc: ggherdovich@suse.cz Cc: hpa@zytor.com Cc: matt@codeblueprint.co.uk Cc: mpe@ellerman.id.au Link: http://lkml.kernel.org/r/20180509163115.6fnnyeg4vdm2ct4v@techsingularity.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-05-09 16:31:15 +00:00
p->numa_migrate_retry = jiffies + interval;
/* Success if task is already running on preferred CPU */
if (task_node(p) == p->numa_preferred_nid)
return;
/* Otherwise, try migrate to a CPU on the preferred node */
task_numa_migrate(p);
}
/*
sched/numa: Spread memory according to CPU and memory use The pseudo-interleaving in NUMA placement has a fundamental problem: using hard usage thresholds to spread memory equally between nodes can prevent workloads from converging, or keep memory "trapped" on nodes where the workload is barely running any more. In order for workloads to properly converge, the memory migration should not be stopped when nodes reach parity, but instead be distributed according to how heavily memory is used from each node. This way memory migration and task migration reinforce each other, instead of one putting the brakes on the other. Remove the hard thresholds from the pseudo-interleaving code, and instead use a more gradual policy on memory placement. This also seems to improve convergence of workloads that do not run flat out, but sleep in between bursts of activity. We still want to slow down NUMA scanning and migration once a workload has settled on a few actively used nodes, so keep the 3/4 hysteresis in place. Keep track of whether a workload is actively running on multiple nodes, so task_numa_migrate does a full scan of the system for better task placement. In the case of running 3 SPECjbb2005 instances on a 4 node system, this code seems to result in fairer distribution of memory between nodes, with more memory bandwidth for each instance. Signed-off-by: Rik van Riel <riel@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: mgorman@suse.de Link: http://lkml.kernel.org/r/20160125170739.2fc9a641@annuminas.surriel.com [ Minor readability tweaks. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-01-25 22:07:39 +00:00
* Find out how many nodes on the workload is actively running on. Do this by
* tracking the nodes from which NUMA hinting faults are triggered. This can
* be different from the set of nodes where the workload's memory is currently
* located.
*/
sched/numa: Spread memory according to CPU and memory use The pseudo-interleaving in NUMA placement has a fundamental problem: using hard usage thresholds to spread memory equally between nodes can prevent workloads from converging, or keep memory "trapped" on nodes where the workload is barely running any more. In order for workloads to properly converge, the memory migration should not be stopped when nodes reach parity, but instead be distributed according to how heavily memory is used from each node. This way memory migration and task migration reinforce each other, instead of one putting the brakes on the other. Remove the hard thresholds from the pseudo-interleaving code, and instead use a more gradual policy on memory placement. This also seems to improve convergence of workloads that do not run flat out, but sleep in between bursts of activity. We still want to slow down NUMA scanning and migration once a workload has settled on a few actively used nodes, so keep the 3/4 hysteresis in place. Keep track of whether a workload is actively running on multiple nodes, so task_numa_migrate does a full scan of the system for better task placement. In the case of running 3 SPECjbb2005 instances on a 4 node system, this code seems to result in fairer distribution of memory between nodes, with more memory bandwidth for each instance. Signed-off-by: Rik van Riel <riel@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: mgorman@suse.de Link: http://lkml.kernel.org/r/20160125170739.2fc9a641@annuminas.surriel.com [ Minor readability tweaks. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-01-25 22:07:39 +00:00
static void numa_group_count_active_nodes(struct numa_group *numa_group)
{
unsigned long faults, max_faults = 0;
sched/numa: Spread memory according to CPU and memory use The pseudo-interleaving in NUMA placement has a fundamental problem: using hard usage thresholds to spread memory equally between nodes can prevent workloads from converging, or keep memory "trapped" on nodes where the workload is barely running any more. In order for workloads to properly converge, the memory migration should not be stopped when nodes reach parity, but instead be distributed according to how heavily memory is used from each node. This way memory migration and task migration reinforce each other, instead of one putting the brakes on the other. Remove the hard thresholds from the pseudo-interleaving code, and instead use a more gradual policy on memory placement. This also seems to improve convergence of workloads that do not run flat out, but sleep in between bursts of activity. We still want to slow down NUMA scanning and migration once a workload has settled on a few actively used nodes, so keep the 3/4 hysteresis in place. Keep track of whether a workload is actively running on multiple nodes, so task_numa_migrate does a full scan of the system for better task placement. In the case of running 3 SPECjbb2005 instances on a 4 node system, this code seems to result in fairer distribution of memory between nodes, with more memory bandwidth for each instance. Signed-off-by: Rik van Riel <riel@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: mgorman@suse.de Link: http://lkml.kernel.org/r/20160125170739.2fc9a641@annuminas.surriel.com [ Minor readability tweaks. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-01-25 22:07:39 +00:00
int nid, active_nodes = 0;
for_each_online_node(nid) {
faults = group_faults_cpu(numa_group, nid);
if (faults > max_faults)
max_faults = faults;
}
for_each_online_node(nid) {
faults = group_faults_cpu(numa_group, nid);
sched/numa: Spread memory according to CPU and memory use The pseudo-interleaving in NUMA placement has a fundamental problem: using hard usage thresholds to spread memory equally between nodes can prevent workloads from converging, or keep memory "trapped" on nodes where the workload is barely running any more. In order for workloads to properly converge, the memory migration should not be stopped when nodes reach parity, but instead be distributed according to how heavily memory is used from each node. This way memory migration and task migration reinforce each other, instead of one putting the brakes on the other. Remove the hard thresholds from the pseudo-interleaving code, and instead use a more gradual policy on memory placement. This also seems to improve convergence of workloads that do not run flat out, but sleep in between bursts of activity. We still want to slow down NUMA scanning and migration once a workload has settled on a few actively used nodes, so keep the 3/4 hysteresis in place. Keep track of whether a workload is actively running on multiple nodes, so task_numa_migrate does a full scan of the system for better task placement. In the case of running 3 SPECjbb2005 instances on a 4 node system, this code seems to result in fairer distribution of memory between nodes, with more memory bandwidth for each instance. Signed-off-by: Rik van Riel <riel@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: mgorman@suse.de Link: http://lkml.kernel.org/r/20160125170739.2fc9a641@annuminas.surriel.com [ Minor readability tweaks. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-01-25 22:07:39 +00:00
if (faults * ACTIVE_NODE_FRACTION > max_faults)
active_nodes++;
}
sched/numa: Spread memory according to CPU and memory use The pseudo-interleaving in NUMA placement has a fundamental problem: using hard usage thresholds to spread memory equally between nodes can prevent workloads from converging, or keep memory "trapped" on nodes where the workload is barely running any more. In order for workloads to properly converge, the memory migration should not be stopped when nodes reach parity, but instead be distributed according to how heavily memory is used from each node. This way memory migration and task migration reinforce each other, instead of one putting the brakes on the other. Remove the hard thresholds from the pseudo-interleaving code, and instead use a more gradual policy on memory placement. This also seems to improve convergence of workloads that do not run flat out, but sleep in between bursts of activity. We still want to slow down NUMA scanning and migration once a workload has settled on a few actively used nodes, so keep the 3/4 hysteresis in place. Keep track of whether a workload is actively running on multiple nodes, so task_numa_migrate does a full scan of the system for better task placement. In the case of running 3 SPECjbb2005 instances on a 4 node system, this code seems to result in fairer distribution of memory between nodes, with more memory bandwidth for each instance. Signed-off-by: Rik van Riel <riel@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: mgorman@suse.de Link: http://lkml.kernel.org/r/20160125170739.2fc9a641@annuminas.surriel.com [ Minor readability tweaks. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-01-25 22:07:39 +00:00
numa_group->max_faults_cpu = max_faults;
numa_group->active_nodes = active_nodes;
}
/*
* When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
* increments. The more local the fault statistics are, the higher the scan
* period will be for the next scan window. If local/(local+remote) ratio is
* below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
* the scan period will decrease. Aim for 70% local accesses.
*/
#define NUMA_PERIOD_SLOTS 10
#define NUMA_PERIOD_THRESHOLD 7
/*
* Increase the scan period (slow down scanning) if the majority of
* our memory is already on our local node, or if the majority of
* the page accesses are shared with other processes.
* Otherwise, decrease the scan period.
*/
static void update_task_scan_period(struct task_struct *p,
unsigned long shared, unsigned long private)
{
unsigned int period_slot;
int lr_ratio, ps_ratio;
int diff;
unsigned long remote = p->numa_faults_locality[0];
unsigned long local = p->numa_faults_locality[1];
/*
* If there were no record hinting faults then either the task is
* completely idle or all activity is areas that are not of interest
mm: numa: slow PTE scan rate if migration failures occur Dave Chinner reported the following on https://lkml.org/lkml/2015/3/1/226 Across the board the 4.0-rc1 numbers are much slower, and the degradation is far worse when using the large memory footprint configs. Perf points straight at the cause - this is from 4.0-rc1 on the "-o bhash=101073" config: - 56.07% 56.07% [kernel] [k] default_send_IPI_mask_sequence_phys - default_send_IPI_mask_sequence_phys - 99.99% physflat_send_IPI_mask - 99.37% native_send_call_func_ipi smp_call_function_many - native_flush_tlb_others - 99.85% flush_tlb_page ptep_clear_flush try_to_unmap_one rmap_walk try_to_unmap migrate_pages migrate_misplaced_page - handle_mm_fault - 99.73% __do_page_fault trace_do_page_fault do_async_page_fault + async_page_fault 0.63% native_send_call_func_single_ipi generic_exec_single smp_call_function_single This is showing excessive migration activity even though excessive migrations are meant to get throttled. Normally, the scan rate is tuned on a per-task basis depending on the locality of faults. However, if migrations fail for any reason then the PTE scanner may scan faster if the faults continue to be remote. This means there is higher system CPU overhead and fault trapping at exactly the time we know that migrations cannot happen. This patch tracks when migration failures occur and slows the PTE scanner. Signed-off-by: Mel Gorman <mgorman@suse.de> Reported-by: Dave Chinner <david@fromorbit.com> Tested-by: Dave Chinner <david@fromorbit.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Aneesh Kumar <aneesh.kumar@linux.vnet.ibm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-03-25 22:55:42 +00:00
* to automatic numa balancing. Related to that, if there were failed
* migration then it implies we are migrating too quickly or the local
* node is overloaded. In either case, scan slower
*/
mm: numa: slow PTE scan rate if migration failures occur Dave Chinner reported the following on https://lkml.org/lkml/2015/3/1/226 Across the board the 4.0-rc1 numbers are much slower, and the degradation is far worse when using the large memory footprint configs. Perf points straight at the cause - this is from 4.0-rc1 on the "-o bhash=101073" config: - 56.07% 56.07% [kernel] [k] default_send_IPI_mask_sequence_phys - default_send_IPI_mask_sequence_phys - 99.99% physflat_send_IPI_mask - 99.37% native_send_call_func_ipi smp_call_function_many - native_flush_tlb_others - 99.85% flush_tlb_page ptep_clear_flush try_to_unmap_one rmap_walk try_to_unmap migrate_pages migrate_misplaced_page - handle_mm_fault - 99.73% __do_page_fault trace_do_page_fault do_async_page_fault + async_page_fault 0.63% native_send_call_func_single_ipi generic_exec_single smp_call_function_single This is showing excessive migration activity even though excessive migrations are meant to get throttled. Normally, the scan rate is tuned on a per-task basis depending on the locality of faults. However, if migrations fail for any reason then the PTE scanner may scan faster if the faults continue to be remote. This means there is higher system CPU overhead and fault trapping at exactly the time we know that migrations cannot happen. This patch tracks when migration failures occur and slows the PTE scanner. Signed-off-by: Mel Gorman <mgorman@suse.de> Reported-by: Dave Chinner <david@fromorbit.com> Tested-by: Dave Chinner <david@fromorbit.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Aneesh Kumar <aneesh.kumar@linux.vnet.ibm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-03-25 22:55:42 +00:00
if (local + shared == 0 || p->numa_faults_locality[2]) {
p->numa_scan_period = min(p->numa_scan_period_max,
p->numa_scan_period << 1);
p->mm->numa_next_scan = jiffies +
msecs_to_jiffies(p->numa_scan_period);
return;
}
/*
* Prepare to scale scan period relative to the current period.
* == NUMA_PERIOD_THRESHOLD scan period stays the same
* < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
* >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
*/
period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
/*
* Most memory accesses are local. There is no need to
* do fast NUMA scanning, since memory is already local.
*/
int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
if (!slot)
slot = 1;
diff = slot * period_slot;
} else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
/*
* Most memory accesses are shared with other tasks.
* There is no point in continuing fast NUMA scanning,
* since other tasks may just move the memory elsewhere.
*/
int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
if (!slot)
slot = 1;
diff = slot * period_slot;
} else {
/*
* Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
* yet they are not on the local NUMA node. Speed up
* NUMA scanning to get the memory moved over.
*/
int ratio = max(lr_ratio, ps_ratio);
diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
}
p->numa_scan_period = clamp(p->numa_scan_period + diff,
task_scan_min(p), task_scan_max(p));
memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
}
sched/numa: Normalize faults_cpu stats and weigh by CPU use Tracing the code that decides the active nodes has made it abundantly clear that the naive implementation of the faults_from code has issues. Specifically, the garbage collector in some workloads will access orders of magnitudes more memory than the threads that do all the active work. This resulted in the node with the garbage collector being marked the only active node in the group. This issue is avoided if we weigh the statistics by CPU use of each task in the numa group, instead of by how many faults each thread has occurred. To achieve this, we normalize the number of faults to the fraction of faults that occurred on each node, and then multiply that fraction by the fraction of CPU time the task has used since the last time task_numa_placement was invoked. This way the nodes in the active node mask will be the ones where the tasks from the numa group are most actively running, and the influence of eg. the garbage collector and other do-little threads is properly minimized. On a 4 node system, using CPU use statistics calculated over a longer interval results in about 1% fewer page migrations with two 32-warehouse specjbb runs on a 4 node system, and about 5% fewer page migrations, as well as 1% better throughput, with two 8-warehouse specjbb runs, as compared with the shorter term statistics kept by the scheduler. Signed-off-by: Rik van Riel <riel@redhat.com> Acked-by: Mel Gorman <mgorman@suse.de> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Cc: Chegu Vinod <chegu_vinod@hp.com> Link: http://lkml.kernel.org/r/1390860228-21539-7-git-send-email-riel@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-01-27 22:03:45 +00:00
/*
* Get the fraction of time the task has been running since the last
* NUMA placement cycle. The scheduler keeps similar statistics, but
* decays those on a 32ms period, which is orders of magnitude off
* from the dozens-of-seconds NUMA balancing period. Use the scheduler
* stats only if the task is so new there are no NUMA statistics yet.
*/
static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
{
u64 runtime, delta, now;
/* Use the start of this time slice to avoid calculations. */
now = p->se.exec_start;
runtime = p->se.sum_exec_runtime;
if (p->last_task_numa_placement) {
delta = runtime - p->last_sum_exec_runtime;
*period = now - p->last_task_numa_placement;
/* Avoid time going backwards, prevent potential divide error: */
if (unlikely((s64)*period < 0))
*period = 0;
sched/numa: Normalize faults_cpu stats and weigh by CPU use Tracing the code that decides the active nodes has made it abundantly clear that the naive implementation of the faults_from code has issues. Specifically, the garbage collector in some workloads will access orders of magnitudes more memory than the threads that do all the active work. This resulted in the node with the garbage collector being marked the only active node in the group. This issue is avoided if we weigh the statistics by CPU use of each task in the numa group, instead of by how many faults each thread has occurred. To achieve this, we normalize the number of faults to the fraction of faults that occurred on each node, and then multiply that fraction by the fraction of CPU time the task has used since the last time task_numa_placement was invoked. This way the nodes in the active node mask will be the ones where the tasks from the numa group are most actively running, and the influence of eg. the garbage collector and other do-little threads is properly minimized. On a 4 node system, using CPU use statistics calculated over a longer interval results in about 1% fewer page migrations with two 32-warehouse specjbb runs on a 4 node system, and about 5% fewer page migrations, as well as 1% better throughput, with two 8-warehouse specjbb runs, as compared with the shorter term statistics kept by the scheduler. Signed-off-by: Rik van Riel <riel@redhat.com> Acked-by: Mel Gorman <mgorman@suse.de> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Cc: Chegu Vinod <chegu_vinod@hp.com> Link: http://lkml.kernel.org/r/1390860228-21539-7-git-send-email-riel@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-01-27 22:03:45 +00:00
} else {
delta = p->se.avg.load_sum;
sched/fair: Rewrite runnable load and utilization average tracking The idea of runnable load average (let runnable time contribute to weight) was proposed by Paul Turner and Ben Segall, and it is still followed by this rewrite. This rewrite aims to solve the following issues: 1. cfs_rq's load average (namely runnable_load_avg and blocked_load_avg) is updated at the granularity of an entity at a time, which results in the cfs_rq's load average is stale or partially updated: at any time, only one entity is up to date, all other entities are effectively lagging behind. This is undesirable. To illustrate, if we have n runnable entities in the cfs_rq, as time elapses, they certainly become outdated: t0: cfs_rq { e1_old, e2_old, ..., en_old } and when we update: t1: update e1, then we have cfs_rq { e1_new, e2_old, ..., en_old } t2: update e2, then we have cfs_rq { e1_old, e2_new, ..., en_old } ... We solve this by combining all runnable entities' load averages together in cfs_rq's avg, and update the cfs_rq's avg as a whole. This is based on the fact that if we regard the update as a function, then: w * update(e) = update(w * e) and update(e1) + update(e2) = update(e1 + e2), then w1 * update(e1) + w2 * update(e2) = update(w1 * e1 + w2 * e2) therefore, by this rewrite, we have an entirely updated cfs_rq at the time we update it: t1: update cfs_rq { e1_new, e2_new, ..., en_new } t2: update cfs_rq { e1_new, e2_new, ..., en_new } ... 2. cfs_rq's load average is different between top rq->cfs_rq and other task_group's per CPU cfs_rqs in whether or not blocked_load_average contributes to the load. The basic idea behind runnable load average (the same for utilization) is that the blocked state is taken into account as opposed to only accounting for the currently runnable state. Therefore, the average should include both the runnable/running and blocked load averages. This rewrite does that. In addition, we also combine runnable/running and blocked averages of all entities into the cfs_rq's average, and update it together at once. This is based on the fact that: update(runnable) + update(blocked) = update(runnable + blocked) This significantly reduces the code as we don't need to separately maintain/update runnable/running load and blocked load. 3. How task_group entities' share is calculated is complex and imprecise. We reduce the complexity in this rewrite to allow a very simple rule: the task_group's load_avg is aggregated from its per CPU cfs_rqs's load_avgs. Then group entity's weight is simply proportional to its own cfs_rq's load_avg / task_group's load_avg. To illustrate, if a task_group has { cfs_rq1, cfs_rq2, ..., cfs_rqn }, then, task_group_avg = cfs_rq1_avg + cfs_rq2_avg + ... + cfs_rqn_avg, then cfs_rqx's entity's share = cfs_rqx_avg / task_group_avg * task_group's share To sum up, this rewrite in principle is equivalent to the current one, but fixes the issues described above. Turns out, it significantly reduces the code complexity and hence increases clarity and efficiency. In addition, the new averages are more smooth/continuous (no spurious spikes and valleys) and updated more consistently and quickly to reflect the load dynamics. As a result, we have less load tracking overhead, better performance, and especially better power efficiency due to more balanced load. Signed-off-by: Yuyang Du <yuyang.du@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: arjan@linux.intel.com Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: fengguang.wu@intel.com Cc: len.brown@intel.com Cc: morten.rasmussen@arm.com Cc: pjt@google.com Cc: rafael.j.wysocki@intel.com Cc: umgwanakikbuti@gmail.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1436918682-4971-3-git-send-email-yuyang.du@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-07-15 00:04:37 +00:00
*period = LOAD_AVG_MAX;
sched/numa: Normalize faults_cpu stats and weigh by CPU use Tracing the code that decides the active nodes has made it abundantly clear that the naive implementation of the faults_from code has issues. Specifically, the garbage collector in some workloads will access orders of magnitudes more memory than the threads that do all the active work. This resulted in the node with the garbage collector being marked the only active node in the group. This issue is avoided if we weigh the statistics by CPU use of each task in the numa group, instead of by how many faults each thread has occurred. To achieve this, we normalize the number of faults to the fraction of faults that occurred on each node, and then multiply that fraction by the fraction of CPU time the task has used since the last time task_numa_placement was invoked. This way the nodes in the active node mask will be the ones where the tasks from the numa group are most actively running, and the influence of eg. the garbage collector and other do-little threads is properly minimized. On a 4 node system, using CPU use statistics calculated over a longer interval results in about 1% fewer page migrations with two 32-warehouse specjbb runs on a 4 node system, and about 5% fewer page migrations, as well as 1% better throughput, with two 8-warehouse specjbb runs, as compared with the shorter term statistics kept by the scheduler. Signed-off-by: Rik van Riel <riel@redhat.com> Acked-by: Mel Gorman <mgorman@suse.de> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Cc: Chegu Vinod <chegu_vinod@hp.com> Link: http://lkml.kernel.org/r/1390860228-21539-7-git-send-email-riel@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-01-27 22:03:45 +00:00
}
p->last_sum_exec_runtime = runtime;
p->last_task_numa_placement = now;
return delta;
}
/*
* Determine the preferred nid for a task in a numa_group. This needs to
* be done in a way that produces consistent results with group_weight,
* otherwise workloads might not converge.
*/
static int preferred_group_nid(struct task_struct *p, int nid)
{
nodemask_t nodes;
int dist;
/* Direct connections between all NUMA nodes. */
if (sched_numa_topology_type == NUMA_DIRECT)
return nid;
/*
* On a system with glueless mesh NUMA topology, group_weight
* scores nodes according to the number of NUMA hinting faults on
* both the node itself, and on nearby nodes.
*/
if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
unsigned long score, max_score = 0;
int node, max_node = nid;
dist = sched_max_numa_distance;
for_each_online_node(node) {
score = group_weight(p, node, dist);
if (score > max_score) {
max_score = score;
max_node = node;
}
}
return max_node;
}
/*
* Finding the preferred nid in a system with NUMA backplane
* interconnect topology is more involved. The goal is to locate
* tasks from numa_groups near each other in the system, and
* untangle workloads from different sides of the system. This requires
* searching down the hierarchy of node groups, recursively searching
* inside the highest scoring group of nodes. The nodemask tricks
* keep the complexity of the search down.
*/
nodes = node_online_map;
for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
unsigned long max_faults = 0;
2015-01-23 08:25:38 +00:00
nodemask_t max_group = NODE_MASK_NONE;
int a, b;
/* Are there nodes at this distance from each other? */
if (!find_numa_distance(dist))
continue;
for_each_node_mask(a, nodes) {
unsigned long faults = 0;
nodemask_t this_group;
nodes_clear(this_group);
/* Sum group's NUMA faults; includes a==b case. */
for_each_node_mask(b, nodes) {
if (node_distance(a, b) < dist) {
faults += group_faults(p, b);
node_set(b, this_group);
node_clear(b, nodes);
}
}
/* Remember the top group. */
if (faults > max_faults) {
max_faults = faults;
max_group = this_group;
/*
* subtle: at the smallest distance there is
* just one node left in each "group", the
* winner is the preferred nid.
*/
nid = a;
}
}
/* Next round, evaluate the nodes within max_group. */
if (!max_faults)
break;
nodes = max_group;
}
return nid;
}
static void task_numa_placement(struct task_struct *p)
{
mm: replace all open encodings for NUMA_NO_NODE Patch series "Replace all open encodings for NUMA_NO_NODE", v3. All these places for replacement were found by running the following grep patterns on the entire kernel code. Please let me know if this might have missed some instances. This might also have replaced some false positives. I will appreciate suggestions, inputs and review. 1. git grep "nid == -1" 2. git grep "node == -1" 3. git grep "nid = -1" 4. git grep "node = -1" This patch (of 2): At present there are multiple places where invalid node number is encoded as -1. Even though implicitly understood it is always better to have macros in there. Replace these open encodings for an invalid node number with the global macro NUMA_NO_NODE. This helps remove NUMA related assumptions like 'invalid node' from various places redirecting them to a common definition. Link: http://lkml.kernel.org/r/1545127933-10711-2-git-send-email-anshuman.khandual@arm.com Signed-off-by: Anshuman Khandual <anshuman.khandual@arm.com> Reviewed-by: David Hildenbrand <david@redhat.com> Acked-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com> [ixgbe] Acked-by: Jens Axboe <axboe@kernel.dk> [mtip32xx] Acked-by: Vinod Koul <vkoul@kernel.org> [dmaengine.c] Acked-by: Michael Ellerman <mpe@ellerman.id.au> [powerpc] Acked-by: Doug Ledford <dledford@redhat.com> [drivers/infiniband] Cc: Joseph Qi <jiangqi903@gmail.com> Cc: Hans Verkuil <hverkuil@xs4all.nl> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:42:58 +00:00
int seq, nid, max_nid = NUMA_NO_NODE;
sched/numa: Use task faults only if numa_group is not yet set up When numa_group faults are available, task_numa_placement only uses numa_group faults to evaluate preferred node. However it still accounts task faults and even evaluates the preferred node just based on task faults just to discard it in favour of preferred node chosen on the basis of numa_group. Instead use task faults only if numa_group is not set. Running SPECjbb2005 on a 4 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 16 25549.6 25215.7 -1.30 1 73190 72107 -1.47 Running SPECjbb2005 on a 16 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 8 113437 113372 -0.05 1 196130 177403 -9.54 (numbers from v1 based on v4.17-rc5) Testcase Time: Min Max Avg StdDev numa01.sh Real: 506.35 794.46 599.06 104.26 numa01.sh Sys: 150.37 223.56 195.99 24.94 numa01.sh User: 43450.69 61752.04 49281.50 6635.33 numa02.sh Real: 60.33 62.40 61.31 0.90 numa02.sh Sys: 18.12 31.66 24.28 5.89 numa02.sh User: 5203.91 5325.32 5260.29 49.98 numa03.sh Real: 696.47 853.62 745.80 57.28 numa03.sh Sys: 85.68 123.71 97.89 13.48 numa03.sh User: 55978.45 66418.63 59254.94 3737.97 numa04.sh Real: 444.05 514.83 497.06 26.85 numa04.sh Sys: 230.39 375.79 316.23 48.58 numa04.sh User: 35403.12 41004.10 39720.80 2163.08 numa05.sh Real: 423.09 460.41 439.57 13.92 numa05.sh Sys: 287.38 480.15 369.37 68.52 numa05.sh User: 34732.12 38016.80 36255.85 1070.51 Testcase Time: Min Max Avg StdDev %Change numa01.sh Real: 478.45 565.90 515.11 30.87 16.29% numa01.sh Sys: 207.79 271.04 232.94 21.33 -15.8% numa01.sh User: 39763.93 47303.12 43210.73 2644.86 14.04% numa02.sh Real: 60.00 61.46 60.78 0.49 0.871% numa02.sh Sys: 15.71 25.31 20.69 3.42 17.35% numa02.sh User: 5175.92 5265.86 5235.97 32.82 0.464% numa03.sh Real: 776.42 834.85 806.01 23.22 -7.47% numa03.sh Sys: 114.43 128.75 121.65 5.49 -19.5% numa03.sh User: 60773.93 64855.25 62616.91 1576.39 -5.36% numa04.sh Real: 456.93 511.95 482.91 20.88 2.930% numa04.sh Sys: 178.09 460.89 356.86 94.58 -11.3% numa04.sh User: 36312.09 42553.24 39623.21 2247.96 0.246% numa05.sh Real: 393.98 493.48 436.61 35.59 0.677% numa05.sh Sys: 164.49 329.15 265.87 61.78 38.92% numa05.sh User: 33182.65 36654.53 35074.51 1187.71 3.368% Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mel Gorman <mgorman@techsingularity.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1529514181-9842-6-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-20 17:02:46 +00:00
unsigned long max_faults = 0;
unsigned long fault_types[2] = { 0, 0 };
sched/numa: Normalize faults_cpu stats and weigh by CPU use Tracing the code that decides the active nodes has made it abundantly clear that the naive implementation of the faults_from code has issues. Specifically, the garbage collector in some workloads will access orders of magnitudes more memory than the threads that do all the active work. This resulted in the node with the garbage collector being marked the only active node in the group. This issue is avoided if we weigh the statistics by CPU use of each task in the numa group, instead of by how many faults each thread has occurred. To achieve this, we normalize the number of faults to the fraction of faults that occurred on each node, and then multiply that fraction by the fraction of CPU time the task has used since the last time task_numa_placement was invoked. This way the nodes in the active node mask will be the ones where the tasks from the numa group are most actively running, and the influence of eg. the garbage collector and other do-little threads is properly minimized. On a 4 node system, using CPU use statistics calculated over a longer interval results in about 1% fewer page migrations with two 32-warehouse specjbb runs on a 4 node system, and about 5% fewer page migrations, as well as 1% better throughput, with two 8-warehouse specjbb runs, as compared with the shorter term statistics kept by the scheduler. Signed-off-by: Rik van Riel <riel@redhat.com> Acked-by: Mel Gorman <mgorman@suse.de> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Cc: Chegu Vinod <chegu_vinod@hp.com> Link: http://lkml.kernel.org/r/1390860228-21539-7-git-send-email-riel@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-01-27 22:03:45 +00:00
unsigned long total_faults;
u64 runtime, period;
spinlock_t *group_lock = NULL;
struct numa_group *ng;
/*
* The p->mm->numa_scan_seq field gets updated without
* exclusive access. Use READ_ONCE() here to ensure
* that the field is read in a single access:
*/
seq = READ_ONCE(p->mm->numa_scan_seq);
if (p->numa_scan_seq == seq)
return;
p->numa_scan_seq = seq;
sched/numa: Set the scan rate proportional to the memory usage of the task being scanned The NUMA PTE scan rate is controlled with a combination of the numa_balancing_scan_period_min, numa_balancing_scan_period_max and numa_balancing_scan_size. This scan rate is independent of the size of the task and as an aside it is further complicated by the fact that numa_balancing_scan_size controls how many pages are marked pte_numa and not how much virtual memory is scanned. In combination, it is almost impossible to meaningfully tune the min and max scan periods and reasoning about performance is complex when the time to complete a full scan is is partially a function of the tasks memory size. This patch alters the semantic of the min and max tunables to be about tuning the length time it takes to complete a scan of a tasks occupied virtual address space. Conceptually this is a lot easier to understand. There is a "sanity" check to ensure the scan rate is never extremely fast based on the amount of virtual memory that should be scanned in a second. The default of 2.5G seems arbitrary but it is to have the maximum scan rate after the patch roughly match the maximum scan rate before the patch was applied. On a similar note, numa_scan_period is in milliseconds and not jiffies. Properly placed pages slow the scanning rate but adding 10 jiffies to numa_scan_period means that the rate scanning slows depends on HZ which is confusing. Get rid of the jiffies_to_msec conversion and treat it as ms. Signed-off-by: Mel Gorman <mgorman@suse.de> Reviewed-by: Rik van Riel <riel@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1381141781-10992-18-git-send-email-mgorman@suse.de Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-10-07 10:28:55 +00:00
p->numa_scan_period_max = task_scan_max(p);
sched/numa: Normalize faults_cpu stats and weigh by CPU use Tracing the code that decides the active nodes has made it abundantly clear that the naive implementation of the faults_from code has issues. Specifically, the garbage collector in some workloads will access orders of magnitudes more memory than the threads that do all the active work. This resulted in the node with the garbage collector being marked the only active node in the group. This issue is avoided if we weigh the statistics by CPU use of each task in the numa group, instead of by how many faults each thread has occurred. To achieve this, we normalize the number of faults to the fraction of faults that occurred on each node, and then multiply that fraction by the fraction of CPU time the task has used since the last time task_numa_placement was invoked. This way the nodes in the active node mask will be the ones where the tasks from the numa group are most actively running, and the influence of eg. the garbage collector and other do-little threads is properly minimized. On a 4 node system, using CPU use statistics calculated over a longer interval results in about 1% fewer page migrations with two 32-warehouse specjbb runs on a 4 node system, and about 5% fewer page migrations, as well as 1% better throughput, with two 8-warehouse specjbb runs, as compared with the shorter term statistics kept by the scheduler. Signed-off-by: Rik van Riel <riel@redhat.com> Acked-by: Mel Gorman <mgorman@suse.de> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Cc: Chegu Vinod <chegu_vinod@hp.com> Link: http://lkml.kernel.org/r/1390860228-21539-7-git-send-email-riel@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-01-27 22:03:45 +00:00
total_faults = p->numa_faults_locality[0] +
p->numa_faults_locality[1];
runtime = numa_get_avg_runtime(p, &period);
/* If the task is part of a group prevent parallel updates to group stats */
ng = deref_curr_numa_group(p);
if (ng) {
group_lock = &ng->lock;
spin_lock_irq(group_lock);
}
/* Find the node with the highest number of faults */
for_each_online_node(nid) {
/* Keep track of the offsets in numa_faults array */
int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
unsigned long faults = 0, group_faults = 0;
int priv;
for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
sched/numa: Normalize faults_cpu stats and weigh by CPU use Tracing the code that decides the active nodes has made it abundantly clear that the naive implementation of the faults_from code has issues. Specifically, the garbage collector in some workloads will access orders of magnitudes more memory than the threads that do all the active work. This resulted in the node with the garbage collector being marked the only active node in the group. This issue is avoided if we weigh the statistics by CPU use of each task in the numa group, instead of by how many faults each thread has occurred. To achieve this, we normalize the number of faults to the fraction of faults that occurred on each node, and then multiply that fraction by the fraction of CPU time the task has used since the last time task_numa_placement was invoked. This way the nodes in the active node mask will be the ones where the tasks from the numa group are most actively running, and the influence of eg. the garbage collector and other do-little threads is properly minimized. On a 4 node system, using CPU use statistics calculated over a longer interval results in about 1% fewer page migrations with two 32-warehouse specjbb runs on a 4 node system, and about 5% fewer page migrations, as well as 1% better throughput, with two 8-warehouse specjbb runs, as compared with the shorter term statistics kept by the scheduler. Signed-off-by: Rik van Riel <riel@redhat.com> Acked-by: Mel Gorman <mgorman@suse.de> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Cc: Chegu Vinod <chegu_vinod@hp.com> Link: http://lkml.kernel.org/r/1390860228-21539-7-git-send-email-riel@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-01-27 22:03:45 +00:00
long diff, f_diff, f_weight;
mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
/* Decay existing window, copy faults since last scan */
diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
fault_types[priv] += p->numa_faults[membuf_idx];
p->numa_faults[membuf_idx] = 0;
sched/numa: Normalize faults_cpu stats and weigh by CPU use Tracing the code that decides the active nodes has made it abundantly clear that the naive implementation of the faults_from code has issues. Specifically, the garbage collector in some workloads will access orders of magnitudes more memory than the threads that do all the active work. This resulted in the node with the garbage collector being marked the only active node in the group. This issue is avoided if we weigh the statistics by CPU use of each task in the numa group, instead of by how many faults each thread has occurred. To achieve this, we normalize the number of faults to the fraction of faults that occurred on each node, and then multiply that fraction by the fraction of CPU time the task has used since the last time task_numa_placement was invoked. This way the nodes in the active node mask will be the ones where the tasks from the numa group are most actively running, and the influence of eg. the garbage collector and other do-little threads is properly minimized. On a 4 node system, using CPU use statistics calculated over a longer interval results in about 1% fewer page migrations with two 32-warehouse specjbb runs on a 4 node system, and about 5% fewer page migrations, as well as 1% better throughput, with two 8-warehouse specjbb runs, as compared with the shorter term statistics kept by the scheduler. Signed-off-by: Rik van Riel <riel@redhat.com> Acked-by: Mel Gorman <mgorman@suse.de> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Cc: Chegu Vinod <chegu_vinod@hp.com> Link: http://lkml.kernel.org/r/1390860228-21539-7-git-send-email-riel@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-01-27 22:03:45 +00:00
/*
* Normalize the faults_from, so all tasks in a group
* count according to CPU use, instead of by the raw
* number of faults. Tasks with little runtime have
* little over-all impact on throughput, and thus their
* faults are less important.
*/
f_weight = div64_u64(runtime << 16, period + 1);
f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
sched/numa: Normalize faults_cpu stats and weigh by CPU use Tracing the code that decides the active nodes has made it abundantly clear that the naive implementation of the faults_from code has issues. Specifically, the garbage collector in some workloads will access orders of magnitudes more memory than the threads that do all the active work. This resulted in the node with the garbage collector being marked the only active node in the group. This issue is avoided if we weigh the statistics by CPU use of each task in the numa group, instead of by how many faults each thread has occurred. To achieve this, we normalize the number of faults to the fraction of faults that occurred on each node, and then multiply that fraction by the fraction of CPU time the task has used since the last time task_numa_placement was invoked. This way the nodes in the active node mask will be the ones where the tasks from the numa group are most actively running, and the influence of eg. the garbage collector and other do-little threads is properly minimized. On a 4 node system, using CPU use statistics calculated over a longer interval results in about 1% fewer page migrations with two 32-warehouse specjbb runs on a 4 node system, and about 5% fewer page migrations, as well as 1% better throughput, with two 8-warehouse specjbb runs, as compared with the shorter term statistics kept by the scheduler. Signed-off-by: Rik van Riel <riel@redhat.com> Acked-by: Mel Gorman <mgorman@suse.de> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Cc: Chegu Vinod <chegu_vinod@hp.com> Link: http://lkml.kernel.org/r/1390860228-21539-7-git-send-email-riel@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-01-27 22:03:45 +00:00
(total_faults + 1);
f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
p->numa_faults[cpubuf_idx] = 0;
p->numa_faults[mem_idx] += diff;
p->numa_faults[cpu_idx] += f_diff;
faults += p->numa_faults[mem_idx];
p->total_numa_faults += diff;
if (ng) {
/*
* safe because we can only change our own group
*
* mem_idx represents the offset for a given
* nid and priv in a specific region because it
* is at the beginning of the numa_faults array.
*/
ng->faults[mem_idx] += diff;
ng->faults_cpu[mem_idx] += f_diff;
ng->total_faults += diff;
group_faults += ng->faults[mem_idx];
}
}
if (!ng) {
sched/numa: Use task faults only if numa_group is not yet set up When numa_group faults are available, task_numa_placement only uses numa_group faults to evaluate preferred node. However it still accounts task faults and even evaluates the preferred node just based on task faults just to discard it in favour of preferred node chosen on the basis of numa_group. Instead use task faults only if numa_group is not set. Running SPECjbb2005 on a 4 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 16 25549.6 25215.7 -1.30 1 73190 72107 -1.47 Running SPECjbb2005 on a 16 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 8 113437 113372 -0.05 1 196130 177403 -9.54 (numbers from v1 based on v4.17-rc5) Testcase Time: Min Max Avg StdDev numa01.sh Real: 506.35 794.46 599.06 104.26 numa01.sh Sys: 150.37 223.56 195.99 24.94 numa01.sh User: 43450.69 61752.04 49281.50 6635.33 numa02.sh Real: 60.33 62.40 61.31 0.90 numa02.sh Sys: 18.12 31.66 24.28 5.89 numa02.sh User: 5203.91 5325.32 5260.29 49.98 numa03.sh Real: 696.47 853.62 745.80 57.28 numa03.sh Sys: 85.68 123.71 97.89 13.48 numa03.sh User: 55978.45 66418.63 59254.94 3737.97 numa04.sh Real: 444.05 514.83 497.06 26.85 numa04.sh Sys: 230.39 375.79 316.23 48.58 numa04.sh User: 35403.12 41004.10 39720.80 2163.08 numa05.sh Real: 423.09 460.41 439.57 13.92 numa05.sh Sys: 287.38 480.15 369.37 68.52 numa05.sh User: 34732.12 38016.80 36255.85 1070.51 Testcase Time: Min Max Avg StdDev %Change numa01.sh Real: 478.45 565.90 515.11 30.87 16.29% numa01.sh Sys: 207.79 271.04 232.94 21.33 -15.8% numa01.sh User: 39763.93 47303.12 43210.73 2644.86 14.04% numa02.sh Real: 60.00 61.46 60.78 0.49 0.871% numa02.sh Sys: 15.71 25.31 20.69 3.42 17.35% numa02.sh User: 5175.92 5265.86 5235.97 32.82 0.464% numa03.sh Real: 776.42 834.85 806.01 23.22 -7.47% numa03.sh Sys: 114.43 128.75 121.65 5.49 -19.5% numa03.sh User: 60773.93 64855.25 62616.91 1576.39 -5.36% numa04.sh Real: 456.93 511.95 482.91 20.88 2.930% numa04.sh Sys: 178.09 460.89 356.86 94.58 -11.3% numa04.sh User: 36312.09 42553.24 39623.21 2247.96 0.246% numa05.sh Real: 393.98 493.48 436.61 35.59 0.677% numa05.sh Sys: 164.49 329.15 265.87 61.78 38.92% numa05.sh User: 33182.65 36654.53 35074.51 1187.71 3.368% Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mel Gorman <mgorman@techsingularity.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1529514181-9842-6-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-20 17:02:46 +00:00
if (faults > max_faults) {
max_faults = faults;
max_nid = nid;
}
} else if (group_faults > max_faults) {
max_faults = group_faults;
max_nid = nid;
}
}
if (ng) {
numa_group_count_active_nodes(ng);
spin_unlock_irq(group_lock);
sched/numa: Use task faults only if numa_group is not yet set up When numa_group faults are available, task_numa_placement only uses numa_group faults to evaluate preferred node. However it still accounts task faults and even evaluates the preferred node just based on task faults just to discard it in favour of preferred node chosen on the basis of numa_group. Instead use task faults only if numa_group is not set. Running SPECjbb2005 on a 4 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 16 25549.6 25215.7 -1.30 1 73190 72107 -1.47 Running SPECjbb2005 on a 16 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 8 113437 113372 -0.05 1 196130 177403 -9.54 (numbers from v1 based on v4.17-rc5) Testcase Time: Min Max Avg StdDev numa01.sh Real: 506.35 794.46 599.06 104.26 numa01.sh Sys: 150.37 223.56 195.99 24.94 numa01.sh User: 43450.69 61752.04 49281.50 6635.33 numa02.sh Real: 60.33 62.40 61.31 0.90 numa02.sh Sys: 18.12 31.66 24.28 5.89 numa02.sh User: 5203.91 5325.32 5260.29 49.98 numa03.sh Real: 696.47 853.62 745.80 57.28 numa03.sh Sys: 85.68 123.71 97.89 13.48 numa03.sh User: 55978.45 66418.63 59254.94 3737.97 numa04.sh Real: 444.05 514.83 497.06 26.85 numa04.sh Sys: 230.39 375.79 316.23 48.58 numa04.sh User: 35403.12 41004.10 39720.80 2163.08 numa05.sh Real: 423.09 460.41 439.57 13.92 numa05.sh Sys: 287.38 480.15 369.37 68.52 numa05.sh User: 34732.12 38016.80 36255.85 1070.51 Testcase Time: Min Max Avg StdDev %Change numa01.sh Real: 478.45 565.90 515.11 30.87 16.29% numa01.sh Sys: 207.79 271.04 232.94 21.33 -15.8% numa01.sh User: 39763.93 47303.12 43210.73 2644.86 14.04% numa02.sh Real: 60.00 61.46 60.78 0.49 0.871% numa02.sh Sys: 15.71 25.31 20.69 3.42 17.35% numa02.sh User: 5175.92 5265.86 5235.97 32.82 0.464% numa03.sh Real: 776.42 834.85 806.01 23.22 -7.47% numa03.sh Sys: 114.43 128.75 121.65 5.49 -19.5% numa03.sh User: 60773.93 64855.25 62616.91 1576.39 -5.36% numa04.sh Real: 456.93 511.95 482.91 20.88 2.930% numa04.sh Sys: 178.09 460.89 356.86 94.58 -11.3% numa04.sh User: 36312.09 42553.24 39623.21 2247.96 0.246% numa05.sh Real: 393.98 493.48 436.61 35.59 0.677% numa05.sh Sys: 164.49 329.15 265.87 61.78 38.92% numa05.sh User: 33182.65 36654.53 35074.51 1187.71 3.368% Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mel Gorman <mgorman@techsingularity.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1529514181-9842-6-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-20 17:02:46 +00:00
max_nid = preferred_group_nid(p, max_nid);
}
if (max_faults) {
/* Set the new preferred node */
if (max_nid != p->numa_preferred_nid)
sched_setnuma(p, max_nid);
}
update_task_scan_period(p, fault_types[0], fault_types[1]);
}
static inline int get_numa_group(struct numa_group *grp)
{
sched/fair: Convert numa_group.refcount to refcount_t atomic_t variables are currently used to implement reference counters with the following properties: - counter is initialized to 1 using atomic_set() - a resource is freed upon counter reaching zero - once counter reaches zero, its further increments aren't allowed - counter schema uses basic atomic operations (set, inc, inc_not_zero, dec_and_test, etc.) Such atomic variables should be converted to a newly provided refcount_t type and API that prevents accidental counter overflows and underflows. This is important since overflows and underflows can lead to use-after-free situation and be exploitable. The variable numa_group.refcount is used as pure reference counter. Convert it to refcount_t and fix up the operations. ** Important note for maintainers: Some functions from refcount_t API defined in lib/refcount.c have different memory ordering guarantees than their atomic counterparts. The full comparison can be seen in https://lkml.org/lkml/2017/11/15/57 and it is hopefully soon in state to be merged to the documentation tree. Normally the differences should not matter since refcount_t provides enough guarantees to satisfy the refcounting use cases, but in some rare cases it might matter. Please double check that you don't have some undocumented memory guarantees for this variable usage. For the numa_group.refcount it might make a difference in following places: - get_numa_group(): increment in refcount_inc_not_zero() only guarantees control dependency on success vs. fully ordered atomic counterpart - put_numa_group(): decrement in refcount_dec_and_test() only provides RELEASE ordering and control dependency on success vs. fully ordered atomic counterpart Suggested-by: Kees Cook <keescook@chromium.org> Signed-off-by: Elena Reshetova <elena.reshetova@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: David Windsor <dwindsor@gmail.com> Reviewed-by: Hans Liljestrand <ishkamiel@gmail.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: akpm@linux-foundation.org Cc: viro@zeniv.linux.org.uk Link: https://lkml.kernel.org/r/1547814450-18902-4-git-send-email-elena.reshetova@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-01-18 12:27:28 +00:00
return refcount_inc_not_zero(&grp->refcount);
}
static inline void put_numa_group(struct numa_group *grp)
{
sched/fair: Convert numa_group.refcount to refcount_t atomic_t variables are currently used to implement reference counters with the following properties: - counter is initialized to 1 using atomic_set() - a resource is freed upon counter reaching zero - once counter reaches zero, its further increments aren't allowed - counter schema uses basic atomic operations (set, inc, inc_not_zero, dec_and_test, etc.) Such atomic variables should be converted to a newly provided refcount_t type and API that prevents accidental counter overflows and underflows. This is important since overflows and underflows can lead to use-after-free situation and be exploitable. The variable numa_group.refcount is used as pure reference counter. Convert it to refcount_t and fix up the operations. ** Important note for maintainers: Some functions from refcount_t API defined in lib/refcount.c have different memory ordering guarantees than their atomic counterparts. The full comparison can be seen in https://lkml.org/lkml/2017/11/15/57 and it is hopefully soon in state to be merged to the documentation tree. Normally the differences should not matter since refcount_t provides enough guarantees to satisfy the refcounting use cases, but in some rare cases it might matter. Please double check that you don't have some undocumented memory guarantees for this variable usage. For the numa_group.refcount it might make a difference in following places: - get_numa_group(): increment in refcount_inc_not_zero() only guarantees control dependency on success vs. fully ordered atomic counterpart - put_numa_group(): decrement in refcount_dec_and_test() only provides RELEASE ordering and control dependency on success vs. fully ordered atomic counterpart Suggested-by: Kees Cook <keescook@chromium.org> Signed-off-by: Elena Reshetova <elena.reshetova@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: David Windsor <dwindsor@gmail.com> Reviewed-by: Hans Liljestrand <ishkamiel@gmail.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: akpm@linux-foundation.org Cc: viro@zeniv.linux.org.uk Link: https://lkml.kernel.org/r/1547814450-18902-4-git-send-email-elena.reshetova@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-01-18 12:27:28 +00:00
if (refcount_dec_and_test(&grp->refcount))
kfree_rcu(grp, rcu);
}
static void task_numa_group(struct task_struct *p, int cpupid, int flags,
int *priv)
{
struct numa_group *grp, *my_grp;
struct task_struct *tsk;
bool join = false;
int cpu = cpupid_to_cpu(cpupid);
int i;
if (unlikely(!deref_curr_numa_group(p))) {
unsigned int size = sizeof(struct numa_group) +
4*nr_node_ids*sizeof(unsigned long);
grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
if (!grp)
return;
sched/fair: Convert numa_group.refcount to refcount_t atomic_t variables are currently used to implement reference counters with the following properties: - counter is initialized to 1 using atomic_set() - a resource is freed upon counter reaching zero - once counter reaches zero, its further increments aren't allowed - counter schema uses basic atomic operations (set, inc, inc_not_zero, dec_and_test, etc.) Such atomic variables should be converted to a newly provided refcount_t type and API that prevents accidental counter overflows and underflows. This is important since overflows and underflows can lead to use-after-free situation and be exploitable. The variable numa_group.refcount is used as pure reference counter. Convert it to refcount_t and fix up the operations. ** Important note for maintainers: Some functions from refcount_t API defined in lib/refcount.c have different memory ordering guarantees than their atomic counterparts. The full comparison can be seen in https://lkml.org/lkml/2017/11/15/57 and it is hopefully soon in state to be merged to the documentation tree. Normally the differences should not matter since refcount_t provides enough guarantees to satisfy the refcounting use cases, but in some rare cases it might matter. Please double check that you don't have some undocumented memory guarantees for this variable usage. For the numa_group.refcount it might make a difference in following places: - get_numa_group(): increment in refcount_inc_not_zero() only guarantees control dependency on success vs. fully ordered atomic counterpart - put_numa_group(): decrement in refcount_dec_and_test() only provides RELEASE ordering and control dependency on success vs. fully ordered atomic counterpart Suggested-by: Kees Cook <keescook@chromium.org> Signed-off-by: Elena Reshetova <elena.reshetova@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: David Windsor <dwindsor@gmail.com> Reviewed-by: Hans Liljestrand <ishkamiel@gmail.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: akpm@linux-foundation.org Cc: viro@zeniv.linux.org.uk Link: https://lkml.kernel.org/r/1547814450-18902-4-git-send-email-elena.reshetova@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-01-18 12:27:28 +00:00
refcount_set(&grp->refcount, 1);
sched/numa: Spread memory according to CPU and memory use The pseudo-interleaving in NUMA placement has a fundamental problem: using hard usage thresholds to spread memory equally between nodes can prevent workloads from converging, or keep memory "trapped" on nodes where the workload is barely running any more. In order for workloads to properly converge, the memory migration should not be stopped when nodes reach parity, but instead be distributed according to how heavily memory is used from each node. This way memory migration and task migration reinforce each other, instead of one putting the brakes on the other. Remove the hard thresholds from the pseudo-interleaving code, and instead use a more gradual policy on memory placement. This also seems to improve convergence of workloads that do not run flat out, but sleep in between bursts of activity. We still want to slow down NUMA scanning and migration once a workload has settled on a few actively used nodes, so keep the 3/4 hysteresis in place. Keep track of whether a workload is actively running on multiple nodes, so task_numa_migrate does a full scan of the system for better task placement. In the case of running 3 SPECjbb2005 instances on a 4 node system, this code seems to result in fairer distribution of memory between nodes, with more memory bandwidth for each instance. Signed-off-by: Rik van Riel <riel@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: mgorman@suse.de Link: http://lkml.kernel.org/r/20160125170739.2fc9a641@annuminas.surriel.com [ Minor readability tweaks. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-01-25 22:07:39 +00:00
grp->active_nodes = 1;
grp->max_faults_cpu = 0;
spin_lock_init(&grp->lock);
grp->gid = p->pid;
/* Second half of the array tracks nids where faults happen */
grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
nr_node_ids;
for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
grp->faults[i] = p->numa_faults[i];
grp->total_faults = p->total_numa_faults;
grp->nr_tasks++;
rcu_assign_pointer(p->numa_group, grp);
}
rcu_read_lock();
tsk = READ_ONCE(cpu_rq(cpu)->curr);
if (!cpupid_match_pid(tsk, cpupid))
goto no_join;
grp = rcu_dereference(tsk->numa_group);
if (!grp)
goto no_join;
my_grp = deref_curr_numa_group(p);
if (grp == my_grp)
goto no_join;
/*
* Only join the other group if its bigger; if we're the bigger group,
* the other task will join us.
*/
if (my_grp->nr_tasks > grp->nr_tasks)
goto no_join;
/*
* Tie-break on the grp address.
*/
if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
goto no_join;
/* Always join threads in the same process. */
if (tsk->mm == current->mm)
join = true;
/* Simple filter to avoid false positives due to PID collisions */
if (flags & TNF_SHARED)
join = true;
/* Update priv based on whether false sharing was detected */
*priv = !join;
if (join && !get_numa_group(grp))
goto no_join;
rcu_read_unlock();
if (!join)
return;
BUG_ON(irqs_disabled());
double_lock_irq(&my_grp->lock, &grp->lock);
for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
my_grp->faults[i] -= p->numa_faults[i];
grp->faults[i] += p->numa_faults[i];
}
my_grp->total_faults -= p->total_numa_faults;
grp->total_faults += p->total_numa_faults;
my_grp->nr_tasks--;
grp->nr_tasks++;
spin_unlock(&my_grp->lock);
spin_unlock_irq(&grp->lock);
rcu_assign_pointer(p->numa_group, grp);
put_numa_group(my_grp);
return;
no_join:
rcu_read_unlock();
return;
}
/*
* Get rid of NUMA staticstics associated with a task (either current or dead).
* If @final is set, the task is dead and has reached refcount zero, so we can
* safely free all relevant data structures. Otherwise, there might be
* concurrent reads from places like load balancing and procfs, and we should
* reset the data back to default state without freeing ->numa_faults.
*/
void task_numa_free(struct task_struct *p, bool final)
{
/* safe: p either is current or is being freed by current */
struct numa_group *grp = rcu_dereference_raw(p->numa_group);
unsigned long *numa_faults = p->numa_faults;
unsigned long flags;
int i;
if (!numa_faults)
return;
if (grp) {
spin_lock_irqsave(&grp->lock, flags);
for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
grp->faults[i] -= p->numa_faults[i];
grp->total_faults -= p->total_numa_faults;
grp->nr_tasks--;
spin_unlock_irqrestore(&grp->lock, flags);
RCU_INIT_POINTER(p->numa_group, NULL);
put_numa_group(grp);
}
if (final) {
p->numa_faults = NULL;
kfree(numa_faults);
} else {
p->total_numa_faults = 0;
for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
numa_faults[i] = 0;
}
}
/*
* Got a PROT_NONE fault for a page on @node.
*/
void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
{
struct task_struct *p = current;
bool migrated = flags & TNF_MIGRATED;
int cpu_node = task_node(current);
int local = !!(flags & TNF_FAULT_LOCAL);
sched/numa: Spread memory according to CPU and memory use The pseudo-interleaving in NUMA placement has a fundamental problem: using hard usage thresholds to spread memory equally between nodes can prevent workloads from converging, or keep memory "trapped" on nodes where the workload is barely running any more. In order for workloads to properly converge, the memory migration should not be stopped when nodes reach parity, but instead be distributed according to how heavily memory is used from each node. This way memory migration and task migration reinforce each other, instead of one putting the brakes on the other. Remove the hard thresholds from the pseudo-interleaving code, and instead use a more gradual policy on memory placement. This also seems to improve convergence of workloads that do not run flat out, but sleep in between bursts of activity. We still want to slow down NUMA scanning and migration once a workload has settled on a few actively used nodes, so keep the 3/4 hysteresis in place. Keep track of whether a workload is actively running on multiple nodes, so task_numa_migrate does a full scan of the system for better task placement. In the case of running 3 SPECjbb2005 instances on a 4 node system, this code seems to result in fairer distribution of memory between nodes, with more memory bandwidth for each instance. Signed-off-by: Rik van Riel <riel@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: mgorman@suse.de Link: http://lkml.kernel.org/r/20160125170739.2fc9a641@annuminas.surriel.com [ Minor readability tweaks. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-01-25 22:07:39 +00:00
struct numa_group *ng;
int priv;
if (!static_branch_likely(&sched_numa_balancing))
return;
/* for example, ksmd faulting in a user's mm */
if (!p->mm)
return;
/* Allocate buffer to track faults on a per-node basis */
if (unlikely(!p->numa_faults)) {
int size = sizeof(*p->numa_faults) *
NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
if (!p->numa_faults)
return;
p->total_numa_faults = 0;
memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
}
/*
* First accesses are treated as private, otherwise consider accesses
* to be private if the accessing pid has not changed
*/
if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
priv = 1;
} else {
priv = cpupid_match_pid(p, last_cpupid);
if (!priv && !(flags & TNF_NO_GROUP))
task_numa_group(p, last_cpupid, flags, &priv);
}
/*
* If a workload spans multiple NUMA nodes, a shared fault that
* occurs wholly within the set of nodes that the workload is
* actively using should be counted as local. This allows the
* scan rate to slow down when a workload has settled down.
*/
ng = deref_curr_numa_group(p);
sched/numa: Spread memory according to CPU and memory use The pseudo-interleaving in NUMA placement has a fundamental problem: using hard usage thresholds to spread memory equally between nodes can prevent workloads from converging, or keep memory "trapped" on nodes where the workload is barely running any more. In order for workloads to properly converge, the memory migration should not be stopped when nodes reach parity, but instead be distributed according to how heavily memory is used from each node. This way memory migration and task migration reinforce each other, instead of one putting the brakes on the other. Remove the hard thresholds from the pseudo-interleaving code, and instead use a more gradual policy on memory placement. This also seems to improve convergence of workloads that do not run flat out, but sleep in between bursts of activity. We still want to slow down NUMA scanning and migration once a workload has settled on a few actively used nodes, so keep the 3/4 hysteresis in place. Keep track of whether a workload is actively running on multiple nodes, so task_numa_migrate does a full scan of the system for better task placement. In the case of running 3 SPECjbb2005 instances on a 4 node system, this code seems to result in fairer distribution of memory between nodes, with more memory bandwidth for each instance. Signed-off-by: Rik van Riel <riel@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: mgorman@suse.de Link: http://lkml.kernel.org/r/20160125170739.2fc9a641@annuminas.surriel.com [ Minor readability tweaks. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-01-25 22:07:39 +00:00
if (!priv && !local && ng && ng->active_nodes > 1 &&
numa_is_active_node(cpu_node, ng) &&
numa_is_active_node(mem_node, ng))
local = 1;
/*
* Retry to migrate task to preferred node periodically, in case it
* previously failed, or the scheduler moved us.
*/
sched/numa: Move task_numa_placement() closer to numa_migrate_preferred() numa_migrate_preferred() is called periodically or when task preferred node changes. Preferred node evaluations happen once per scan sequence. If the scan completion happens just after the periodic NUMA migration, then we try to migrate to the preferred node and the preferred node might change, needing another node migration. Avoid this by checking for scan sequence completion only when checking for periodic migration. Running SPECjbb2005 on a 4 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 16 25862.6 26158.1 1.14258 1 74357 72725 -2.19482 Running SPECjbb2005 on a 16 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 8 117019 113992 -2.58 1 179095 174947 -2.31 (numbers from v1 based on v4.17-rc5) Testcase Time: Min Max Avg StdDev numa01.sh Real: 449.46 770.77 615.22 101.70 numa01.sh Sys: 132.72 208.17 170.46 24.96 numa01.sh User: 39185.26 60290.89 50066.76 6807.84 numa02.sh Real: 60.85 61.79 61.28 0.37 numa02.sh Sys: 15.34 24.71 21.08 3.61 numa02.sh User: 5204.41 5249.85 5231.21 17.60 numa03.sh Real: 785.50 916.97 840.77 44.98 numa03.sh Sys: 108.08 133.60 119.43 8.82 numa03.sh User: 61422.86 70919.75 64720.87 3310.61 numa04.sh Real: 429.57 587.37 480.80 57.40 numa04.sh Sys: 240.61 321.97 290.84 33.58 numa04.sh User: 34597.65 40498.99 37079.48 2060.72 numa05.sh Real: 392.09 431.25 414.65 13.82 numa05.sh Sys: 229.41 372.48 297.54 53.14 numa05.sh User: 33390.86 34697.49 34222.43 556.42 Testcase Time: Min Max Avg StdDev %Change numa01.sh Real: 424.63 566.18 498.12 59.26 23.50% numa01.sh Sys: 160.19 256.53 208.98 37.02 -18.4% numa01.sh User: 37320.00 46225.58 42001.57 3482.45 19.20% numa02.sh Real: 60.17 62.47 60.91 0.85 0.607% numa02.sh Sys: 15.30 22.82 17.04 2.90 23.70% numa02.sh User: 5202.13 5255.51 5219.08 20.14 0.232% numa03.sh Real: 823.91 844.89 833.86 8.46 0.828% numa03.sh Sys: 130.69 148.29 140.47 6.21 -14.9% numa03.sh User: 62519.15 64262.20 63613.38 620.05 1.740% numa04.sh Real: 515.30 603.74 548.56 30.93 -12.3% numa04.sh Sys: 459.73 525.48 489.18 21.63 -40.5% numa04.sh User: 40561.96 44919.18 42047.87 1526.85 -11.8% numa05.sh Real: 396.58 454.37 421.13 19.71 -1.53% numa05.sh Sys: 208.72 422.02 348.90 73.60 -14.7% numa05.sh User: 33124.08 36109.35 34846.47 1089.74 -1.79% Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mel Gorman <mgorman@techsingularity.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1529514181-9842-20-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-20 17:03:00 +00:00
if (time_after(jiffies, p->numa_migrate_retry)) {
task_numa_placement(p);
numa_migrate_preferred(p);
sched/numa: Move task_numa_placement() closer to numa_migrate_preferred() numa_migrate_preferred() is called periodically or when task preferred node changes. Preferred node evaluations happen once per scan sequence. If the scan completion happens just after the periodic NUMA migration, then we try to migrate to the preferred node and the preferred node might change, needing another node migration. Avoid this by checking for scan sequence completion only when checking for periodic migration. Running SPECjbb2005 on a 4 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 16 25862.6 26158.1 1.14258 1 74357 72725 -2.19482 Running SPECjbb2005 on a 16 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 8 117019 113992 -2.58 1 179095 174947 -2.31 (numbers from v1 based on v4.17-rc5) Testcase Time: Min Max Avg StdDev numa01.sh Real: 449.46 770.77 615.22 101.70 numa01.sh Sys: 132.72 208.17 170.46 24.96 numa01.sh User: 39185.26 60290.89 50066.76 6807.84 numa02.sh Real: 60.85 61.79 61.28 0.37 numa02.sh Sys: 15.34 24.71 21.08 3.61 numa02.sh User: 5204.41 5249.85 5231.21 17.60 numa03.sh Real: 785.50 916.97 840.77 44.98 numa03.sh Sys: 108.08 133.60 119.43 8.82 numa03.sh User: 61422.86 70919.75 64720.87 3310.61 numa04.sh Real: 429.57 587.37 480.80 57.40 numa04.sh Sys: 240.61 321.97 290.84 33.58 numa04.sh User: 34597.65 40498.99 37079.48 2060.72 numa05.sh Real: 392.09 431.25 414.65 13.82 numa05.sh Sys: 229.41 372.48 297.54 53.14 numa05.sh User: 33390.86 34697.49 34222.43 556.42 Testcase Time: Min Max Avg StdDev %Change numa01.sh Real: 424.63 566.18 498.12 59.26 23.50% numa01.sh Sys: 160.19 256.53 208.98 37.02 -18.4% numa01.sh User: 37320.00 46225.58 42001.57 3482.45 19.20% numa02.sh Real: 60.17 62.47 60.91 0.85 0.607% numa02.sh Sys: 15.30 22.82 17.04 2.90 23.70% numa02.sh User: 5202.13 5255.51 5219.08 20.14 0.232% numa03.sh Real: 823.91 844.89 833.86 8.46 0.828% numa03.sh Sys: 130.69 148.29 140.47 6.21 -14.9% numa03.sh User: 62519.15 64262.20 63613.38 620.05 1.740% numa04.sh Real: 515.30 603.74 548.56 30.93 -12.3% numa04.sh Sys: 459.73 525.48 489.18 21.63 -40.5% numa04.sh User: 40561.96 44919.18 42047.87 1526.85 -11.8% numa05.sh Real: 396.58 454.37 421.13 19.71 -1.53% numa05.sh Sys: 208.72 422.02 348.90 73.60 -14.7% numa05.sh User: 33124.08 36109.35 34846.47 1089.74 -1.79% Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mel Gorman <mgorman@techsingularity.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1529514181-9842-20-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-20 17:03:00 +00:00
}
if (migrated)
p->numa_pages_migrated += pages;
mm: numa: slow PTE scan rate if migration failures occur Dave Chinner reported the following on https://lkml.org/lkml/2015/3/1/226 Across the board the 4.0-rc1 numbers are much slower, and the degradation is far worse when using the large memory footprint configs. Perf points straight at the cause - this is from 4.0-rc1 on the "-o bhash=101073" config: - 56.07% 56.07% [kernel] [k] default_send_IPI_mask_sequence_phys - default_send_IPI_mask_sequence_phys - 99.99% physflat_send_IPI_mask - 99.37% native_send_call_func_ipi smp_call_function_many - native_flush_tlb_others - 99.85% flush_tlb_page ptep_clear_flush try_to_unmap_one rmap_walk try_to_unmap migrate_pages migrate_misplaced_page - handle_mm_fault - 99.73% __do_page_fault trace_do_page_fault do_async_page_fault + async_page_fault 0.63% native_send_call_func_single_ipi generic_exec_single smp_call_function_single This is showing excessive migration activity even though excessive migrations are meant to get throttled. Normally, the scan rate is tuned on a per-task basis depending on the locality of faults. However, if migrations fail for any reason then the PTE scanner may scan faster if the faults continue to be remote. This means there is higher system CPU overhead and fault trapping at exactly the time we know that migrations cannot happen. This patch tracks when migration failures occur and slows the PTE scanner. Signed-off-by: Mel Gorman <mgorman@suse.de> Reported-by: Dave Chinner <david@fromorbit.com> Tested-by: Dave Chinner <david@fromorbit.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Aneesh Kumar <aneesh.kumar@linux.vnet.ibm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-03-25 22:55:42 +00:00
if (flags & TNF_MIGRATE_FAIL)
p->numa_faults_locality[2] += pages;
p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
p->numa_faults_locality[local] += pages;
}
mm: sched: numa: Implement constant, per task Working Set Sampling (WSS) rate Previously, to probe the working set of a task, we'd use a very simple and crude method: mark all of its address space PROT_NONE. That method has various (obvious) disadvantages: - it samples the working set at dissimilar rates, giving some tasks a sampling quality advantage over others. - creates performance problems for tasks with very large working sets - over-samples processes with large address spaces but which only very rarely execute Improve that method by keeping a rotating offset into the address space that marks the current position of the scan, and advance it by a constant rate (in a CPU cycles execution proportional manner). If the offset reaches the last mapped address of the mm then it then it starts over at the first address. The per-task nature of the working set sampling functionality in this tree allows such constant rate, per task, execution-weight proportional sampling of the working set, with an adaptive sampling interval/frequency that goes from once per 100ms up to just once per 8 seconds. The current sampling volume is 256 MB per interval. As tasks mature and converge their working set, so does the sampling rate slow down to just a trickle, 256 MB per 8 seconds of CPU time executed. This, beyond being adaptive, also rate-limits rarely executing systems and does not over-sample on overloaded systems. [ In AutoNUMA speak, this patch deals with the effective sampling rate of the 'hinting page fault'. AutoNUMA's scanning is currently rate-limited, but it is also fundamentally single-threaded, executing in the knuma_scand kernel thread, so the limit in AutoNUMA is global and does not scale up with the number of CPUs, nor does it scan tasks in an execution proportional manner. So the idea of rate-limiting the scanning was first implemented in the AutoNUMA tree via a global rate limit. This patch goes beyond that by implementing an execution rate proportional working set sampling rate that is not implemented via a single global scanning daemon. ] [ Dan Carpenter pointed out a possible NULL pointer dereference in the first version of this patch. ] Based-on-idea-by: Andrea Arcangeli <aarcange@redhat.com> Bug-Found-By: Dan Carpenter <dan.carpenter@oracle.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Rik van Riel <riel@redhat.com> [ Wrote changelog and fixed bug. ] Signed-off-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Mel Gorman <mgorman@suse.de> Reviewed-by: Rik van Riel <riel@redhat.com>
2012-10-25 12:16:45 +00:00
static void reset_ptenuma_scan(struct task_struct *p)
{
/*
* We only did a read acquisition of the mmap sem, so
* p->mm->numa_scan_seq is written to without exclusive access
* and the update is not guaranteed to be atomic. That's not
* much of an issue though, since this is just used for
* statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
* expensive, to avoid any form of compiler optimizations:
*/
WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
mm: sched: numa: Implement constant, per task Working Set Sampling (WSS) rate Previously, to probe the working set of a task, we'd use a very simple and crude method: mark all of its address space PROT_NONE. That method has various (obvious) disadvantages: - it samples the working set at dissimilar rates, giving some tasks a sampling quality advantage over others. - creates performance problems for tasks with very large working sets - over-samples processes with large address spaces but which only very rarely execute Improve that method by keeping a rotating offset into the address space that marks the current position of the scan, and advance it by a constant rate (in a CPU cycles execution proportional manner). If the offset reaches the last mapped address of the mm then it then it starts over at the first address. The per-task nature of the working set sampling functionality in this tree allows such constant rate, per task, execution-weight proportional sampling of the working set, with an adaptive sampling interval/frequency that goes from once per 100ms up to just once per 8 seconds. The current sampling volume is 256 MB per interval. As tasks mature and converge their working set, so does the sampling rate slow down to just a trickle, 256 MB per 8 seconds of CPU time executed. This, beyond being adaptive, also rate-limits rarely executing systems and does not over-sample on overloaded systems. [ In AutoNUMA speak, this patch deals with the effective sampling rate of the 'hinting page fault'. AutoNUMA's scanning is currently rate-limited, but it is also fundamentally single-threaded, executing in the knuma_scand kernel thread, so the limit in AutoNUMA is global and does not scale up with the number of CPUs, nor does it scan tasks in an execution proportional manner. So the idea of rate-limiting the scanning was first implemented in the AutoNUMA tree via a global rate limit. This patch goes beyond that by implementing an execution rate proportional working set sampling rate that is not implemented via a single global scanning daemon. ] [ Dan Carpenter pointed out a possible NULL pointer dereference in the first version of this patch. ] Based-on-idea-by: Andrea Arcangeli <aarcange@redhat.com> Bug-Found-By: Dan Carpenter <dan.carpenter@oracle.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Rik van Riel <riel@redhat.com> [ Wrote changelog and fixed bug. ] Signed-off-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Mel Gorman <mgorman@suse.de> Reviewed-by: Rik van Riel <riel@redhat.com>
2012-10-25 12:16:45 +00:00
p->mm->numa_scan_offset = 0;
}
/*
* The expensive part of numa migration is done from task_work context.
* Triggered from task_tick_numa().
*/
static void task_numa_work(struct callback_head *work)
{
unsigned long migrate, next_scan, now = jiffies;
struct task_struct *p = current;
struct mm_struct *mm = p->mm;
sched/numa: Cap PTE scanning overhead to 3% of run time There is a fundamental mismatch between the runtime based NUMA scanning at the task level, and the wall clock time NUMA scanning at the mm level. On a severely overloaded system, with very large processes, this mismatch can cause the system to spend all of its time in change_prot_numa(). This can happen if the task spends at least two ticks in change_prot_numa(), and only gets two ticks of CPU time in the real time between two scan intervals of the mm. This patch ensures that a task never spends more than 3% of run time scanning PTEs. It does that by ensuring that in-between task_numa_work() runs, the task spends at least 32x as much time on other things than it did on task_numa_work(). This is done stochastically: if a timer tick happens, or the task gets rescheduled during task_numa_work(), we delay a future run of task_numa_work() until the task has spent at least 32x the amount of CPU time doing something else, as it spent inside task_numa_work(). The longer task_numa_work() takes, the more likely it is this happens. If task_numa_work() takes very little time, chances are low that that code will do anything, but we will not care. Reported-and-tested-by: Jan Stancek <jstancek@redhat.com> Signed-off-by: Rik van Riel <riel@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: mgorman@suse.de Link: http://lkml.kernel.org/r/1446756983-28173-3-git-send-email-riel@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-11-05 20:56:23 +00:00
u64 runtime = p->se.sum_exec_runtime;
mm: sched: numa: Implement constant, per task Working Set Sampling (WSS) rate Previously, to probe the working set of a task, we'd use a very simple and crude method: mark all of its address space PROT_NONE. That method has various (obvious) disadvantages: - it samples the working set at dissimilar rates, giving some tasks a sampling quality advantage over others. - creates performance problems for tasks with very large working sets - over-samples processes with large address spaces but which only very rarely execute Improve that method by keeping a rotating offset into the address space that marks the current position of the scan, and advance it by a constant rate (in a CPU cycles execution proportional manner). If the offset reaches the last mapped address of the mm then it then it starts over at the first address. The per-task nature of the working set sampling functionality in this tree allows such constant rate, per task, execution-weight proportional sampling of the working set, with an adaptive sampling interval/frequency that goes from once per 100ms up to just once per 8 seconds. The current sampling volume is 256 MB per interval. As tasks mature and converge their working set, so does the sampling rate slow down to just a trickle, 256 MB per 8 seconds of CPU time executed. This, beyond being adaptive, also rate-limits rarely executing systems and does not over-sample on overloaded systems. [ In AutoNUMA speak, this patch deals with the effective sampling rate of the 'hinting page fault'. AutoNUMA's scanning is currently rate-limited, but it is also fundamentally single-threaded, executing in the knuma_scand kernel thread, so the limit in AutoNUMA is global and does not scale up with the number of CPUs, nor does it scan tasks in an execution proportional manner. So the idea of rate-limiting the scanning was first implemented in the AutoNUMA tree via a global rate limit. This patch goes beyond that by implementing an execution rate proportional working set sampling rate that is not implemented via a single global scanning daemon. ] [ Dan Carpenter pointed out a possible NULL pointer dereference in the first version of this patch. ] Based-on-idea-by: Andrea Arcangeli <aarcange@redhat.com> Bug-Found-By: Dan Carpenter <dan.carpenter@oracle.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Rik van Riel <riel@redhat.com> [ Wrote changelog and fixed bug. ] Signed-off-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Mel Gorman <mgorman@suse.de> Reviewed-by: Rik van Riel <riel@redhat.com>
2012-10-25 12:16:45 +00:00
struct vm_area_struct *vma;
unsigned long start, end;
sched/numa: Set the scan rate proportional to the memory usage of the task being scanned The NUMA PTE scan rate is controlled with a combination of the numa_balancing_scan_period_min, numa_balancing_scan_period_max and numa_balancing_scan_size. This scan rate is independent of the size of the task and as an aside it is further complicated by the fact that numa_balancing_scan_size controls how many pages are marked pte_numa and not how much virtual memory is scanned. In combination, it is almost impossible to meaningfully tune the min and max scan periods and reasoning about performance is complex when the time to complete a full scan is is partially a function of the tasks memory size. This patch alters the semantic of the min and max tunables to be about tuning the length time it takes to complete a scan of a tasks occupied virtual address space. Conceptually this is a lot easier to understand. There is a "sanity" check to ensure the scan rate is never extremely fast based on the amount of virtual memory that should be scanned in a second. The default of 2.5G seems arbitrary but it is to have the maximum scan rate after the patch roughly match the maximum scan rate before the patch was applied. On a similar note, numa_scan_period is in milliseconds and not jiffies. Properly placed pages slow the scanning rate but adding 10 jiffies to numa_scan_period means that the rate scanning slows depends on HZ which is confusing. Get rid of the jiffies_to_msec conversion and treat it as ms. Signed-off-by: Mel Gorman <mgorman@suse.de> Reviewed-by: Rik van Riel <riel@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1381141781-10992-18-git-send-email-mgorman@suse.de Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-10-07 10:28:55 +00:00
unsigned long nr_pte_updates = 0;
sched/numa: Limit the amount of virtual memory scanned in task_numa_work() Currently task_numa_work() scans up to numa_balancing_scan_size_mb worth of memory per invocation, but only counts memory areas that have at least one PTE that is still present and not marked for numa hint faulting. It will skip over arbitarily large amounts of memory that are either unused, full of swap ptes, or full of PTEs that were already marked for NUMA hint faults but have not been faulted on yet. This can cause excessive amounts of CPU use, due to there being essentially no upper limit on the scan rate of very large processes that are not yet in a phase where they are actively accessing old memory pages (eg. they are still initializing their data). Avoid that problem by placing an upper limit on the amount of virtual memory that task_numa_work() scans in each invocation. This can be a higher limit than "pages", to ensure the task still skips over unused areas fairly quickly. While we are here, also fix the "nr_pte_updates" logic, so it only counts page ranges with ptes in them. Reported-by: Andrea Arcangeli <aarcange@redhat.com> Reported-by: Jan Stancek <jstancek@redhat.com> Signed-off-by: Rik van Riel <riel@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/20150911090027.4a7987bd@annuminas.surriel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-09-11 13:00:27 +00:00
long pages, virtpages;
SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
work->next = work;
/*
* Who cares about NUMA placement when they're dying.
*
* NOTE: make sure not to dereference p->mm before this check,
* exit_task_work() happens _after_ exit_mm() so we could be called
* without p->mm even though we still had it when we enqueued this
* work.
*/
if (p->flags & PF_EXITING)
return;
if (!mm->numa_next_scan) {
mm->numa_next_scan = now +
msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
}
/*
* Enforce maximal scan/migration frequency..
*/
migrate = mm->numa_next_scan;
if (time_before(now, migrate))
return;
sched/numa: Set the scan rate proportional to the memory usage of the task being scanned The NUMA PTE scan rate is controlled with a combination of the numa_balancing_scan_period_min, numa_balancing_scan_period_max and numa_balancing_scan_size. This scan rate is independent of the size of the task and as an aside it is further complicated by the fact that numa_balancing_scan_size controls how many pages are marked pte_numa and not how much virtual memory is scanned. In combination, it is almost impossible to meaningfully tune the min and max scan periods and reasoning about performance is complex when the time to complete a full scan is is partially a function of the tasks memory size. This patch alters the semantic of the min and max tunables to be about tuning the length time it takes to complete a scan of a tasks occupied virtual address space. Conceptually this is a lot easier to understand. There is a "sanity" check to ensure the scan rate is never extremely fast based on the amount of virtual memory that should be scanned in a second. The default of 2.5G seems arbitrary but it is to have the maximum scan rate after the patch roughly match the maximum scan rate before the patch was applied. On a similar note, numa_scan_period is in milliseconds and not jiffies. Properly placed pages slow the scanning rate but adding 10 jiffies to numa_scan_period means that the rate scanning slows depends on HZ which is confusing. Get rid of the jiffies_to_msec conversion and treat it as ms. Signed-off-by: Mel Gorman <mgorman@suse.de> Reviewed-by: Rik van Riel <riel@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1381141781-10992-18-git-send-email-mgorman@suse.de Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-10-07 10:28:55 +00:00
if (p->numa_scan_period == 0) {
p->numa_scan_period_max = task_scan_max(p);
p->numa_scan_period = task_scan_start(p);
sched/numa: Set the scan rate proportional to the memory usage of the task being scanned The NUMA PTE scan rate is controlled with a combination of the numa_balancing_scan_period_min, numa_balancing_scan_period_max and numa_balancing_scan_size. This scan rate is independent of the size of the task and as an aside it is further complicated by the fact that numa_balancing_scan_size controls how many pages are marked pte_numa and not how much virtual memory is scanned. In combination, it is almost impossible to meaningfully tune the min and max scan periods and reasoning about performance is complex when the time to complete a full scan is is partially a function of the tasks memory size. This patch alters the semantic of the min and max tunables to be about tuning the length time it takes to complete a scan of a tasks occupied virtual address space. Conceptually this is a lot easier to understand. There is a "sanity" check to ensure the scan rate is never extremely fast based on the amount of virtual memory that should be scanned in a second. The default of 2.5G seems arbitrary but it is to have the maximum scan rate after the patch roughly match the maximum scan rate before the patch was applied. On a similar note, numa_scan_period is in milliseconds and not jiffies. Properly placed pages slow the scanning rate but adding 10 jiffies to numa_scan_period means that the rate scanning slows depends on HZ which is confusing. Get rid of the jiffies_to_msec conversion and treat it as ms. Signed-off-by: Mel Gorman <mgorman@suse.de> Reviewed-by: Rik van Riel <riel@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1381141781-10992-18-git-send-email-mgorman@suse.de Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-10-07 10:28:55 +00:00
}
next_scan = now + msecs_to_jiffies(p->numa_scan_period);
if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
return;
sched/numa: Mitigate chance that same task always updates PTEs With a trace_printk("working\n"); right after the cmpxchg in task_numa_work() we can see that of a 4 thread process, its always the same task winning the race and doing the protection change. This is a problem since the task doing the protection change has a penalty for taking faults -- it is busy when marking the PTEs. If its always the same task the ->numa_faults[] get severely skewed. Avoid this by delaying the task doing the protection change such that it is unlikely to win the privilege again. Before: root@interlagos:~# grep "thread 0/.*working" /debug/tracing/trace | tail -15 thread 0/0-3232 [022] .... 212.787402: task_numa_work: working thread 0/0-3232 [022] .... 212.888473: task_numa_work: working thread 0/0-3232 [022] .... 212.989538: task_numa_work: working thread 0/0-3232 [022] .... 213.090602: task_numa_work: working thread 0/0-3232 [022] .... 213.191667: task_numa_work: working thread 0/0-3232 [022] .... 213.292734: task_numa_work: working thread 0/0-3232 [022] .... 213.393804: task_numa_work: working thread 0/0-3232 [022] .... 213.494869: task_numa_work: working thread 0/0-3232 [022] .... 213.596937: task_numa_work: working thread 0/0-3232 [022] .... 213.699000: task_numa_work: working thread 0/0-3232 [022] .... 213.801067: task_numa_work: working thread 0/0-3232 [022] .... 213.903155: task_numa_work: working thread 0/0-3232 [022] .... 214.005201: task_numa_work: working thread 0/0-3232 [022] .... 214.107266: task_numa_work: working thread 0/0-3232 [022] .... 214.209342: task_numa_work: working After: root@interlagos:~# grep "thread 0/.*working" /debug/tracing/trace | tail -15 thread 0/0-3253 [005] .... 136.865051: task_numa_work: working thread 0/2-3255 [026] .... 136.965134: task_numa_work: working thread 0/3-3256 [024] .... 137.065217: task_numa_work: working thread 0/3-3256 [024] .... 137.165302: task_numa_work: working thread 0/3-3256 [024] .... 137.265382: task_numa_work: working thread 0/0-3253 [004] .... 137.366465: task_numa_work: working thread 0/2-3255 [026] .... 137.466549: task_numa_work: working thread 0/0-3253 [004] .... 137.566629: task_numa_work: working thread 0/0-3253 [004] .... 137.666711: task_numa_work: working thread 0/1-3254 [028] .... 137.766799: task_numa_work: working thread 0/0-3253 [004] .... 137.866876: task_numa_work: working thread 0/2-3255 [026] .... 137.966960: task_numa_work: working thread 0/1-3254 [028] .... 138.067041: task_numa_work: working thread 0/2-3255 [026] .... 138.167123: task_numa_work: working thread 0/3-3256 [024] .... 138.267207: task_numa_work: working Signed-off-by: Peter Zijlstra <peterz@infradead.org> Signed-off-by: Mel Gorman <mgorman@suse.de> Reviewed-by: Rik van Riel <riel@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Link: http://lkml.kernel.org/r/1381141781-10992-14-git-send-email-mgorman@suse.de Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-10-07 10:28:51 +00:00
/*
* Delay this task enough that another task of this mm will likely win
* the next time around.
*/
p->node_stamp += 2 * TICK_NSEC;
start = mm->numa_scan_offset;
pages = sysctl_numa_balancing_scan_size;
pages <<= 20 - PAGE_SHIFT; /* MB in pages */
sched/numa: Limit the amount of virtual memory scanned in task_numa_work() Currently task_numa_work() scans up to numa_balancing_scan_size_mb worth of memory per invocation, but only counts memory areas that have at least one PTE that is still present and not marked for numa hint faulting. It will skip over arbitarily large amounts of memory that are either unused, full of swap ptes, or full of PTEs that were already marked for NUMA hint faults but have not been faulted on yet. This can cause excessive amounts of CPU use, due to there being essentially no upper limit on the scan rate of very large processes that are not yet in a phase where they are actively accessing old memory pages (eg. they are still initializing their data). Avoid that problem by placing an upper limit on the amount of virtual memory that task_numa_work() scans in each invocation. This can be a higher limit than "pages", to ensure the task still skips over unused areas fairly quickly. While we are here, also fix the "nr_pte_updates" logic, so it only counts page ranges with ptes in them. Reported-by: Andrea Arcangeli <aarcange@redhat.com> Reported-by: Jan Stancek <jstancek@redhat.com> Signed-off-by: Rik van Riel <riel@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/20150911090027.4a7987bd@annuminas.surriel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-09-11 13:00:27 +00:00
virtpages = pages * 8; /* Scan up to this much virtual space */
if (!pages)
return;
sched/numa: Limit the amount of virtual memory scanned in task_numa_work() Currently task_numa_work() scans up to numa_balancing_scan_size_mb worth of memory per invocation, but only counts memory areas that have at least one PTE that is still present and not marked for numa hint faulting. It will skip over arbitarily large amounts of memory that are either unused, full of swap ptes, or full of PTEs that were already marked for NUMA hint faults but have not been faulted on yet. This can cause excessive amounts of CPU use, due to there being essentially no upper limit on the scan rate of very large processes that are not yet in a phase where they are actively accessing old memory pages (eg. they are still initializing their data). Avoid that problem by placing an upper limit on the amount of virtual memory that task_numa_work() scans in each invocation. This can be a higher limit than "pages", to ensure the task still skips over unused areas fairly quickly. While we are here, also fix the "nr_pte_updates" logic, so it only counts page ranges with ptes in them. Reported-by: Andrea Arcangeli <aarcange@redhat.com> Reported-by: Jan Stancek <jstancek@redhat.com> Signed-off-by: Rik van Riel <riel@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/20150911090027.4a7987bd@annuminas.surriel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-09-11 13:00:27 +00:00
sched/numa: Use down_read_trylock() for the mmap_sem A customer has reported a soft-lockup when running an intensive memory stress test, where the trace on multiple CPU's looks like this: RIP: 0010:[<ffffffff810c53fe>] [<ffffffff810c53fe>] native_queued_spin_lock_slowpath+0x10e/0x190 ... Call Trace: [<ffffffff81182d07>] queued_spin_lock_slowpath+0x7/0xa [<ffffffff811bc331>] change_protection_range+0x3b1/0x930 [<ffffffff811d4be8>] change_prot_numa+0x18/0x30 [<ffffffff810adefe>] task_numa_work+0x1fe/0x310 [<ffffffff81098322>] task_work_run+0x72/0x90 Further investigation showed that the lock contention here is pmd_lock(). The task_numa_work() function makes sure that only one thread is let to perform the work in a single scan period (via cmpxchg), but if there's a thread with mmap_sem locked for writing for several periods, multiple threads in task_numa_work() can build up a convoy waiting for mmap_sem for read and then all get unblocked at once. This patch changes the down_read() to the trylock version, which prevents the build up. For a workload experiencing mmap_sem contention, it's probably better to postpone the NUMA balancing work anyway. This seems to have fixed the soft lockups involving pmd_lock(), which is in line with the convoy theory. Signed-off-by: Vlastimil Babka <vbabka@suse.cz> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Rik van Riel <riel@redhat.com> Acked-by: Mel Gorman <mgorman@techsingularity.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/20170515131316.21909-1-vbabka@suse.cz Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-05-15 13:13:16 +00:00
if (!down_read_trylock(&mm->mmap_sem))
return;
vma = find_vma(mm, start);
mm: sched: numa: Implement constant, per task Working Set Sampling (WSS) rate Previously, to probe the working set of a task, we'd use a very simple and crude method: mark all of its address space PROT_NONE. That method has various (obvious) disadvantages: - it samples the working set at dissimilar rates, giving some tasks a sampling quality advantage over others. - creates performance problems for tasks with very large working sets - over-samples processes with large address spaces but which only very rarely execute Improve that method by keeping a rotating offset into the address space that marks the current position of the scan, and advance it by a constant rate (in a CPU cycles execution proportional manner). If the offset reaches the last mapped address of the mm then it then it starts over at the first address. The per-task nature of the working set sampling functionality in this tree allows such constant rate, per task, execution-weight proportional sampling of the working set, with an adaptive sampling interval/frequency that goes from once per 100ms up to just once per 8 seconds. The current sampling volume is 256 MB per interval. As tasks mature and converge their working set, so does the sampling rate slow down to just a trickle, 256 MB per 8 seconds of CPU time executed. This, beyond being adaptive, also rate-limits rarely executing systems and does not over-sample on overloaded systems. [ In AutoNUMA speak, this patch deals with the effective sampling rate of the 'hinting page fault'. AutoNUMA's scanning is currently rate-limited, but it is also fundamentally single-threaded, executing in the knuma_scand kernel thread, so the limit in AutoNUMA is global and does not scale up with the number of CPUs, nor does it scan tasks in an execution proportional manner. So the idea of rate-limiting the scanning was first implemented in the AutoNUMA tree via a global rate limit. This patch goes beyond that by implementing an execution rate proportional working set sampling rate that is not implemented via a single global scanning daemon. ] [ Dan Carpenter pointed out a possible NULL pointer dereference in the first version of this patch. ] Based-on-idea-by: Andrea Arcangeli <aarcange@redhat.com> Bug-Found-By: Dan Carpenter <dan.carpenter@oracle.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Rik van Riel <riel@redhat.com> [ Wrote changelog and fixed bug. ] Signed-off-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Mel Gorman <mgorman@suse.de> Reviewed-by: Rik van Riel <riel@redhat.com>
2012-10-25 12:16:45 +00:00
if (!vma) {
reset_ptenuma_scan(p);
start = 0;
mm: sched: numa: Implement constant, per task Working Set Sampling (WSS) rate Previously, to probe the working set of a task, we'd use a very simple and crude method: mark all of its address space PROT_NONE. That method has various (obvious) disadvantages: - it samples the working set at dissimilar rates, giving some tasks a sampling quality advantage over others. - creates performance problems for tasks with very large working sets - over-samples processes with large address spaces but which only very rarely execute Improve that method by keeping a rotating offset into the address space that marks the current position of the scan, and advance it by a constant rate (in a CPU cycles execution proportional manner). If the offset reaches the last mapped address of the mm then it then it starts over at the first address. The per-task nature of the working set sampling functionality in this tree allows such constant rate, per task, execution-weight proportional sampling of the working set, with an adaptive sampling interval/frequency that goes from once per 100ms up to just once per 8 seconds. The current sampling volume is 256 MB per interval. As tasks mature and converge their working set, so does the sampling rate slow down to just a trickle, 256 MB per 8 seconds of CPU time executed. This, beyond being adaptive, also rate-limits rarely executing systems and does not over-sample on overloaded systems. [ In AutoNUMA speak, this patch deals with the effective sampling rate of the 'hinting page fault'. AutoNUMA's scanning is currently rate-limited, but it is also fundamentally single-threaded, executing in the knuma_scand kernel thread, so the limit in AutoNUMA is global and does not scale up with the number of CPUs, nor does it scan tasks in an execution proportional manner. So the idea of rate-limiting the scanning was first implemented in the AutoNUMA tree via a global rate limit. This patch goes beyond that by implementing an execution rate proportional working set sampling rate that is not implemented via a single global scanning daemon. ] [ Dan Carpenter pointed out a possible NULL pointer dereference in the first version of this patch. ] Based-on-idea-by: Andrea Arcangeli <aarcange@redhat.com> Bug-Found-By: Dan Carpenter <dan.carpenter@oracle.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Rik van Riel <riel@redhat.com> [ Wrote changelog and fixed bug. ] Signed-off-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Mel Gorman <mgorman@suse.de> Reviewed-by: Rik van Riel <riel@redhat.com>
2012-10-25 12:16:45 +00:00
vma = mm->mmap;
}
for (; vma; vma = vma->vm_next) {
if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
sched, numa: do not hint for NUMA balancing on VM_MIXEDMAP mappings Jovi Zhangwei reported the following problem Below kernel vm bug can be triggered by tcpdump which mmaped a lot of pages with GFP_COMP flag. [Mon May 25 05:29:33 2015] page:ffffea0015414000 count:66 mapcount:1 mapping: (null) index:0x0 [Mon May 25 05:29:33 2015] flags: 0x20047580004000(head) [Mon May 25 05:29:33 2015] page dumped because: VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page)) [Mon May 25 05:29:33 2015] ------------[ cut here ]------------ [Mon May 25 05:29:33 2015] kernel BUG at mm/migrate.c:1661! [Mon May 25 05:29:33 2015] invalid opcode: 0000 [#1] SMP In this case it was triggered by running tcpdump but it's not necessary reproducible on all systems. sudo tcpdump -i bond0.100 'tcp port 4242' -c 100000000000 -w 4242.pcap Compound pages cannot be migrated and it was not expected that such pages be marked for NUMA balancing. This did not take into account that drivers such as net/packet/af_packet.c may insert compound pages into userspace with vm_insert_page. This patch tells the NUMA balancing protection scanner to skip all VM_MIXEDMAP mappings which avoids the possibility that compound pages are marked for migration. Signed-off-by: Mel Gorman <mgorman@suse.de> Reported-by: Jovi Zhangwei <jovi@cloudflare.com> Cc: Ingo Molnar <mingo@elte.hu> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: <stable@vger.kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-06-10 18:15:00 +00:00
is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
mm: sched: numa: Implement constant, per task Working Set Sampling (WSS) rate Previously, to probe the working set of a task, we'd use a very simple and crude method: mark all of its address space PROT_NONE. That method has various (obvious) disadvantages: - it samples the working set at dissimilar rates, giving some tasks a sampling quality advantage over others. - creates performance problems for tasks with very large working sets - over-samples processes with large address spaces but which only very rarely execute Improve that method by keeping a rotating offset into the address space that marks the current position of the scan, and advance it by a constant rate (in a CPU cycles execution proportional manner). If the offset reaches the last mapped address of the mm then it then it starts over at the first address. The per-task nature of the working set sampling functionality in this tree allows such constant rate, per task, execution-weight proportional sampling of the working set, with an adaptive sampling interval/frequency that goes from once per 100ms up to just once per 8 seconds. The current sampling volume is 256 MB per interval. As tasks mature and converge their working set, so does the sampling rate slow down to just a trickle, 256 MB per 8 seconds of CPU time executed. This, beyond being adaptive, also rate-limits rarely executing systems and does not over-sample on overloaded systems. [ In AutoNUMA speak, this patch deals with the effective sampling rate of the 'hinting page fault'. AutoNUMA's scanning is currently rate-limited, but it is also fundamentally single-threaded, executing in the knuma_scand kernel thread, so the limit in AutoNUMA is global and does not scale up with the number of CPUs, nor does it scan tasks in an execution proportional manner. So the idea of rate-limiting the scanning was first implemented in the AutoNUMA tree via a global rate limit. This patch goes beyond that by implementing an execution rate proportional working set sampling rate that is not implemented via a single global scanning daemon. ] [ Dan Carpenter pointed out a possible NULL pointer dereference in the first version of this patch. ] Based-on-idea-by: Andrea Arcangeli <aarcange@redhat.com> Bug-Found-By: Dan Carpenter <dan.carpenter@oracle.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Rik van Riel <riel@redhat.com> [ Wrote changelog and fixed bug. ] Signed-off-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Mel Gorman <mgorman@suse.de> Reviewed-by: Rik van Riel <riel@redhat.com>
2012-10-25 12:16:45 +00:00
continue;
}
mm: sched: numa: Implement constant, per task Working Set Sampling (WSS) rate Previously, to probe the working set of a task, we'd use a very simple and crude method: mark all of its address space PROT_NONE. That method has various (obvious) disadvantages: - it samples the working set at dissimilar rates, giving some tasks a sampling quality advantage over others. - creates performance problems for tasks with very large working sets - over-samples processes with large address spaces but which only very rarely execute Improve that method by keeping a rotating offset into the address space that marks the current position of the scan, and advance it by a constant rate (in a CPU cycles execution proportional manner). If the offset reaches the last mapped address of the mm then it then it starts over at the first address. The per-task nature of the working set sampling functionality in this tree allows such constant rate, per task, execution-weight proportional sampling of the working set, with an adaptive sampling interval/frequency that goes from once per 100ms up to just once per 8 seconds. The current sampling volume is 256 MB per interval. As tasks mature and converge their working set, so does the sampling rate slow down to just a trickle, 256 MB per 8 seconds of CPU time executed. This, beyond being adaptive, also rate-limits rarely executing systems and does not over-sample on overloaded systems. [ In AutoNUMA speak, this patch deals with the effective sampling rate of the 'hinting page fault'. AutoNUMA's scanning is currently rate-limited, but it is also fundamentally single-threaded, executing in the knuma_scand kernel thread, so the limit in AutoNUMA is global and does not scale up with the number of CPUs, nor does it scan tasks in an execution proportional manner. So the idea of rate-limiting the scanning was first implemented in the AutoNUMA tree via a global rate limit. This patch goes beyond that by implementing an execution rate proportional working set sampling rate that is not implemented via a single global scanning daemon. ] [ Dan Carpenter pointed out a possible NULL pointer dereference in the first version of this patch. ] Based-on-idea-by: Andrea Arcangeli <aarcange@redhat.com> Bug-Found-By: Dan Carpenter <dan.carpenter@oracle.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Rik van Riel <riel@redhat.com> [ Wrote changelog and fixed bug. ] Signed-off-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Mel Gorman <mgorman@suse.de> Reviewed-by: Rik van Riel <riel@redhat.com>
2012-10-25 12:16:45 +00:00
/*
* Shared library pages mapped by multiple processes are not
* migrated as it is expected they are cache replicated. Avoid
* hinting faults in read-only file-backed mappings or the vdso
* as migrating the pages will be of marginal benefit.
*/
if (!vma->vm_mm ||
(vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
continue;
/*
* Skip inaccessible VMAs to avoid any confusion between
* PROT_NONE and NUMA hinting ptes
*/
mm/vma: make vma_is_accessible() available for general use Lets move vma_is_accessible() helper to include/linux/mm.h which makes it available for general use. While here, this replaces all remaining open encodings for VMA access check with vma_is_accessible(). Signed-off-by: Anshuman Khandual <anshuman.khandual@arm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Acked-by: Geert Uytterhoeven <geert@linux-m68k.org> Acked-by: Guo Ren <guoren@kernel.org> Acked-by: Vlastimil Babka <vbabka@suse.cz> Cc: Guo Ren <guoren@kernel.org> Cc: Geert Uytterhoeven <geert@linux-m68k.org> Cc: Ralf Baechle <ralf@linux-mips.org> Cc: Paul Burton <paulburton@kernel.org> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: Michael Ellerman <mpe@ellerman.id.au> Cc: Yoshinori Sato <ysato@users.sourceforge.jp> Cc: Rich Felker <dalias@libc.org> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Andy Lutomirski <luto@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Ingo Molnar <mingo@redhat.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: "Aneesh Kumar K.V" <aneesh.kumar@linux.ibm.com> Cc: Arnaldo Carvalho de Melo <acme@kernel.org> Cc: Arnd Bergmann <arnd@arndb.de> Cc: Nick Piggin <npiggin@gmail.com> Cc: Paul Mackerras <paulus@ozlabs.org> Cc: Will Deacon <will@kernel.org> Link: http://lkml.kernel.org/r/1582520593-30704-3-git-send-email-anshuman.khandual@arm.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-07 03:03:47 +00:00
if (!vma_is_accessible(vma))
continue;
do {
start = max(start, vma->vm_start);
end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
end = min(end, vma->vm_end);
sched/numa: Limit the amount of virtual memory scanned in task_numa_work() Currently task_numa_work() scans up to numa_balancing_scan_size_mb worth of memory per invocation, but only counts memory areas that have at least one PTE that is still present and not marked for numa hint faulting. It will skip over arbitarily large amounts of memory that are either unused, full of swap ptes, or full of PTEs that were already marked for NUMA hint faults but have not been faulted on yet. This can cause excessive amounts of CPU use, due to there being essentially no upper limit on the scan rate of very large processes that are not yet in a phase where they are actively accessing old memory pages (eg. they are still initializing their data). Avoid that problem by placing an upper limit on the amount of virtual memory that task_numa_work() scans in each invocation. This can be a higher limit than "pages", to ensure the task still skips over unused areas fairly quickly. While we are here, also fix the "nr_pte_updates" logic, so it only counts page ranges with ptes in them. Reported-by: Andrea Arcangeli <aarcange@redhat.com> Reported-by: Jan Stancek <jstancek@redhat.com> Signed-off-by: Rik van Riel <riel@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/20150911090027.4a7987bd@annuminas.surriel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-09-11 13:00:27 +00:00
nr_pte_updates = change_prot_numa(vma, start, end);
sched/numa: Set the scan rate proportional to the memory usage of the task being scanned The NUMA PTE scan rate is controlled with a combination of the numa_balancing_scan_period_min, numa_balancing_scan_period_max and numa_balancing_scan_size. This scan rate is independent of the size of the task and as an aside it is further complicated by the fact that numa_balancing_scan_size controls how many pages are marked pte_numa and not how much virtual memory is scanned. In combination, it is almost impossible to meaningfully tune the min and max scan periods and reasoning about performance is complex when the time to complete a full scan is is partially a function of the tasks memory size. This patch alters the semantic of the min and max tunables to be about tuning the length time it takes to complete a scan of a tasks occupied virtual address space. Conceptually this is a lot easier to understand. There is a "sanity" check to ensure the scan rate is never extremely fast based on the amount of virtual memory that should be scanned in a second. The default of 2.5G seems arbitrary but it is to have the maximum scan rate after the patch roughly match the maximum scan rate before the patch was applied. On a similar note, numa_scan_period is in milliseconds and not jiffies. Properly placed pages slow the scanning rate but adding 10 jiffies to numa_scan_period means that the rate scanning slows depends on HZ which is confusing. Get rid of the jiffies_to_msec conversion and treat it as ms. Signed-off-by: Mel Gorman <mgorman@suse.de> Reviewed-by: Rik van Riel <riel@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1381141781-10992-18-git-send-email-mgorman@suse.de Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-10-07 10:28:55 +00:00
/*
sched/numa: Limit the amount of virtual memory scanned in task_numa_work() Currently task_numa_work() scans up to numa_balancing_scan_size_mb worth of memory per invocation, but only counts memory areas that have at least one PTE that is still present and not marked for numa hint faulting. It will skip over arbitarily large amounts of memory that are either unused, full of swap ptes, or full of PTEs that were already marked for NUMA hint faults but have not been faulted on yet. This can cause excessive amounts of CPU use, due to there being essentially no upper limit on the scan rate of very large processes that are not yet in a phase where they are actively accessing old memory pages (eg. they are still initializing their data). Avoid that problem by placing an upper limit on the amount of virtual memory that task_numa_work() scans in each invocation. This can be a higher limit than "pages", to ensure the task still skips over unused areas fairly quickly. While we are here, also fix the "nr_pte_updates" logic, so it only counts page ranges with ptes in them. Reported-by: Andrea Arcangeli <aarcange@redhat.com> Reported-by: Jan Stancek <jstancek@redhat.com> Signed-off-by: Rik van Riel <riel@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/20150911090027.4a7987bd@annuminas.surriel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-09-11 13:00:27 +00:00
* Try to scan sysctl_numa_balancing_size worth of
* hpages that have at least one present PTE that
* is not already pte-numa. If the VMA contains
* areas that are unused or already full of prot_numa
* PTEs, scan up to virtpages, to skip through those
* areas faster.
sched/numa: Set the scan rate proportional to the memory usage of the task being scanned The NUMA PTE scan rate is controlled with a combination of the numa_balancing_scan_period_min, numa_balancing_scan_period_max and numa_balancing_scan_size. This scan rate is independent of the size of the task and as an aside it is further complicated by the fact that numa_balancing_scan_size controls how many pages are marked pte_numa and not how much virtual memory is scanned. In combination, it is almost impossible to meaningfully tune the min and max scan periods and reasoning about performance is complex when the time to complete a full scan is is partially a function of the tasks memory size. This patch alters the semantic of the min and max tunables to be about tuning the length time it takes to complete a scan of a tasks occupied virtual address space. Conceptually this is a lot easier to understand. There is a "sanity" check to ensure the scan rate is never extremely fast based on the amount of virtual memory that should be scanned in a second. The default of 2.5G seems arbitrary but it is to have the maximum scan rate after the patch roughly match the maximum scan rate before the patch was applied. On a similar note, numa_scan_period is in milliseconds and not jiffies. Properly placed pages slow the scanning rate but adding 10 jiffies to numa_scan_period means that the rate scanning slows depends on HZ which is confusing. Get rid of the jiffies_to_msec conversion and treat it as ms. Signed-off-by: Mel Gorman <mgorman@suse.de> Reviewed-by: Rik van Riel <riel@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1381141781-10992-18-git-send-email-mgorman@suse.de Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-10-07 10:28:55 +00:00
*/
if (nr_pte_updates)
pages -= (end - start) >> PAGE_SHIFT;
sched/numa: Limit the amount of virtual memory scanned in task_numa_work() Currently task_numa_work() scans up to numa_balancing_scan_size_mb worth of memory per invocation, but only counts memory areas that have at least one PTE that is still present and not marked for numa hint faulting. It will skip over arbitarily large amounts of memory that are either unused, full of swap ptes, or full of PTEs that were already marked for NUMA hint faults but have not been faulted on yet. This can cause excessive amounts of CPU use, due to there being essentially no upper limit on the scan rate of very large processes that are not yet in a phase where they are actively accessing old memory pages (eg. they are still initializing their data). Avoid that problem by placing an upper limit on the amount of virtual memory that task_numa_work() scans in each invocation. This can be a higher limit than "pages", to ensure the task still skips over unused areas fairly quickly. While we are here, also fix the "nr_pte_updates" logic, so it only counts page ranges with ptes in them. Reported-by: Andrea Arcangeli <aarcange@redhat.com> Reported-by: Jan Stancek <jstancek@redhat.com> Signed-off-by: Rik van Riel <riel@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/20150911090027.4a7987bd@annuminas.surriel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-09-11 13:00:27 +00:00
virtpages -= (end - start) >> PAGE_SHIFT;
mm: sched: numa: Implement constant, per task Working Set Sampling (WSS) rate Previously, to probe the working set of a task, we'd use a very simple and crude method: mark all of its address space PROT_NONE. That method has various (obvious) disadvantages: - it samples the working set at dissimilar rates, giving some tasks a sampling quality advantage over others. - creates performance problems for tasks with very large working sets - over-samples processes with large address spaces but which only very rarely execute Improve that method by keeping a rotating offset into the address space that marks the current position of the scan, and advance it by a constant rate (in a CPU cycles execution proportional manner). If the offset reaches the last mapped address of the mm then it then it starts over at the first address. The per-task nature of the working set sampling functionality in this tree allows such constant rate, per task, execution-weight proportional sampling of the working set, with an adaptive sampling interval/frequency that goes from once per 100ms up to just once per 8 seconds. The current sampling volume is 256 MB per interval. As tasks mature and converge their working set, so does the sampling rate slow down to just a trickle, 256 MB per 8 seconds of CPU time executed. This, beyond being adaptive, also rate-limits rarely executing systems and does not over-sample on overloaded systems. [ In AutoNUMA speak, this patch deals with the effective sampling rate of the 'hinting page fault'. AutoNUMA's scanning is currently rate-limited, but it is also fundamentally single-threaded, executing in the knuma_scand kernel thread, so the limit in AutoNUMA is global and does not scale up with the number of CPUs, nor does it scan tasks in an execution proportional manner. So the idea of rate-limiting the scanning was first implemented in the AutoNUMA tree via a global rate limit. This patch goes beyond that by implementing an execution rate proportional working set sampling rate that is not implemented via a single global scanning daemon. ] [ Dan Carpenter pointed out a possible NULL pointer dereference in the first version of this patch. ] Based-on-idea-by: Andrea Arcangeli <aarcange@redhat.com> Bug-Found-By: Dan Carpenter <dan.carpenter@oracle.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Rik van Riel <riel@redhat.com> [ Wrote changelog and fixed bug. ] Signed-off-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Mel Gorman <mgorman@suse.de> Reviewed-by: Rik van Riel <riel@redhat.com>
2012-10-25 12:16:45 +00:00
start = end;
sched/numa: Limit the amount of virtual memory scanned in task_numa_work() Currently task_numa_work() scans up to numa_balancing_scan_size_mb worth of memory per invocation, but only counts memory areas that have at least one PTE that is still present and not marked for numa hint faulting. It will skip over arbitarily large amounts of memory that are either unused, full of swap ptes, or full of PTEs that were already marked for NUMA hint faults but have not been faulted on yet. This can cause excessive amounts of CPU use, due to there being essentially no upper limit on the scan rate of very large processes that are not yet in a phase where they are actively accessing old memory pages (eg. they are still initializing their data). Avoid that problem by placing an upper limit on the amount of virtual memory that task_numa_work() scans in each invocation. This can be a higher limit than "pages", to ensure the task still skips over unused areas fairly quickly. While we are here, also fix the "nr_pte_updates" logic, so it only counts page ranges with ptes in them. Reported-by: Andrea Arcangeli <aarcange@redhat.com> Reported-by: Jan Stancek <jstancek@redhat.com> Signed-off-by: Rik van Riel <riel@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/20150911090027.4a7987bd@annuminas.surriel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-09-11 13:00:27 +00:00
if (pages <= 0 || virtpages <= 0)
goto out;
cond_resched();
} while (end != vma->vm_end);
}
mm: sched: numa: Implement constant, per task Working Set Sampling (WSS) rate Previously, to probe the working set of a task, we'd use a very simple and crude method: mark all of its address space PROT_NONE. That method has various (obvious) disadvantages: - it samples the working set at dissimilar rates, giving some tasks a sampling quality advantage over others. - creates performance problems for tasks with very large working sets - over-samples processes with large address spaces but which only very rarely execute Improve that method by keeping a rotating offset into the address space that marks the current position of the scan, and advance it by a constant rate (in a CPU cycles execution proportional manner). If the offset reaches the last mapped address of the mm then it then it starts over at the first address. The per-task nature of the working set sampling functionality in this tree allows such constant rate, per task, execution-weight proportional sampling of the working set, with an adaptive sampling interval/frequency that goes from once per 100ms up to just once per 8 seconds. The current sampling volume is 256 MB per interval. As tasks mature and converge their working set, so does the sampling rate slow down to just a trickle, 256 MB per 8 seconds of CPU time executed. This, beyond being adaptive, also rate-limits rarely executing systems and does not over-sample on overloaded systems. [ In AutoNUMA speak, this patch deals with the effective sampling rate of the 'hinting page fault'. AutoNUMA's scanning is currently rate-limited, but it is also fundamentally single-threaded, executing in the knuma_scand kernel thread, so the limit in AutoNUMA is global and does not scale up with the number of CPUs, nor does it scan tasks in an execution proportional manner. So the idea of rate-limiting the scanning was first implemented in the AutoNUMA tree via a global rate limit. This patch goes beyond that by implementing an execution rate proportional working set sampling rate that is not implemented via a single global scanning daemon. ] [ Dan Carpenter pointed out a possible NULL pointer dereference in the first version of this patch. ] Based-on-idea-by: Andrea Arcangeli <aarcange@redhat.com> Bug-Found-By: Dan Carpenter <dan.carpenter@oracle.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Rik van Riel <riel@redhat.com> [ Wrote changelog and fixed bug. ] Signed-off-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Mel Gorman <mgorman@suse.de> Reviewed-by: Rik van Riel <riel@redhat.com>
2012-10-25 12:16:45 +00:00
out:
mm: sched: numa: Implement constant, per task Working Set Sampling (WSS) rate Previously, to probe the working set of a task, we'd use a very simple and crude method: mark all of its address space PROT_NONE. That method has various (obvious) disadvantages: - it samples the working set at dissimilar rates, giving some tasks a sampling quality advantage over others. - creates performance problems for tasks with very large working sets - over-samples processes with large address spaces but which only very rarely execute Improve that method by keeping a rotating offset into the address space that marks the current position of the scan, and advance it by a constant rate (in a CPU cycles execution proportional manner). If the offset reaches the last mapped address of the mm then it then it starts over at the first address. The per-task nature of the working set sampling functionality in this tree allows such constant rate, per task, execution-weight proportional sampling of the working set, with an adaptive sampling interval/frequency that goes from once per 100ms up to just once per 8 seconds. The current sampling volume is 256 MB per interval. As tasks mature and converge their working set, so does the sampling rate slow down to just a trickle, 256 MB per 8 seconds of CPU time executed. This, beyond being adaptive, also rate-limits rarely executing systems and does not over-sample on overloaded systems. [ In AutoNUMA speak, this patch deals with the effective sampling rate of the 'hinting page fault'. AutoNUMA's scanning is currently rate-limited, but it is also fundamentally single-threaded, executing in the knuma_scand kernel thread, so the limit in AutoNUMA is global and does not scale up with the number of CPUs, nor does it scan tasks in an execution proportional manner. So the idea of rate-limiting the scanning was first implemented in the AutoNUMA tree via a global rate limit. This patch goes beyond that by implementing an execution rate proportional working set sampling rate that is not implemented via a single global scanning daemon. ] [ Dan Carpenter pointed out a possible NULL pointer dereference in the first version of this patch. ] Based-on-idea-by: Andrea Arcangeli <aarcange@redhat.com> Bug-Found-By: Dan Carpenter <dan.carpenter@oracle.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Rik van Riel <riel@redhat.com> [ Wrote changelog and fixed bug. ] Signed-off-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Mel Gorman <mgorman@suse.de> Reviewed-by: Rik van Riel <riel@redhat.com>
2012-10-25 12:16:45 +00:00
/*
* It is possible to reach the end of the VMA list but the last few
* VMAs are not guaranteed to the vma_migratable. If they are not, we
* would find the !migratable VMA on the next scan but not reset the
* scanner to the start so check it now.
mm: sched: numa: Implement constant, per task Working Set Sampling (WSS) rate Previously, to probe the working set of a task, we'd use a very simple and crude method: mark all of its address space PROT_NONE. That method has various (obvious) disadvantages: - it samples the working set at dissimilar rates, giving some tasks a sampling quality advantage over others. - creates performance problems for tasks with very large working sets - over-samples processes with large address spaces but which only very rarely execute Improve that method by keeping a rotating offset into the address space that marks the current position of the scan, and advance it by a constant rate (in a CPU cycles execution proportional manner). If the offset reaches the last mapped address of the mm then it then it starts over at the first address. The per-task nature of the working set sampling functionality in this tree allows such constant rate, per task, execution-weight proportional sampling of the working set, with an adaptive sampling interval/frequency that goes from once per 100ms up to just once per 8 seconds. The current sampling volume is 256 MB per interval. As tasks mature and converge their working set, so does the sampling rate slow down to just a trickle, 256 MB per 8 seconds of CPU time executed. This, beyond being adaptive, also rate-limits rarely executing systems and does not over-sample on overloaded systems. [ In AutoNUMA speak, this patch deals with the effective sampling rate of the 'hinting page fault'. AutoNUMA's scanning is currently rate-limited, but it is also fundamentally single-threaded, executing in the knuma_scand kernel thread, so the limit in AutoNUMA is global and does not scale up with the number of CPUs, nor does it scan tasks in an execution proportional manner. So the idea of rate-limiting the scanning was first implemented in the AutoNUMA tree via a global rate limit. This patch goes beyond that by implementing an execution rate proportional working set sampling rate that is not implemented via a single global scanning daemon. ] [ Dan Carpenter pointed out a possible NULL pointer dereference in the first version of this patch. ] Based-on-idea-by: Andrea Arcangeli <aarcange@redhat.com> Bug-Found-By: Dan Carpenter <dan.carpenter@oracle.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Rik van Riel <riel@redhat.com> [ Wrote changelog and fixed bug. ] Signed-off-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Mel Gorman <mgorman@suse.de> Reviewed-by: Rik van Riel <riel@redhat.com>
2012-10-25 12:16:45 +00:00
*/
if (vma)
mm->numa_scan_offset = start;
mm: sched: numa: Implement constant, per task Working Set Sampling (WSS) rate Previously, to probe the working set of a task, we'd use a very simple and crude method: mark all of its address space PROT_NONE. That method has various (obvious) disadvantages: - it samples the working set at dissimilar rates, giving some tasks a sampling quality advantage over others. - creates performance problems for tasks with very large working sets - over-samples processes with large address spaces but which only very rarely execute Improve that method by keeping a rotating offset into the address space that marks the current position of the scan, and advance it by a constant rate (in a CPU cycles execution proportional manner). If the offset reaches the last mapped address of the mm then it then it starts over at the first address. The per-task nature of the working set sampling functionality in this tree allows such constant rate, per task, execution-weight proportional sampling of the working set, with an adaptive sampling interval/frequency that goes from once per 100ms up to just once per 8 seconds. The current sampling volume is 256 MB per interval. As tasks mature and converge their working set, so does the sampling rate slow down to just a trickle, 256 MB per 8 seconds of CPU time executed. This, beyond being adaptive, also rate-limits rarely executing systems and does not over-sample on overloaded systems. [ In AutoNUMA speak, this patch deals with the effective sampling rate of the 'hinting page fault'. AutoNUMA's scanning is currently rate-limited, but it is also fundamentally single-threaded, executing in the knuma_scand kernel thread, so the limit in AutoNUMA is global and does not scale up with the number of CPUs, nor does it scan tasks in an execution proportional manner. So the idea of rate-limiting the scanning was first implemented in the AutoNUMA tree via a global rate limit. This patch goes beyond that by implementing an execution rate proportional working set sampling rate that is not implemented via a single global scanning daemon. ] [ Dan Carpenter pointed out a possible NULL pointer dereference in the first version of this patch. ] Based-on-idea-by: Andrea Arcangeli <aarcange@redhat.com> Bug-Found-By: Dan Carpenter <dan.carpenter@oracle.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Rik van Riel <riel@redhat.com> [ Wrote changelog and fixed bug. ] Signed-off-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Mel Gorman <mgorman@suse.de> Reviewed-by: Rik van Riel <riel@redhat.com>
2012-10-25 12:16:45 +00:00
else
reset_ptenuma_scan(p);
up_read(&mm->mmap_sem);
sched/numa: Cap PTE scanning overhead to 3% of run time There is a fundamental mismatch between the runtime based NUMA scanning at the task level, and the wall clock time NUMA scanning at the mm level. On a severely overloaded system, with very large processes, this mismatch can cause the system to spend all of its time in change_prot_numa(). This can happen if the task spends at least two ticks in change_prot_numa(), and only gets two ticks of CPU time in the real time between two scan intervals of the mm. This patch ensures that a task never spends more than 3% of run time scanning PTEs. It does that by ensuring that in-between task_numa_work() runs, the task spends at least 32x as much time on other things than it did on task_numa_work(). This is done stochastically: if a timer tick happens, or the task gets rescheduled during task_numa_work(), we delay a future run of task_numa_work() until the task has spent at least 32x the amount of CPU time doing something else, as it spent inside task_numa_work(). The longer task_numa_work() takes, the more likely it is this happens. If task_numa_work() takes very little time, chances are low that that code will do anything, but we will not care. Reported-and-tested-by: Jan Stancek <jstancek@redhat.com> Signed-off-by: Rik van Riel <riel@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: mgorman@suse.de Link: http://lkml.kernel.org/r/1446756983-28173-3-git-send-email-riel@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-11-05 20:56:23 +00:00
/*
* Make sure tasks use at least 32x as much time to run other code
* than they used here, to limit NUMA PTE scanning overhead to 3% max.
* Usually update_task_scan_period slows down scanning enough; on an
* overloaded system we need to limit overhead on a per task basis.
*/
if (unlikely(p->se.sum_exec_runtime != runtime)) {
u64 diff = p->se.sum_exec_runtime - runtime;
p->node_stamp += 32 * diff;
}
}
void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
{
int mm_users = 0;
struct mm_struct *mm = p->mm;
if (mm) {
mm_users = atomic_read(&mm->mm_users);
if (mm_users == 1) {
mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
mm->numa_scan_seq = 0;
}
}
p->node_stamp = 0;
p->numa_scan_seq = mm ? mm->numa_scan_seq : 0;
p->numa_scan_period = sysctl_numa_balancing_scan_delay;
/* Protect against double add, see task_tick_numa and task_numa_work */
p->numa_work.next = &p->numa_work;
p->numa_faults = NULL;
RCU_INIT_POINTER(p->numa_group, NULL);
p->last_task_numa_placement = 0;
p->last_sum_exec_runtime = 0;
init_task_work(&p->numa_work, task_numa_work);
/* New address space, reset the preferred nid */
if (!(clone_flags & CLONE_VM)) {
p->numa_preferred_nid = NUMA_NO_NODE;
return;
}
/*
* New thread, keep existing numa_preferred_nid which should be copied
* already by arch_dup_task_struct but stagger when scans start.
*/
if (mm) {
unsigned int delay;
delay = min_t(unsigned int, task_scan_max(current),
current->numa_scan_period * mm_users * NSEC_PER_MSEC);
delay += 2 * TICK_NSEC;
p->node_stamp = delay;
}
}
/*
* Drive the periodic memory faults..
*/
static void task_tick_numa(struct rq *rq, struct task_struct *curr)
{
struct callback_head *work = &curr->numa_work;
u64 period, now;
/*
* We don't care about NUMA placement if we don't have memory.
*/
if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
return;
/*
* Using runtime rather than walltime has the dual advantage that
* we (mostly) drive the selection from busy threads and that the
* task needs to have done some actual work before we bother with
* NUMA placement.
*/
now = curr->se.sum_exec_runtime;
period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
if (now > curr->node_stamp + period) {
mm: sched: numa: Implement slow start for working set sampling Add a 1 second delay before starting to scan the working set of a task and starting to balance it amongst nodes. [ note that before the constant per task WSS sampling rate patch the initial scan would happen much later still, in effect that patch caused this regression. ] The theory is that short-run tasks benefit very little from NUMA placement: they come and go, and they better stick to the node they were started on. As tasks mature and rebalance to other CPUs and nodes, so does their NUMA placement have to change and so does it start to matter more and more. In practice this change fixes an observable kbuild regression: # [ a perf stat --null --repeat 10 test of ten bzImage builds to /dev/shm ] !NUMA: 45.291088843 seconds time elapsed ( +- 0.40% ) 45.154231752 seconds time elapsed ( +- 0.36% ) +NUMA, no slow start: 46.172308123 seconds time elapsed ( +- 0.30% ) 46.343168745 seconds time elapsed ( +- 0.25% ) +NUMA, 1 sec slow start: 45.224189155 seconds time elapsed ( +- 0.25% ) 45.160866532 seconds time elapsed ( +- 0.17% ) and it also fixes an observable perf bench (hackbench) regression: # perf stat --null --repeat 10 perf bench sched messaging -NUMA: -NUMA: 0.246225691 seconds time elapsed ( +- 1.31% ) +NUMA no slow start: 0.252620063 seconds time elapsed ( +- 1.13% ) +NUMA 1sec delay: 0.248076230 seconds time elapsed ( +- 1.35% ) The implementation is simple and straightforward, most of the patch deals with adding the /proc/sys/kernel/numa_balancing_scan_delay_ms tunable knob. Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Rik van Riel <riel@redhat.com> [ Wrote the changelog, ran measurements, tuned the default. ] Signed-off-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Mel Gorman <mgorman@suse.de> Reviewed-by: Rik van Riel <riel@redhat.com>
2012-10-25 12:16:47 +00:00
if (!curr->node_stamp)
curr->numa_scan_period = task_scan_start(curr);
sched/numa: Mitigate chance that same task always updates PTEs With a trace_printk("working\n"); right after the cmpxchg in task_numa_work() we can see that of a 4 thread process, its always the same task winning the race and doing the protection change. This is a problem since the task doing the protection change has a penalty for taking faults -- it is busy when marking the PTEs. If its always the same task the ->numa_faults[] get severely skewed. Avoid this by delaying the task doing the protection change such that it is unlikely to win the privilege again. Before: root@interlagos:~# grep "thread 0/.*working" /debug/tracing/trace | tail -15 thread 0/0-3232 [022] .... 212.787402: task_numa_work: working thread 0/0-3232 [022] .... 212.888473: task_numa_work: working thread 0/0-3232 [022] .... 212.989538: task_numa_work: working thread 0/0-3232 [022] .... 213.090602: task_numa_work: working thread 0/0-3232 [022] .... 213.191667: task_numa_work: working thread 0/0-3232 [022] .... 213.292734: task_numa_work: working thread 0/0-3232 [022] .... 213.393804: task_numa_work: working thread 0/0-3232 [022] .... 213.494869: task_numa_work: working thread 0/0-3232 [022] .... 213.596937: task_numa_work: working thread 0/0-3232 [022] .... 213.699000: task_numa_work: working thread 0/0-3232 [022] .... 213.801067: task_numa_work: working thread 0/0-3232 [022] .... 213.903155: task_numa_work: working thread 0/0-3232 [022] .... 214.005201: task_numa_work: working thread 0/0-3232 [022] .... 214.107266: task_numa_work: working thread 0/0-3232 [022] .... 214.209342: task_numa_work: working After: root@interlagos:~# grep "thread 0/.*working" /debug/tracing/trace | tail -15 thread 0/0-3253 [005] .... 136.865051: task_numa_work: working thread 0/2-3255 [026] .... 136.965134: task_numa_work: working thread 0/3-3256 [024] .... 137.065217: task_numa_work: working thread 0/3-3256 [024] .... 137.165302: task_numa_work: working thread 0/3-3256 [024] .... 137.265382: task_numa_work: working thread 0/0-3253 [004] .... 137.366465: task_numa_work: working thread 0/2-3255 [026] .... 137.466549: task_numa_work: working thread 0/0-3253 [004] .... 137.566629: task_numa_work: working thread 0/0-3253 [004] .... 137.666711: task_numa_work: working thread 0/1-3254 [028] .... 137.766799: task_numa_work: working thread 0/0-3253 [004] .... 137.866876: task_numa_work: working thread 0/2-3255 [026] .... 137.966960: task_numa_work: working thread 0/1-3254 [028] .... 138.067041: task_numa_work: working thread 0/2-3255 [026] .... 138.167123: task_numa_work: working thread 0/3-3256 [024] .... 138.267207: task_numa_work: working Signed-off-by: Peter Zijlstra <peterz@infradead.org> Signed-off-by: Mel Gorman <mgorman@suse.de> Reviewed-by: Rik van Riel <riel@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Link: http://lkml.kernel.org/r/1381141781-10992-14-git-send-email-mgorman@suse.de Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-10-07 10:28:51 +00:00
curr->node_stamp += period;
if (!time_before(jiffies, curr->mm->numa_next_scan))
task_work_add(curr, work, true);
}
}
sched/numa: Reset scan rate whenever task moves across nodes Currently task scan rate is reset when NUMA balancer migrates the task to a different node. If NUMA balancer initiates a swap, reset is only applicable to the task that initiates the swap. Similarly no scan rate reset is done if the task is migrated across nodes by traditional load balancer. Instead move the scan reset to the migrate_task_rq. This ensures the task moved out of its preferred node, either gets back to its preferred node quickly or finds a new preferred node. Doing so, would be fair to all tasks migrating across nodes. Specjbb2005 results (8 warehouses) Higher bops are better 2 Socket - 2 Node Haswell - X86 JVMS Prev Current %Change 4 200668 203370 1.3465 1 321791 328431 2.06345 2 Socket - 4 Node Power8 - PowerNV JVMS Prev Current %Change 1 204848 206070 0.59654 2 Socket - 2 Node Power9 - PowerNV JVMS Prev Current %Change 4 188098 188386 0.153112 1 200351 201566 0.606436 4 Socket - 4 Node Power7 - PowerVM JVMS Prev Current %Change 8 58145.9 59157.4 1.73959 1 103798 105495 1.63491 Some events stats before and after applying the patch. perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 13,912,183 13,825,492 migrations 1,155,931 1,152,509 faults 367,139 371,948 cache-misses 54,240,196,814 55,654,206,041 sched:sched_move_numa 1,571 1,856 sched:sched_stick_numa 9 4 sched:sched_swap_numa 463 428 migrate:mm_migrate_pages 703 898 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 50155 57146 numa_hint_faults_local 45264 51612 numa_hit 239652 238164 numa_huge_pte_updates 36 16 numa_interleave 68 63 numa_local 239576 238085 numa_other 76 79 numa_pages_migrated 680 883 numa_pte_updates 71146 67540 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 3,156,720 3,288,525 migrations 30,354 38,652 faults 97,261 111,678 cache-misses 12,400,026,826 12,111,197,376 sched:sched_move_numa 4 900 sched:sched_stick_numa 0 0 sched:sched_swap_numa 1 5 migrate:mm_migrate_pages 20 714 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 272 18572 numa_hint_faults_local 186 14850 numa_hit 71362 73197 numa_huge_pte_updates 0 11 numa_interleave 23 25 numa_local 71299 73138 numa_other 63 59 numa_pages_migrated 2 712 numa_pte_updates 0 24021 perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 8,606,824 8,451,543 migrations 155,352 202,804 faults 301,409 310,024 cache-misses 157,759,224 253,522,507 sched:sched_move_numa 168 213 sched:sched_stick_numa 0 0 sched:sched_swap_numa 3 2 migrate:mm_migrate_pages 125 88 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 4650 11830 numa_hint_faults_local 3946 11301 numa_hit 90489 90038 numa_huge_pte_updates 0 0 numa_interleave 892 855 numa_local 90034 89796 numa_other 455 242 numa_pages_migrated 124 88 numa_pte_updates 4818 12039 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 2,113,167 2,049,153 migrations 10,533 11,405 faults 142,727 162,309 cache-misses 5,594,192 7,203,343 sched:sched_move_numa 10 22 sched:sched_stick_numa 0 0 sched:sched_swap_numa 0 0 migrate:mm_migrate_pages 6 1 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 744 1693 numa_hint_faults_local 584 1669 numa_hit 25551 25177 numa_huge_pte_updates 0 0 numa_interleave 263 194 numa_local 25302 24993 numa_other 249 184 numa_pages_migrated 6 1 numa_pte_updates 744 1577 perf stats 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 101,227,352 94,515,937 migrations 4,151,829 4,203,554 faults 745,233 832,697 cache-misses 224,669,561,766 226,248,698,331 sched:sched_move_numa 617 1,730 sched:sched_stick_numa 2 14 sched:sched_swap_numa 187 432 migrate:mm_migrate_pages 316 1,398 vmstat 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 24195 80079 numa_hint_faults_local 21639 68620 numa_hit 238331 241187 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 238331 241186 numa_other 0 1 numa_pages_migrated 204 1347 numa_pte_updates 24561 80729 perf stats 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 62,738,978 63,704,961 migrations 562,702 573,404 faults 228,465 230,878 cache-misses 75,778,067,952 76,568,222,781 sched:sched_move_numa 648 509 sched:sched_stick_numa 13 31 sched:sched_swap_numa 137 182 migrate:mm_migrate_pages 733 541 vmstat 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 10281 8501 numa_hint_faults_local 3242 2960 numa_hit 36338 35526 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 36338 35526 numa_other 0 0 numa_pages_migrated 706 539 numa_pte_updates 10176 8433 Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Jirka Hladky <jhladky@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1537552141-27815-4-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-09-21 17:48:58 +00:00
static void update_scan_period(struct task_struct *p, int new_cpu)
{
int src_nid = cpu_to_node(task_cpu(p));
int dst_nid = cpu_to_node(new_cpu);
sched/numa: Limit the conditions where scan period is reset migrate_task_rq_fair() resets the scan rate for NUMA balancing on every cross-node migration. In the event of excessive load balancing due to saturation, this may result in the scan rate being pegged at maximum and further overloading the machine. This patch only resets the scan if NUMA balancing is active, a preferred node has been selected and the task is being migrated from the preferred node as these are the most harmful. For example, a migration to the preferred node does not justify a faster scan rate. Similarly, a migration between two nodes that are not preferred is probably bouncing due to over-saturation of the machine. In that case, scanning faster and trapping more NUMA faults will further overload the machine. Specjbb2005 results (8 warehouses) Higher bops are better 2 Socket - 2 Node Haswell - X86 JVMS Prev Current %Change 4 203370 205332 0.964744 1 328431 319785 -2.63252 2 Socket - 4 Node Power8 - PowerNV JVMS Prev Current %Change 1 206070 206585 0.249915 2 Socket - 2 Node Power9 - PowerNV JVMS Prev Current %Change 4 188386 189162 0.41192 1 201566 213760 6.04963 4 Socket - 4 Node Power7 - PowerVM JVMS Prev Current %Change 8 59157.4 58736.8 -0.710985 1 105495 105419 -0.0720413 Some events stats before and after applying the patch. perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 13,825,492 14,285,708 migrations 1,152,509 1,180,621 faults 371,948 339,114 cache-misses 55,654,206,041 55,205,631,894 sched:sched_move_numa 1,856 843 sched:sched_stick_numa 4 6 sched:sched_swap_numa 428 219 migrate:mm_migrate_pages 898 365 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 57146 26907 numa_hint_faults_local 51612 24279 numa_hit 238164 239771 numa_huge_pte_updates 16 0 numa_interleave 63 68 numa_local 238085 239688 numa_other 79 83 numa_pages_migrated 883 363 numa_pte_updates 67540 27415 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 3,288,525 3,202,779 migrations 38,652 37,186 faults 111,678 106,076 cache-misses 12,111,197,376 12,024,873,744 sched:sched_move_numa 900 931 sched:sched_stick_numa 0 0 sched:sched_swap_numa 5 1 migrate:mm_migrate_pages 714 637 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 18572 17409 numa_hint_faults_local 14850 14367 numa_hit 73197 73953 numa_huge_pte_updates 11 20 numa_interleave 25 25 numa_local 73138 73892 numa_other 59 61 numa_pages_migrated 712 668 numa_pte_updates 24021 27276 perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 8,451,543 8,474,013 migrations 202,804 254,934 faults 310,024 320,506 cache-misses 253,522,507 110,580,458 sched:sched_move_numa 213 725 sched:sched_stick_numa 0 0 sched:sched_swap_numa 2 7 migrate:mm_migrate_pages 88 145 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 11830 22797 numa_hint_faults_local 11301 21539 numa_hit 90038 89308 numa_huge_pte_updates 0 0 numa_interleave 855 865 numa_local 89796 88955 numa_other 242 353 numa_pages_migrated 88 149 numa_pte_updates 12039 22930 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 2,049,153 2,195,628 migrations 11,405 11,179 faults 162,309 149,656 cache-misses 7,203,343 8,117,515 sched:sched_move_numa 22 49 sched:sched_stick_numa 0 0 sched:sched_swap_numa 0 0 migrate:mm_migrate_pages 1 5 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 1693 3577 numa_hint_faults_local 1669 3476 numa_hit 25177 26142 numa_huge_pte_updates 0 0 numa_interleave 194 358 numa_local 24993 26042 numa_other 184 100 numa_pages_migrated 1 5 numa_pte_updates 1577 3587 perf stats 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 94,515,937 100,602,296 migrations 4,203,554 4,135,630 faults 832,697 789,256 cache-misses 226,248,698,331 226,160,621,058 sched:sched_move_numa 1,730 1,366 sched:sched_stick_numa 14 16 sched:sched_swap_numa 432 374 migrate:mm_migrate_pages 1,398 1,350 vmstat 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 80079 47857 numa_hint_faults_local 68620 39768 numa_hit 241187 240165 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 241186 240165 numa_other 1 0 numa_pages_migrated 1347 1224 numa_pte_updates 80729 48354 perf stats 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 63,704,961 58,515,496 migrations 573,404 564,845 faults 230,878 245,807 cache-misses 76,568,222,781 73,603,757,976 sched:sched_move_numa 509 996 sched:sched_stick_numa 31 10 sched:sched_swap_numa 182 193 migrate:mm_migrate_pages 541 646 vmstat 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 8501 13422 numa_hint_faults_local 2960 5619 numa_hit 35526 36118 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 35526 36116 numa_other 0 2 numa_pages_migrated 539 616 numa_pte_updates 8433 13374 Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Jirka Hladky <jhladky@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1537552141-27815-5-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-09-21 17:48:59 +00:00
if (!static_branch_likely(&sched_numa_balancing))
return;
sched/numa: Reset scan rate whenever task moves across nodes Currently task scan rate is reset when NUMA balancer migrates the task to a different node. If NUMA balancer initiates a swap, reset is only applicable to the task that initiates the swap. Similarly no scan rate reset is done if the task is migrated across nodes by traditional load balancer. Instead move the scan reset to the migrate_task_rq. This ensures the task moved out of its preferred node, either gets back to its preferred node quickly or finds a new preferred node. Doing so, would be fair to all tasks migrating across nodes. Specjbb2005 results (8 warehouses) Higher bops are better 2 Socket - 2 Node Haswell - X86 JVMS Prev Current %Change 4 200668 203370 1.3465 1 321791 328431 2.06345 2 Socket - 4 Node Power8 - PowerNV JVMS Prev Current %Change 1 204848 206070 0.59654 2 Socket - 2 Node Power9 - PowerNV JVMS Prev Current %Change 4 188098 188386 0.153112 1 200351 201566 0.606436 4 Socket - 4 Node Power7 - PowerVM JVMS Prev Current %Change 8 58145.9 59157.4 1.73959 1 103798 105495 1.63491 Some events stats before and after applying the patch. perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 13,912,183 13,825,492 migrations 1,155,931 1,152,509 faults 367,139 371,948 cache-misses 54,240,196,814 55,654,206,041 sched:sched_move_numa 1,571 1,856 sched:sched_stick_numa 9 4 sched:sched_swap_numa 463 428 migrate:mm_migrate_pages 703 898 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 50155 57146 numa_hint_faults_local 45264 51612 numa_hit 239652 238164 numa_huge_pte_updates 36 16 numa_interleave 68 63 numa_local 239576 238085 numa_other 76 79 numa_pages_migrated 680 883 numa_pte_updates 71146 67540 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 3,156,720 3,288,525 migrations 30,354 38,652 faults 97,261 111,678 cache-misses 12,400,026,826 12,111,197,376 sched:sched_move_numa 4 900 sched:sched_stick_numa 0 0 sched:sched_swap_numa 1 5 migrate:mm_migrate_pages 20 714 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 272 18572 numa_hint_faults_local 186 14850 numa_hit 71362 73197 numa_huge_pte_updates 0 11 numa_interleave 23 25 numa_local 71299 73138 numa_other 63 59 numa_pages_migrated 2 712 numa_pte_updates 0 24021 perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 8,606,824 8,451,543 migrations 155,352 202,804 faults 301,409 310,024 cache-misses 157,759,224 253,522,507 sched:sched_move_numa 168 213 sched:sched_stick_numa 0 0 sched:sched_swap_numa 3 2 migrate:mm_migrate_pages 125 88 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 4650 11830 numa_hint_faults_local 3946 11301 numa_hit 90489 90038 numa_huge_pte_updates 0 0 numa_interleave 892 855 numa_local 90034 89796 numa_other 455 242 numa_pages_migrated 124 88 numa_pte_updates 4818 12039 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 2,113,167 2,049,153 migrations 10,533 11,405 faults 142,727 162,309 cache-misses 5,594,192 7,203,343 sched:sched_move_numa 10 22 sched:sched_stick_numa 0 0 sched:sched_swap_numa 0 0 migrate:mm_migrate_pages 6 1 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 744 1693 numa_hint_faults_local 584 1669 numa_hit 25551 25177 numa_huge_pte_updates 0 0 numa_interleave 263 194 numa_local 25302 24993 numa_other 249 184 numa_pages_migrated 6 1 numa_pte_updates 744 1577 perf stats 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 101,227,352 94,515,937 migrations 4,151,829 4,203,554 faults 745,233 832,697 cache-misses 224,669,561,766 226,248,698,331 sched:sched_move_numa 617 1,730 sched:sched_stick_numa 2 14 sched:sched_swap_numa 187 432 migrate:mm_migrate_pages 316 1,398 vmstat 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 24195 80079 numa_hint_faults_local 21639 68620 numa_hit 238331 241187 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 238331 241186 numa_other 0 1 numa_pages_migrated 204 1347 numa_pte_updates 24561 80729 perf stats 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 62,738,978 63,704,961 migrations 562,702 573,404 faults 228,465 230,878 cache-misses 75,778,067,952 76,568,222,781 sched:sched_move_numa 648 509 sched:sched_stick_numa 13 31 sched:sched_swap_numa 137 182 migrate:mm_migrate_pages 733 541 vmstat 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 10281 8501 numa_hint_faults_local 3242 2960 numa_hit 36338 35526 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 36338 35526 numa_other 0 0 numa_pages_migrated 706 539 numa_pte_updates 10176 8433 Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Jirka Hladky <jhladky@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1537552141-27815-4-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-09-21 17:48:58 +00:00
if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
return;
sched/numa: Limit the conditions where scan period is reset migrate_task_rq_fair() resets the scan rate for NUMA balancing on every cross-node migration. In the event of excessive load balancing due to saturation, this may result in the scan rate being pegged at maximum and further overloading the machine. This patch only resets the scan if NUMA balancing is active, a preferred node has been selected and the task is being migrated from the preferred node as these are the most harmful. For example, a migration to the preferred node does not justify a faster scan rate. Similarly, a migration between two nodes that are not preferred is probably bouncing due to over-saturation of the machine. In that case, scanning faster and trapping more NUMA faults will further overload the machine. Specjbb2005 results (8 warehouses) Higher bops are better 2 Socket - 2 Node Haswell - X86 JVMS Prev Current %Change 4 203370 205332 0.964744 1 328431 319785 -2.63252 2 Socket - 4 Node Power8 - PowerNV JVMS Prev Current %Change 1 206070 206585 0.249915 2 Socket - 2 Node Power9 - PowerNV JVMS Prev Current %Change 4 188386 189162 0.41192 1 201566 213760 6.04963 4 Socket - 4 Node Power7 - PowerVM JVMS Prev Current %Change 8 59157.4 58736.8 -0.710985 1 105495 105419 -0.0720413 Some events stats before and after applying the patch. perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 13,825,492 14,285,708 migrations 1,152,509 1,180,621 faults 371,948 339,114 cache-misses 55,654,206,041 55,205,631,894 sched:sched_move_numa 1,856 843 sched:sched_stick_numa 4 6 sched:sched_swap_numa 428 219 migrate:mm_migrate_pages 898 365 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 57146 26907 numa_hint_faults_local 51612 24279 numa_hit 238164 239771 numa_huge_pte_updates 16 0 numa_interleave 63 68 numa_local 238085 239688 numa_other 79 83 numa_pages_migrated 883 363 numa_pte_updates 67540 27415 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 3,288,525 3,202,779 migrations 38,652 37,186 faults 111,678 106,076 cache-misses 12,111,197,376 12,024,873,744 sched:sched_move_numa 900 931 sched:sched_stick_numa 0 0 sched:sched_swap_numa 5 1 migrate:mm_migrate_pages 714 637 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 18572 17409 numa_hint_faults_local 14850 14367 numa_hit 73197 73953 numa_huge_pte_updates 11 20 numa_interleave 25 25 numa_local 73138 73892 numa_other 59 61 numa_pages_migrated 712 668 numa_pte_updates 24021 27276 perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 8,451,543 8,474,013 migrations 202,804 254,934 faults 310,024 320,506 cache-misses 253,522,507 110,580,458 sched:sched_move_numa 213 725 sched:sched_stick_numa 0 0 sched:sched_swap_numa 2 7 migrate:mm_migrate_pages 88 145 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 11830 22797 numa_hint_faults_local 11301 21539 numa_hit 90038 89308 numa_huge_pte_updates 0 0 numa_interleave 855 865 numa_local 89796 88955 numa_other 242 353 numa_pages_migrated 88 149 numa_pte_updates 12039 22930 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 2,049,153 2,195,628 migrations 11,405 11,179 faults 162,309 149,656 cache-misses 7,203,343 8,117,515 sched:sched_move_numa 22 49 sched:sched_stick_numa 0 0 sched:sched_swap_numa 0 0 migrate:mm_migrate_pages 1 5 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 1693 3577 numa_hint_faults_local 1669 3476 numa_hit 25177 26142 numa_huge_pte_updates 0 0 numa_interleave 194 358 numa_local 24993 26042 numa_other 184 100 numa_pages_migrated 1 5 numa_pte_updates 1577 3587 perf stats 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 94,515,937 100,602,296 migrations 4,203,554 4,135,630 faults 832,697 789,256 cache-misses 226,248,698,331 226,160,621,058 sched:sched_move_numa 1,730 1,366 sched:sched_stick_numa 14 16 sched:sched_swap_numa 432 374 migrate:mm_migrate_pages 1,398 1,350 vmstat 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 80079 47857 numa_hint_faults_local 68620 39768 numa_hit 241187 240165 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 241186 240165 numa_other 1 0 numa_pages_migrated 1347 1224 numa_pte_updates 80729 48354 perf stats 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 63,704,961 58,515,496 migrations 573,404 564,845 faults 230,878 245,807 cache-misses 76,568,222,781 73,603,757,976 sched:sched_move_numa 509 996 sched:sched_stick_numa 31 10 sched:sched_swap_numa 182 193 migrate:mm_migrate_pages 541 646 vmstat 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 8501 13422 numa_hint_faults_local 2960 5619 numa_hit 35526 36118 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 35526 36116 numa_other 0 2 numa_pages_migrated 539 616 numa_pte_updates 8433 13374 Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Jirka Hladky <jhladky@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1537552141-27815-5-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-09-21 17:48:59 +00:00
if (src_nid == dst_nid)
return;
/*
* Allow resets if faults have been trapped before one scan
* has completed. This is most likely due to a new task that
* is pulled cross-node due to wakeups or load balancing.
*/
if (p->numa_scan_seq) {
/*
* Avoid scan adjustments if moving to the preferred
* node or if the task was not previously running on
* the preferred node.
*/
if (dst_nid == p->numa_preferred_nid ||
mm: replace all open encodings for NUMA_NO_NODE Patch series "Replace all open encodings for NUMA_NO_NODE", v3. All these places for replacement were found by running the following grep patterns on the entire kernel code. Please let me know if this might have missed some instances. This might also have replaced some false positives. I will appreciate suggestions, inputs and review. 1. git grep "nid == -1" 2. git grep "node == -1" 3. git grep "nid = -1" 4. git grep "node = -1" This patch (of 2): At present there are multiple places where invalid node number is encoded as -1. Even though implicitly understood it is always better to have macros in there. Replace these open encodings for an invalid node number with the global macro NUMA_NO_NODE. This helps remove NUMA related assumptions like 'invalid node' from various places redirecting them to a common definition. Link: http://lkml.kernel.org/r/1545127933-10711-2-git-send-email-anshuman.khandual@arm.com Signed-off-by: Anshuman Khandual <anshuman.khandual@arm.com> Reviewed-by: David Hildenbrand <david@redhat.com> Acked-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com> [ixgbe] Acked-by: Jens Axboe <axboe@kernel.dk> [mtip32xx] Acked-by: Vinod Koul <vkoul@kernel.org> [dmaengine.c] Acked-by: Michael Ellerman <mpe@ellerman.id.au> [powerpc] Acked-by: Doug Ledford <dledford@redhat.com> [drivers/infiniband] Cc: Joseph Qi <jiangqi903@gmail.com> Cc: Hans Verkuil <hverkuil@xs4all.nl> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:42:58 +00:00
(p->numa_preferred_nid != NUMA_NO_NODE &&
src_nid != p->numa_preferred_nid))
sched/numa: Limit the conditions where scan period is reset migrate_task_rq_fair() resets the scan rate for NUMA balancing on every cross-node migration. In the event of excessive load balancing due to saturation, this may result in the scan rate being pegged at maximum and further overloading the machine. This patch only resets the scan if NUMA balancing is active, a preferred node has been selected and the task is being migrated from the preferred node as these are the most harmful. For example, a migration to the preferred node does not justify a faster scan rate. Similarly, a migration between two nodes that are not preferred is probably bouncing due to over-saturation of the machine. In that case, scanning faster and trapping more NUMA faults will further overload the machine. Specjbb2005 results (8 warehouses) Higher bops are better 2 Socket - 2 Node Haswell - X86 JVMS Prev Current %Change 4 203370 205332 0.964744 1 328431 319785 -2.63252 2 Socket - 4 Node Power8 - PowerNV JVMS Prev Current %Change 1 206070 206585 0.249915 2 Socket - 2 Node Power9 - PowerNV JVMS Prev Current %Change 4 188386 189162 0.41192 1 201566 213760 6.04963 4 Socket - 4 Node Power7 - PowerVM JVMS Prev Current %Change 8 59157.4 58736.8 -0.710985 1 105495 105419 -0.0720413 Some events stats before and after applying the patch. perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 13,825,492 14,285,708 migrations 1,152,509 1,180,621 faults 371,948 339,114 cache-misses 55,654,206,041 55,205,631,894 sched:sched_move_numa 1,856 843 sched:sched_stick_numa 4 6 sched:sched_swap_numa 428 219 migrate:mm_migrate_pages 898 365 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 57146 26907 numa_hint_faults_local 51612 24279 numa_hit 238164 239771 numa_huge_pte_updates 16 0 numa_interleave 63 68 numa_local 238085 239688 numa_other 79 83 numa_pages_migrated 883 363 numa_pte_updates 67540 27415 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 3,288,525 3,202,779 migrations 38,652 37,186 faults 111,678 106,076 cache-misses 12,111,197,376 12,024,873,744 sched:sched_move_numa 900 931 sched:sched_stick_numa 0 0 sched:sched_swap_numa 5 1 migrate:mm_migrate_pages 714 637 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 18572 17409 numa_hint_faults_local 14850 14367 numa_hit 73197 73953 numa_huge_pte_updates 11 20 numa_interleave 25 25 numa_local 73138 73892 numa_other 59 61 numa_pages_migrated 712 668 numa_pte_updates 24021 27276 perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 8,451,543 8,474,013 migrations 202,804 254,934 faults 310,024 320,506 cache-misses 253,522,507 110,580,458 sched:sched_move_numa 213 725 sched:sched_stick_numa 0 0 sched:sched_swap_numa 2 7 migrate:mm_migrate_pages 88 145 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 11830 22797 numa_hint_faults_local 11301 21539 numa_hit 90038 89308 numa_huge_pte_updates 0 0 numa_interleave 855 865 numa_local 89796 88955 numa_other 242 353 numa_pages_migrated 88 149 numa_pte_updates 12039 22930 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 2,049,153 2,195,628 migrations 11,405 11,179 faults 162,309 149,656 cache-misses 7,203,343 8,117,515 sched:sched_move_numa 22 49 sched:sched_stick_numa 0 0 sched:sched_swap_numa 0 0 migrate:mm_migrate_pages 1 5 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 1693 3577 numa_hint_faults_local 1669 3476 numa_hit 25177 26142 numa_huge_pte_updates 0 0 numa_interleave 194 358 numa_local 24993 26042 numa_other 184 100 numa_pages_migrated 1 5 numa_pte_updates 1577 3587 perf stats 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 94,515,937 100,602,296 migrations 4,203,554 4,135,630 faults 832,697 789,256 cache-misses 226,248,698,331 226,160,621,058 sched:sched_move_numa 1,730 1,366 sched:sched_stick_numa 14 16 sched:sched_swap_numa 432 374 migrate:mm_migrate_pages 1,398 1,350 vmstat 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 80079 47857 numa_hint_faults_local 68620 39768 numa_hit 241187 240165 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 241186 240165 numa_other 1 0 numa_pages_migrated 1347 1224 numa_pte_updates 80729 48354 perf stats 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 63,704,961 58,515,496 migrations 573,404 564,845 faults 230,878 245,807 cache-misses 76,568,222,781 73,603,757,976 sched:sched_move_numa 509 996 sched:sched_stick_numa 31 10 sched:sched_swap_numa 182 193 migrate:mm_migrate_pages 541 646 vmstat 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 8501 13422 numa_hint_faults_local 2960 5619 numa_hit 35526 36118 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 35526 36116 numa_other 0 2 numa_pages_migrated 539 616 numa_pte_updates 8433 13374 Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Jirka Hladky <jhladky@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1537552141-27815-5-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-09-21 17:48:59 +00:00
return;
}
p->numa_scan_period = task_scan_start(p);
sched/numa: Reset scan rate whenever task moves across nodes Currently task scan rate is reset when NUMA balancer migrates the task to a different node. If NUMA balancer initiates a swap, reset is only applicable to the task that initiates the swap. Similarly no scan rate reset is done if the task is migrated across nodes by traditional load balancer. Instead move the scan reset to the migrate_task_rq. This ensures the task moved out of its preferred node, either gets back to its preferred node quickly or finds a new preferred node. Doing so, would be fair to all tasks migrating across nodes. Specjbb2005 results (8 warehouses) Higher bops are better 2 Socket - 2 Node Haswell - X86 JVMS Prev Current %Change 4 200668 203370 1.3465 1 321791 328431 2.06345 2 Socket - 4 Node Power8 - PowerNV JVMS Prev Current %Change 1 204848 206070 0.59654 2 Socket - 2 Node Power9 - PowerNV JVMS Prev Current %Change 4 188098 188386 0.153112 1 200351 201566 0.606436 4 Socket - 4 Node Power7 - PowerVM JVMS Prev Current %Change 8 58145.9 59157.4 1.73959 1 103798 105495 1.63491 Some events stats before and after applying the patch. perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 13,912,183 13,825,492 migrations 1,155,931 1,152,509 faults 367,139 371,948 cache-misses 54,240,196,814 55,654,206,041 sched:sched_move_numa 1,571 1,856 sched:sched_stick_numa 9 4 sched:sched_swap_numa 463 428 migrate:mm_migrate_pages 703 898 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 50155 57146 numa_hint_faults_local 45264 51612 numa_hit 239652 238164 numa_huge_pte_updates 36 16 numa_interleave 68 63 numa_local 239576 238085 numa_other 76 79 numa_pages_migrated 680 883 numa_pte_updates 71146 67540 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 3,156,720 3,288,525 migrations 30,354 38,652 faults 97,261 111,678 cache-misses 12,400,026,826 12,111,197,376 sched:sched_move_numa 4 900 sched:sched_stick_numa 0 0 sched:sched_swap_numa 1 5 migrate:mm_migrate_pages 20 714 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 272 18572 numa_hint_faults_local 186 14850 numa_hit 71362 73197 numa_huge_pte_updates 0 11 numa_interleave 23 25 numa_local 71299 73138 numa_other 63 59 numa_pages_migrated 2 712 numa_pte_updates 0 24021 perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 8,606,824 8,451,543 migrations 155,352 202,804 faults 301,409 310,024 cache-misses 157,759,224 253,522,507 sched:sched_move_numa 168 213 sched:sched_stick_numa 0 0 sched:sched_swap_numa 3 2 migrate:mm_migrate_pages 125 88 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 4650 11830 numa_hint_faults_local 3946 11301 numa_hit 90489 90038 numa_huge_pte_updates 0 0 numa_interleave 892 855 numa_local 90034 89796 numa_other 455 242 numa_pages_migrated 124 88 numa_pte_updates 4818 12039 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 2,113,167 2,049,153 migrations 10,533 11,405 faults 142,727 162,309 cache-misses 5,594,192 7,203,343 sched:sched_move_numa 10 22 sched:sched_stick_numa 0 0 sched:sched_swap_numa 0 0 migrate:mm_migrate_pages 6 1 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 744 1693 numa_hint_faults_local 584 1669 numa_hit 25551 25177 numa_huge_pte_updates 0 0 numa_interleave 263 194 numa_local 25302 24993 numa_other 249 184 numa_pages_migrated 6 1 numa_pte_updates 744 1577 perf stats 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 101,227,352 94,515,937 migrations 4,151,829 4,203,554 faults 745,233 832,697 cache-misses 224,669,561,766 226,248,698,331 sched:sched_move_numa 617 1,730 sched:sched_stick_numa 2 14 sched:sched_swap_numa 187 432 migrate:mm_migrate_pages 316 1,398 vmstat 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 24195 80079 numa_hint_faults_local 21639 68620 numa_hit 238331 241187 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 238331 241186 numa_other 0 1 numa_pages_migrated 204 1347 numa_pte_updates 24561 80729 perf stats 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 62,738,978 63,704,961 migrations 562,702 573,404 faults 228,465 230,878 cache-misses 75,778,067,952 76,568,222,781 sched:sched_move_numa 648 509 sched:sched_stick_numa 13 31 sched:sched_swap_numa 137 182 migrate:mm_migrate_pages 733 541 vmstat 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 10281 8501 numa_hint_faults_local 3242 2960 numa_hit 36338 35526 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 36338 35526 numa_other 0 0 numa_pages_migrated 706 539 numa_pte_updates 10176 8433 Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Jirka Hladky <jhladky@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1537552141-27815-4-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-09-21 17:48:58 +00:00
}
#else
static void task_tick_numa(struct rq *rq, struct task_struct *curr)
{
}
static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
{
}
static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
{
}
sched/numa: Reset scan rate whenever task moves across nodes Currently task scan rate is reset when NUMA balancer migrates the task to a different node. If NUMA balancer initiates a swap, reset is only applicable to the task that initiates the swap. Similarly no scan rate reset is done if the task is migrated across nodes by traditional load balancer. Instead move the scan reset to the migrate_task_rq. This ensures the task moved out of its preferred node, either gets back to its preferred node quickly or finds a new preferred node. Doing so, would be fair to all tasks migrating across nodes. Specjbb2005 results (8 warehouses) Higher bops are better 2 Socket - 2 Node Haswell - X86 JVMS Prev Current %Change 4 200668 203370 1.3465 1 321791 328431 2.06345 2 Socket - 4 Node Power8 - PowerNV JVMS Prev Current %Change 1 204848 206070 0.59654 2 Socket - 2 Node Power9 - PowerNV JVMS Prev Current %Change 4 188098 188386 0.153112 1 200351 201566 0.606436 4 Socket - 4 Node Power7 - PowerVM JVMS Prev Current %Change 8 58145.9 59157.4 1.73959 1 103798 105495 1.63491 Some events stats before and after applying the patch. perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 13,912,183 13,825,492 migrations 1,155,931 1,152,509 faults 367,139 371,948 cache-misses 54,240,196,814 55,654,206,041 sched:sched_move_numa 1,571 1,856 sched:sched_stick_numa 9 4 sched:sched_swap_numa 463 428 migrate:mm_migrate_pages 703 898 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 50155 57146 numa_hint_faults_local 45264 51612 numa_hit 239652 238164 numa_huge_pte_updates 36 16 numa_interleave 68 63 numa_local 239576 238085 numa_other 76 79 numa_pages_migrated 680 883 numa_pte_updates 71146 67540 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 3,156,720 3,288,525 migrations 30,354 38,652 faults 97,261 111,678 cache-misses 12,400,026,826 12,111,197,376 sched:sched_move_numa 4 900 sched:sched_stick_numa 0 0 sched:sched_swap_numa 1 5 migrate:mm_migrate_pages 20 714 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 272 18572 numa_hint_faults_local 186 14850 numa_hit 71362 73197 numa_huge_pte_updates 0 11 numa_interleave 23 25 numa_local 71299 73138 numa_other 63 59 numa_pages_migrated 2 712 numa_pte_updates 0 24021 perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 8,606,824 8,451,543 migrations 155,352 202,804 faults 301,409 310,024 cache-misses 157,759,224 253,522,507 sched:sched_move_numa 168 213 sched:sched_stick_numa 0 0 sched:sched_swap_numa 3 2 migrate:mm_migrate_pages 125 88 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 4650 11830 numa_hint_faults_local 3946 11301 numa_hit 90489 90038 numa_huge_pte_updates 0 0 numa_interleave 892 855 numa_local 90034 89796 numa_other 455 242 numa_pages_migrated 124 88 numa_pte_updates 4818 12039 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 2,113,167 2,049,153 migrations 10,533 11,405 faults 142,727 162,309 cache-misses 5,594,192 7,203,343 sched:sched_move_numa 10 22 sched:sched_stick_numa 0 0 sched:sched_swap_numa 0 0 migrate:mm_migrate_pages 6 1 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 744 1693 numa_hint_faults_local 584 1669 numa_hit 25551 25177 numa_huge_pte_updates 0 0 numa_interleave 263 194 numa_local 25302 24993 numa_other 249 184 numa_pages_migrated 6 1 numa_pte_updates 744 1577 perf stats 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 101,227,352 94,515,937 migrations 4,151,829 4,203,554 faults 745,233 832,697 cache-misses 224,669,561,766 226,248,698,331 sched:sched_move_numa 617 1,730 sched:sched_stick_numa 2 14 sched:sched_swap_numa 187 432 migrate:mm_migrate_pages 316 1,398 vmstat 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 24195 80079 numa_hint_faults_local 21639 68620 numa_hit 238331 241187 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 238331 241186 numa_other 0 1 numa_pages_migrated 204 1347 numa_pte_updates 24561 80729 perf stats 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 62,738,978 63,704,961 migrations 562,702 573,404 faults 228,465 230,878 cache-misses 75,778,067,952 76,568,222,781 sched:sched_move_numa 648 509 sched:sched_stick_numa 13 31 sched:sched_swap_numa 137 182 migrate:mm_migrate_pages 733 541 vmstat 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 10281 8501 numa_hint_faults_local 3242 2960 numa_hit 36338 35526 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 36338 35526 numa_other 0 0 numa_pages_migrated 706 539 numa_pte_updates 10176 8433 Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Jirka Hladky <jhladky@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1537552141-27815-4-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-09-21 17:48:58 +00:00
static inline void update_scan_period(struct task_struct *p, int new_cpu)
{
}
#endif /* CONFIG_NUMA_BALANCING */
static void
account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
{
update_load_add(&cfs_rq->load, se->load.weight);
#ifdef CONFIG_SMP
if (entity_is_task(se)) {
struct rq *rq = rq_of(cfs_rq);
account_numa_enqueue(rq, task_of(se));
list_add(&se->group_node, &rq->cfs_tasks);
}
#endif
cfs_rq->nr_running++;
}
static void
account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
{
update_load_sub(&cfs_rq->load, se->load.weight);
#ifdef CONFIG_SMP
if (entity_is_task(se)) {
account_numa_dequeue(rq_of(cfs_rq), task_of(se));
list_del_init(&se->group_node);
}
#endif
cfs_rq->nr_running--;
}
/*
* Signed add and clamp on underflow.
*
* Explicitly do a load-store to ensure the intermediate value never hits
* memory. This allows lockless observations without ever seeing the negative
* values.
*/
#define add_positive(_ptr, _val) do { \
typeof(_ptr) ptr = (_ptr); \
typeof(_val) val = (_val); \
typeof(*ptr) res, var = READ_ONCE(*ptr); \
\
res = var + val; \
\
if (val < 0 && res > var) \
res = 0; \
\
WRITE_ONCE(*ptr, res); \
} while (0)
/*
* Unsigned subtract and clamp on underflow.
*
* Explicitly do a load-store to ensure the intermediate value never hits
* memory. This allows lockless observations without ever seeing the negative
* values.
*/
#define sub_positive(_ptr, _val) do { \
typeof(_ptr) ptr = (_ptr); \
typeof(*ptr) val = (_val); \
typeof(*ptr) res, var = READ_ONCE(*ptr); \
res = var - val; \
if (res > var) \
res = 0; \
WRITE_ONCE(*ptr, res); \
} while (0)
/*
* Remove and clamp on negative, from a local variable.
*
* A variant of sub_positive(), which does not use explicit load-store
* and is thus optimized for local variable updates.
*/
#define lsub_positive(_ptr, _val) do { \
typeof(_ptr) ptr = (_ptr); \
*ptr -= min_t(typeof(*ptr), *ptr, _val); \
} while (0)
#ifdef CONFIG_SMP
static inline void
enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
{
cfs_rq->avg.load_avg += se->avg.load_avg;
cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
}
static inline void
dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
{
sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
}
#else
static inline void
enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
static inline void
dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
#endif
static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
unsigned long weight)
{
if (se->on_rq) {
/* commit outstanding execution time */
if (cfs_rq->curr == se)
update_curr(cfs_rq);
account_entity_dequeue(cfs_rq, se);
}
dequeue_load_avg(cfs_rq, se);
update_load_set(&se->load, weight);
#ifdef CONFIG_SMP
do {
u32 divider = LOAD_AVG_MAX - 1024 + se->avg.period_contrib;
se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
} while (0);
#endif
enqueue_load_avg(cfs_rq, se);
if (se->on_rq)
account_entity_enqueue(cfs_rq, se);
}
void reweight_task(struct task_struct *p, int prio)
{
struct sched_entity *se = &p->se;
struct cfs_rq *cfs_rq = cfs_rq_of(se);
struct load_weight *load = &se->load;
unsigned long weight = scale_load(sched_prio_to_weight[prio]);
reweight_entity(cfs_rq, se, weight);
load->inv_weight = sched_prio_to_wmult[prio];
}
#ifdef CONFIG_FAIR_GROUP_SCHED
#ifdef CONFIG_SMP
/*
* All this does is approximate the hierarchical proportion which includes that
* global sum we all love to hate.
*
* That is, the weight of a group entity, is the proportional share of the
* group weight based on the group runqueue weights. That is:
*
* tg->weight * grq->load.weight
* ge->load.weight = ----------------------------- (1)
* \Sum grq->load.weight
*
* Now, because computing that sum is prohibitively expensive to compute (been
* there, done that) we approximate it with this average stuff. The average
* moves slower and therefore the approximation is cheaper and more stable.
*
* So instead of the above, we substitute:
*
* grq->load.weight -> grq->avg.load_avg (2)
*
* which yields the following:
*
* tg->weight * grq->avg.load_avg
* ge->load.weight = ------------------------------ (3)
* tg->load_avg
*
* Where: tg->load_avg ~= \Sum grq->avg.load_avg
*
* That is shares_avg, and it is right (given the approximation (2)).
*
* The problem with it is that because the average is slow -- it was designed
* to be exactly that of course -- this leads to transients in boundary
* conditions. In specific, the case where the group was idle and we start the
* one task. It takes time for our CPU's grq->avg.load_avg to build up,
* yielding bad latency etc..
*
* Now, in that special case (1) reduces to:
*
* tg->weight * grq->load.weight
* ge->load.weight = ----------------------------- = tg->weight (4)
* grp->load.weight
*
* That is, the sum collapses because all other CPUs are idle; the UP scenario.
*
* So what we do is modify our approximation (3) to approach (4) in the (near)
* UP case, like:
*
* ge->load.weight =
*
* tg->weight * grq->load.weight
* --------------------------------------------------- (5)
* tg->load_avg - grq->avg.load_avg + grq->load.weight
*
* But because grq->load.weight can drop to 0, resulting in a divide by zero,
* we need to use grq->avg.load_avg as its lower bound, which then gives:
*
*
* tg->weight * grq->load.weight
* ge->load.weight = ----------------------------- (6)
* tg_load_avg'
*
* Where:
*
* tg_load_avg' = tg->load_avg - grq->avg.load_avg +
* max(grq->load.weight, grq->avg.load_avg)
*
* And that is shares_weight and is icky. In the (near) UP case it approaches
* (4) while in the normal case it approaches (3). It consistently
* overestimates the ge->load.weight and therefore:
*
* \Sum ge->load.weight >= tg->weight
*
* hence icky!
*/
static long calc_group_shares(struct cfs_rq *cfs_rq)
{
long tg_weight, tg_shares, load, shares;
struct task_group *tg = cfs_rq->tg;
tg_shares = READ_ONCE(tg->shares);
load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
tg_weight = atomic_long_read(&tg->load_avg);
/* Ensure tg_weight >= load */
tg_weight -= cfs_rq->tg_load_avg_contrib;
tg_weight += load;
shares = (tg_shares * load);
if (tg_weight)
shares /= tg_weight;
/*
* MIN_SHARES has to be unscaled here to support per-CPU partitioning
* of a group with small tg->shares value. It is a floor value which is
* assigned as a minimum load.weight to the sched_entity representing
* the group on a CPU.
*
* E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
* on an 8-core system with 8 tasks each runnable on one CPU shares has
* to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
* case no task is runnable on a CPU MIN_SHARES=2 should be returned
* instead of 0.
*/
return clamp_t(long, shares, MIN_SHARES, tg_shares);
}
#endif /* CONFIG_SMP */
static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
/*
* Recomputes the group entity based on the current state of its group
* runqueue.
*/
static void update_cfs_group(struct sched_entity *se)
{
struct cfs_rq *gcfs_rq = group_cfs_rq(se);
long shares;
if (!gcfs_rq)
sched/core: Fix group_entity's share update The update of the share of a cfs_rq is done when its load_avg is updated but before the group_entity's load_avg has been updated for the past time slot. This generates wrong load_avg accounting which can be significant when small tasks are involved in the scheduling. Let take the example of a task a that is dequeued of its task group A: root (cfs_rq) \ (se) A (cfs_rq) \ (se) a Task "a" was the only task in task group A which becomes idle when a is dequeued. We have the sequence: - dequeue_entity a->se - update_load_avg(a->se) - dequeue_entity_load_avg(A->cfs_rq, a->se) - update_cfs_shares(A->cfs_rq) A->cfs_rq->load.weight == 0 A->se->load.weight is updated with the new share (0 in this case) - dequeue_entity A->se - update_load_avg(A->se) but its weight is now null so the last time slot (up to a tick) will be accounted with a weight of 0 instead of its real weight during the time slot. The last time slot will be accounted as an idle one whereas it was a running one. If the running time of task a is short enough that no tick happens when it runs, all running time of group entity A->se will be accounted as idle time. Instead, we should update the share of a cfs_rq (in fact the weight of its group entity) only after having updated the load_avg of the group_entity. update_cfs_shares() now takes the sched_entity as a parameter instead of the cfs_rq, and the weight of the group_entity is updated only once its load_avg has been synced with current time. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: pjt@google.com Link: http://lkml.kernel.org/r/1482335426-7664-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-12-21 15:50:26 +00:00
return;
if (throttled_hierarchy(gcfs_rq))
return;
sched/core: Fix group_entity's share update The update of the share of a cfs_rq is done when its load_avg is updated but before the group_entity's load_avg has been updated for the past time slot. This generates wrong load_avg accounting which can be significant when small tasks are involved in the scheduling. Let take the example of a task a that is dequeued of its task group A: root (cfs_rq) \ (se) A (cfs_rq) \ (se) a Task "a" was the only task in task group A which becomes idle when a is dequeued. We have the sequence: - dequeue_entity a->se - update_load_avg(a->se) - dequeue_entity_load_avg(A->cfs_rq, a->se) - update_cfs_shares(A->cfs_rq) A->cfs_rq->load.weight == 0 A->se->load.weight is updated with the new share (0 in this case) - dequeue_entity A->se - update_load_avg(A->se) but its weight is now null so the last time slot (up to a tick) will be accounted with a weight of 0 instead of its real weight during the time slot. The last time slot will be accounted as an idle one whereas it was a running one. If the running time of task a is short enough that no tick happens when it runs, all running time of group entity A->se will be accounted as idle time. Instead, we should update the share of a cfs_rq (in fact the weight of its group entity) only after having updated the load_avg of the group_entity. update_cfs_shares() now takes the sched_entity as a parameter instead of the cfs_rq, and the weight of the group_entity is updated only once its load_avg has been synced with current time. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: pjt@google.com Link: http://lkml.kernel.org/r/1482335426-7664-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-12-21 15:50:26 +00:00
#ifndef CONFIG_SMP
shares = READ_ONCE(gcfs_rq->tg->shares);
if (likely(se->load.weight == shares))
return;
#else
shares = calc_group_shares(gcfs_rq);
#endif
reweight_entity(cfs_rq_of(se), se, shares);
}
sched/core: Fix group_entity's share update The update of the share of a cfs_rq is done when its load_avg is updated but before the group_entity's load_avg has been updated for the past time slot. This generates wrong load_avg accounting which can be significant when small tasks are involved in the scheduling. Let take the example of a task a that is dequeued of its task group A: root (cfs_rq) \ (se) A (cfs_rq) \ (se) a Task "a" was the only task in task group A which becomes idle when a is dequeued. We have the sequence: - dequeue_entity a->se - update_load_avg(a->se) - dequeue_entity_load_avg(A->cfs_rq, a->se) - update_cfs_shares(A->cfs_rq) A->cfs_rq->load.weight == 0 A->se->load.weight is updated with the new share (0 in this case) - dequeue_entity A->se - update_load_avg(A->se) but its weight is now null so the last time slot (up to a tick) will be accounted with a weight of 0 instead of its real weight during the time slot. The last time slot will be accounted as an idle one whereas it was a running one. If the running time of task a is short enough that no tick happens when it runs, all running time of group entity A->se will be accounted as idle time. Instead, we should update the share of a cfs_rq (in fact the weight of its group entity) only after having updated the load_avg of the group_entity. update_cfs_shares() now takes the sched_entity as a parameter instead of the cfs_rq, and the weight of the group_entity is updated only once its load_avg has been synced with current time. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: pjt@google.com Link: http://lkml.kernel.org/r/1482335426-7664-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-12-21 15:50:26 +00:00
#else /* CONFIG_FAIR_GROUP_SCHED */
static inline void update_cfs_group(struct sched_entity *se)
{
}
#endif /* CONFIG_FAIR_GROUP_SCHED */
static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
{
Power management updates for v4.14-rc1 - Drop the P-state selection algorithm based on a PID controller from intel_pstate and make it use the same P-state selection method (based on the CPU load) for all types of systems in the active mode (Rafael Wysocki, Srinivas Pandruvada). - Rework the cpufreq core and governors to make it possible to take cross-CPU utilization updates into account and modify the schedutil governor to actually do so (Viresh Kumar). - Clean up the handling of transition latency information in the cpufreq core and untangle it from the information on which drivers cannot do dynamic frequency switching (Viresh Kumar). - Add support for new SoCs (MT2701/MT7623 and MT7622) to the mediatek cpufreq driver and update its DT bindings (Sean Wang). - Modify the cpufreq dt-platdev driver to autimatically create cpufreq devices for the new (v2) Operating Performance Points (OPP) DT bindings and update its whitelist of supported systems (Viresh Kumar, Shubhrajyoti Datta, Marc Gonzalez, Khiem Nguyen, Finley Xiao). - Add support for Ux500 to the cpufreq-dt driver and drop the obsolete dbx500 cpufreq driver (Linus Walleij, Arnd Bergmann). - Add new SoC (R8A7795) support to the cpufreq rcar driver (Khiem Nguyen). - Fix and clean up assorted issues in the cpufreq drivers and core (Arvind Yadav, Christophe Jaillet, Colin Ian King, Gustavo Silva, Julia Lawall, Leonard Crestez, Rob Herring, Sudeep Holla). - Update the IO-wait boost handling in the schedutil governor to make it less aggressive (Joel Fernandes). - Rework system suspend diagnostics to make it print fewer messages to the kernel log by default, add a sysfs knob to allow more suspend-related messages to be printed and add Low Power S0 Idle constraints checks to the ACPI suspend-to-idle code (Rafael Wysocki, Srinivas Pandruvada). - Prefer suspend-to-idle over S3 on ACPI-based systems with the ACPI_FADT_LOW_POWER_S0 flag set and the Low Power Idle S0 _DSM interface present in the ACPI tables (Rafael Wysocki). - Update documentation related to system sleep and rename a number of items in the code to make it cleare that they are related to suspend-to-idle (Rafael Wysocki). - Export a variable allowing device drivers to check the target system sleep state from the core system suspend code (Florian Fainelli). - Clean up the cpuidle subsystem to handle the polling state on x86 in a more straightforward way and to use %pOF instead of full_name (Rafael Wysocki, Rob Herring). - Update the devfreq framework to fix and clean up a few minor issues (Chanwoo Choi, Rob Herring). - Extend diagnostics in the generic power domains (genpd) framework and clean it up slightly (Thara Gopinath, Rob Herring). - Fix and clean up a couple of issues in the operating performance points (OPP) framework (Viresh Kumar, Waldemar Rymarkiewicz). - Add support for RV1108 to the rockchip-io Adaptive Voltage Scaling (AVS) driver (David Wu). - Fix the usage of notifiers in CPU power management on some platforms (Alex Shi). - Update the pm-graph system suspend/hibernation and boot profiling utility (Todd Brandt). - Make it possible to run the cpupower utility without CPU0 (Prarit Bhargava). -----BEGIN PGP SIGNATURE----- Version: GnuPG v2 iQIcBAABCAAGBQJZrcDJAAoJEILEb/54YlRx9FUQAIUKvWBAARc61ZIZXjbqZF1v aEMOBuksFns0CMekdptSic6n4wc81E/XYMS8yDhOOMpyDzfAZsTWjmu+gKwN7w3l E/yf/NVlhob9JZ7MqGgqD4EUFfFIaKBXPlWFdDi2rdCUXE2L8xJ7rla8i7zyZlc5 pYHfAppBbF4qUcEY4OoOVOOGRZCfMdiLXj0iZOhMX8Y6yLBRk/AjnVADYsF33hoj gBEfomU+H0K5V8nQEp0ZFKDArPwL+oElHQj6i+nxBpGfPM5evvLXhHOyR6AsldJ5 J4YI1kMuQNSCmvHMqOTxTYyJf8Jcf3Fj4wcjwaVMVGceY1lz6McAKknnFnCqCvz+ mskn84gFCBCM8EoJDqRf0b9MQHcuRyQKM+yw4tjnR9r8yd32erb85ZWFHcPWYhCT fZatNOwFFv2MU+2vo5J3yeUNSWIKT+uBjy+tKPbrDkUwpKZVRj3Oj+hP3Mq9NE8U YBqltsj7tmrdA634zI8C7jfS6wF221S0fId/iPszwmPJaVn/lq8Ror7pWL5YI8U7 SCJFjiqDiGmAcQEkuWwFAQnscZkyHpO+Y3A+jfXl/izoaZETaI5+ceIHBaocm3+5 XrOOpHS3ik8EHf9ji0KFCKZ/pYDwllday3cBQPWo3sMIzpQ2lrjbqdnE1cVnBrld OtHZAeD/jLUXuY6XW2jN =mAiV -----END PGP SIGNATURE----- Merge tag 'pm-4.14-rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/rafael/linux-pm Pull power management updates from Rafael Wysocki: "This time (again) cpufreq gets the majority of changes which mostly are driver updates (including a major consolidation of intel_pstate), some schedutil governor modifications and core cleanups. There also are some changes in the system suspend area, mostly related to diagnostics and debug messages plus some renames of things related to suspend-to-idle. One major change here is that suspend-to-idle is now going to be preferred over S3 on systems where the ACPI tables indicate to do so and provide requsite support (the Low Power Idle S0 _DSM in particular). The system sleep documentation and the tools related to it are updated too. The rest is a few cpuidle changes (nothing major), devfreq updates, generic power domains (genpd) framework updates and a few assorted modifications elsewhere. Specifics: - Drop the P-state selection algorithm based on a PID controller from intel_pstate and make it use the same P-state selection method (based on the CPU load) for all types of systems in the active mode (Rafael Wysocki, Srinivas Pandruvada). - Rework the cpufreq core and governors to make it possible to take cross-CPU utilization updates into account and modify the schedutil governor to actually do so (Viresh Kumar). - Clean up the handling of transition latency information in the cpufreq core and untangle it from the information on which drivers cannot do dynamic frequency switching (Viresh Kumar). - Add support for new SoCs (MT2701/MT7623 and MT7622) to the mediatek cpufreq driver and update its DT bindings (Sean Wang). - Modify the cpufreq dt-platdev driver to autimatically create cpufreq devices for the new (v2) Operating Performance Points (OPP) DT bindings and update its whitelist of supported systems (Viresh Kumar, Shubhrajyoti Datta, Marc Gonzalez, Khiem Nguyen, Finley Xiao). - Add support for Ux500 to the cpufreq-dt driver and drop the obsolete dbx500 cpufreq driver (Linus Walleij, Arnd Bergmann). - Add new SoC (R8A7795) support to the cpufreq rcar driver (Khiem Nguyen). - Fix and clean up assorted issues in the cpufreq drivers and core (Arvind Yadav, Christophe Jaillet, Colin Ian King, Gustavo Silva, Julia Lawall, Leonard Crestez, Rob Herring, Sudeep Holla). - Update the IO-wait boost handling in the schedutil governor to make it less aggressive (Joel Fernandes). - Rework system suspend diagnostics to make it print fewer messages to the kernel log by default, add a sysfs knob to allow more suspend-related messages to be printed and add Low Power S0 Idle constraints checks to the ACPI suspend-to-idle code (Rafael Wysocki, Srinivas Pandruvada). - Prefer suspend-to-idle over S3 on ACPI-based systems with the ACPI_FADT_LOW_POWER_S0 flag set and the Low Power Idle S0 _DSM interface present in the ACPI tables (Rafael Wysocki). - Update documentation related to system sleep and rename a number of items in the code to make it cleare that they are related to suspend-to-idle (Rafael Wysocki). - Export a variable allowing device drivers to check the target system sleep state from the core system suspend code (Florian Fainelli). - Clean up the cpuidle subsystem to handle the polling state on x86 in a more straightforward way and to use %pOF instead of full_name (Rafael Wysocki, Rob Herring). - Update the devfreq framework to fix and clean up a few minor issues (Chanwoo Choi, Rob Herring). - Extend diagnostics in the generic power domains (genpd) framework and clean it up slightly (Thara Gopinath, Rob Herring). - Fix and clean up a couple of issues in the operating performance points (OPP) framework (Viresh Kumar, Waldemar Rymarkiewicz). - Add support for RV1108 to the rockchip-io Adaptive Voltage Scaling (AVS) driver (David Wu). - Fix the usage of notifiers in CPU power management on some platforms (Alex Shi). - Update the pm-graph system suspend/hibernation and boot profiling utility (Todd Brandt). - Make it possible to run the cpupower utility without CPU0 (Prarit Bhargava)" * tag 'pm-4.14-rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/rafael/linux-pm: (87 commits) cpuidle: Make drivers initialize polling state cpuidle: Move polling state initialization code to separate file cpuidle: Eliminate the CPUIDLE_DRIVER_STATE_START symbol cpufreq: imx6q: Fix imx6sx low frequency support cpufreq: speedstep-lib: make several arrays static, makes code smaller PM: docs: Delete the obsolete states.txt document PM: docs: Describe high-level PM strategies and sleep states PM / devfreq: Fix memory leak when fail to register device PM / devfreq: Add dependency on PM_OPP PM / devfreq: Move private devfreq_update_stats() into devfreq PM / devfreq: Convert to using %pOF instead of full_name PM / AVS: rockchip-io: add io selectors and supplies for RV1108 cpufreq: ti: Fix 'of_node_put' being called twice in error handling path cpufreq: dt-platdev: Drop few entries from whitelist cpufreq: dt-platdev: Automatically create cpufreq device with OPP v2 ARM: ux500: don't select CPUFREQ_DT cpuidle: Convert to using %pOF instead of full_name cpufreq: Convert to using %pOF instead of full_name PM / Domains: Convert to using %pOF instead of full_name cpufreq: Cap the default transition delay value to 10 ms ...
2017-09-05 19:19:08 +00:00
struct rq *rq = rq_of(cfs_rq);
if (&rq->cfs == cfs_rq) {
/*
* There are a few boundary cases this might miss but it should
* get called often enough that that should (hopefully) not be
* a real problem.
*
* It will not get called when we go idle, because the idle
* thread is a different class (!fair), nor will the utilization
* number include things like RT tasks.
*
* As is, the util number is not freq-invariant (we'd have to
* implement arch_scale_freq_capacity() for that).
*
* See cpu_util().
*/
cpufreq_update_util(rq, flags);
}
}
#ifdef CONFIG_SMP
#ifdef CONFIG_FAIR_GROUP_SCHED
/**
* update_tg_load_avg - update the tg's load avg
* @cfs_rq: the cfs_rq whose avg changed
* @force: update regardless of how small the difference
*
* This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
* However, because tg->load_avg is a global value there are performance
* considerations.
*
* In order to avoid having to look at the other cfs_rq's, we use a
* differential update where we store the last value we propagated. This in
* turn allows skipping updates if the differential is 'small'.
*
* Updating tg's load_avg is necessary before update_cfs_share().
*/
sched/fair: Rewrite runnable load and utilization average tracking The idea of runnable load average (let runnable time contribute to weight) was proposed by Paul Turner and Ben Segall, and it is still followed by this rewrite. This rewrite aims to solve the following issues: 1. cfs_rq's load average (namely runnable_load_avg and blocked_load_avg) is updated at the granularity of an entity at a time, which results in the cfs_rq's load average is stale or partially updated: at any time, only one entity is up to date, all other entities are effectively lagging behind. This is undesirable. To illustrate, if we have n runnable entities in the cfs_rq, as time elapses, they certainly become outdated: t0: cfs_rq { e1_old, e2_old, ..., en_old } and when we update: t1: update e1, then we have cfs_rq { e1_new, e2_old, ..., en_old } t2: update e2, then we have cfs_rq { e1_old, e2_new, ..., en_old } ... We solve this by combining all runnable entities' load averages together in cfs_rq's avg, and update the cfs_rq's avg as a whole. This is based on the fact that if we regard the update as a function, then: w * update(e) = update(w * e) and update(e1) + update(e2) = update(e1 + e2), then w1 * update(e1) + w2 * update(e2) = update(w1 * e1 + w2 * e2) therefore, by this rewrite, we have an entirely updated cfs_rq at the time we update it: t1: update cfs_rq { e1_new, e2_new, ..., en_new } t2: update cfs_rq { e1_new, e2_new, ..., en_new } ... 2. cfs_rq's load average is different between top rq->cfs_rq and other task_group's per CPU cfs_rqs in whether or not blocked_load_average contributes to the load. The basic idea behind runnable load average (the same for utilization) is that the blocked state is taken into account as opposed to only accounting for the currently runnable state. Therefore, the average should include both the runnable/running and blocked load averages. This rewrite does that. In addition, we also combine runnable/running and blocked averages of all entities into the cfs_rq's average, and update it together at once. This is based on the fact that: update(runnable) + update(blocked) = update(runnable + blocked) This significantly reduces the code as we don't need to separately maintain/update runnable/running load and blocked load. 3. How task_group entities' share is calculated is complex and imprecise. We reduce the complexity in this rewrite to allow a very simple rule: the task_group's load_avg is aggregated from its per CPU cfs_rqs's load_avgs. Then group entity's weight is simply proportional to its own cfs_rq's load_avg / task_group's load_avg. To illustrate, if a task_group has { cfs_rq1, cfs_rq2, ..., cfs_rqn }, then, task_group_avg = cfs_rq1_avg + cfs_rq2_avg + ... + cfs_rqn_avg, then cfs_rqx's entity's share = cfs_rqx_avg / task_group_avg * task_group's share To sum up, this rewrite in principle is equivalent to the current one, but fixes the issues described above. Turns out, it significantly reduces the code complexity and hence increases clarity and efficiency. In addition, the new averages are more smooth/continuous (no spurious spikes and valleys) and updated more consistently and quickly to reflect the load dynamics. As a result, we have less load tracking overhead, better performance, and especially better power efficiency due to more balanced load. Signed-off-by: Yuyang Du <yuyang.du@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: arjan@linux.intel.com Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: fengguang.wu@intel.com Cc: len.brown@intel.com Cc: morten.rasmussen@arm.com Cc: pjt@google.com Cc: rafael.j.wysocki@intel.com Cc: umgwanakikbuti@gmail.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1436918682-4971-3-git-send-email-yuyang.du@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-07-15 00:04:37 +00:00
static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
{
sched/fair: Rewrite runnable load and utilization average tracking The idea of runnable load average (let runnable time contribute to weight) was proposed by Paul Turner and Ben Segall, and it is still followed by this rewrite. This rewrite aims to solve the following issues: 1. cfs_rq's load average (namely runnable_load_avg and blocked_load_avg) is updated at the granularity of an entity at a time, which results in the cfs_rq's load average is stale or partially updated: at any time, only one entity is up to date, all other entities are effectively lagging behind. This is undesirable. To illustrate, if we have n runnable entities in the cfs_rq, as time elapses, they certainly become outdated: t0: cfs_rq { e1_old, e2_old, ..., en_old } and when we update: t1: update e1, then we have cfs_rq { e1_new, e2_old, ..., en_old } t2: update e2, then we have cfs_rq { e1_old, e2_new, ..., en_old } ... We solve this by combining all runnable entities' load averages together in cfs_rq's avg, and update the cfs_rq's avg as a whole. This is based on the fact that if we regard the update as a function, then: w * update(e) = update(w * e) and update(e1) + update(e2) = update(e1 + e2), then w1 * update(e1) + w2 * update(e2) = update(w1 * e1 + w2 * e2) therefore, by this rewrite, we have an entirely updated cfs_rq at the time we update it: t1: update cfs_rq { e1_new, e2_new, ..., en_new } t2: update cfs_rq { e1_new, e2_new, ..., en_new } ... 2. cfs_rq's load average is different between top rq->cfs_rq and other task_group's per CPU cfs_rqs in whether or not blocked_load_average contributes to the load. The basic idea behind runnable load average (the same for utilization) is that the blocked state is taken into account as opposed to only accounting for the currently runnable state. Therefore, the average should include both the runnable/running and blocked load averages. This rewrite does that. In addition, we also combine runnable/running and blocked averages of all entities into the cfs_rq's average, and update it together at once. This is based on the fact that: update(runnable) + update(blocked) = update(runnable + blocked) This significantly reduces the code as we don't need to separately maintain/update runnable/running load and blocked load. 3. How task_group entities' share is calculated is complex and imprecise. We reduce the complexity in this rewrite to allow a very simple rule: the task_group's load_avg is aggregated from its per CPU cfs_rqs's load_avgs. Then group entity's weight is simply proportional to its own cfs_rq's load_avg / task_group's load_avg. To illustrate, if a task_group has { cfs_rq1, cfs_rq2, ..., cfs_rqn }, then, task_group_avg = cfs_rq1_avg + cfs_rq2_avg + ... + cfs_rqn_avg, then cfs_rqx's entity's share = cfs_rqx_avg / task_group_avg * task_group's share To sum up, this rewrite in principle is equivalent to the current one, but fixes the issues described above. Turns out, it significantly reduces the code complexity and hence increases clarity and efficiency. In addition, the new averages are more smooth/continuous (no spurious spikes and valleys) and updated more consistently and quickly to reflect the load dynamics. As a result, we have less load tracking overhead, better performance, and especially better power efficiency due to more balanced load. Signed-off-by: Yuyang Du <yuyang.du@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: arjan@linux.intel.com Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: fengguang.wu@intel.com Cc: len.brown@intel.com Cc: morten.rasmussen@arm.com Cc: pjt@google.com Cc: rafael.j.wysocki@intel.com Cc: umgwanakikbuti@gmail.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1436918682-4971-3-git-send-email-yuyang.du@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-07-15 00:04:37 +00:00
long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
/*
* No need to update load_avg for root_task_group as it is not used.
*/
if (cfs_rq->tg == &root_task_group)
return;
sched/fair: Rewrite runnable load and utilization average tracking The idea of runnable load average (let runnable time contribute to weight) was proposed by Paul Turner and Ben Segall, and it is still followed by this rewrite. This rewrite aims to solve the following issues: 1. cfs_rq's load average (namely runnable_load_avg and blocked_load_avg) is updated at the granularity of an entity at a time, which results in the cfs_rq's load average is stale or partially updated: at any time, only one entity is up to date, all other entities are effectively lagging behind. This is undesirable. To illustrate, if we have n runnable entities in the cfs_rq, as time elapses, they certainly become outdated: t0: cfs_rq { e1_old, e2_old, ..., en_old } and when we update: t1: update e1, then we have cfs_rq { e1_new, e2_old, ..., en_old } t2: update e2, then we have cfs_rq { e1_old, e2_new, ..., en_old } ... We solve this by combining all runnable entities' load averages together in cfs_rq's avg, and update the cfs_rq's avg as a whole. This is based on the fact that if we regard the update as a function, then: w * update(e) = update(w * e) and update(e1) + update(e2) = update(e1 + e2), then w1 * update(e1) + w2 * update(e2) = update(w1 * e1 + w2 * e2) therefore, by this rewrite, we have an entirely updated cfs_rq at the time we update it: t1: update cfs_rq { e1_new, e2_new, ..., en_new } t2: update cfs_rq { e1_new, e2_new, ..., en_new } ... 2. cfs_rq's load average is different between top rq->cfs_rq and other task_group's per CPU cfs_rqs in whether or not blocked_load_average contributes to the load. The basic idea behind runnable load average (the same for utilization) is that the blocked state is taken into account as opposed to only accounting for the currently runnable state. Therefore, the average should include both the runnable/running and blocked load averages. This rewrite does that. In addition, we also combine runnable/running and blocked averages of all entities into the cfs_rq's average, and update it together at once. This is based on the fact that: update(runnable) + update(blocked) = update(runnable + blocked) This significantly reduces the code as we don't need to separately maintain/update runnable/running load and blocked load. 3. How task_group entities' share is calculated is complex and imprecise. We reduce the complexity in this rewrite to allow a very simple rule: the task_group's load_avg is aggregated from its per CPU cfs_rqs's load_avgs. Then group entity's weight is simply proportional to its own cfs_rq's load_avg / task_group's load_avg. To illustrate, if a task_group has { cfs_rq1, cfs_rq2, ..., cfs_rqn }, then, task_group_avg = cfs_rq1_avg + cfs_rq2_avg + ... + cfs_rqn_avg, then cfs_rqx's entity's share = cfs_rqx_avg / task_group_avg * task_group's share To sum up, this rewrite in principle is equivalent to the current one, but fixes the issues described above. Turns out, it significantly reduces the code complexity and hence increases clarity and efficiency. In addition, the new averages are more smooth/continuous (no spurious spikes and valleys) and updated more consistently and quickly to reflect the load dynamics. As a result, we have less load tracking overhead, better performance, and especially better power efficiency due to more balanced load. Signed-off-by: Yuyang Du <yuyang.du@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: arjan@linux.intel.com Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: fengguang.wu@intel.com Cc: len.brown@intel.com Cc: morten.rasmussen@arm.com Cc: pjt@google.com Cc: rafael.j.wysocki@intel.com Cc: umgwanakikbuti@gmail.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1436918682-4971-3-git-send-email-yuyang.du@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-07-15 00:04:37 +00:00
if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
atomic_long_add(delta, &cfs_rq->tg->load_avg);
cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
}
}
/*
* Called within set_task_rq() right before setting a task's CPU. The
* caller only guarantees p->pi_lock is held; no other assumptions,
* including the state of rq->lock, should be made.
*/
void set_task_rq_fair(struct sched_entity *se,
struct cfs_rq *prev, struct cfs_rq *next)
{
u64 p_last_update_time;
u64 n_last_update_time;
if (!sched_feat(ATTACH_AGE_LOAD))
return;
/*
* We are supposed to update the task to "current" time, then its up to
* date and ready to go to new CPU/cfs_rq. But we have difficulty in
* getting what current time is, so simply throw away the out-of-date
* time. This will result in the wakee task is less decayed, but giving
* the wakee more load sounds not bad.
*/
if (!(se->avg.last_update_time && prev))
return;
#ifndef CONFIG_64BIT
{
u64 p_last_update_time_copy;
u64 n_last_update_time_copy;
do {
p_last_update_time_copy = prev->load_last_update_time_copy;
n_last_update_time_copy = next->load_last_update_time_copy;
smp_rmb();
p_last_update_time = prev->avg.last_update_time;
n_last_update_time = next->avg.last_update_time;
} while (p_last_update_time != p_last_update_time_copy ||
n_last_update_time != n_last_update_time_copy);
}
#else
p_last_update_time = prev->avg.last_update_time;
n_last_update_time = next->avg.last_update_time;
#endif
sched/fair: Update scale invariance of PELT The current implementation of load tracking invariance scales the contribution with current frequency and uarch performance (only for utilization) of the CPU. One main result of this formula is that the figures are capped by current capacity of CPU. Another one is that the load_avg is not invariant because not scaled with uarch. The util_avg of a periodic task that runs r time slots every p time slots varies in the range : U * (1-y^r)/(1-y^p) * y^i < Utilization < U * (1-y^r)/(1-y^p) with U is the max util_avg value = SCHED_CAPACITY_SCALE At a lower capacity, the range becomes: U * C * (1-y^r')/(1-y^p) * y^i' < Utilization < U * C * (1-y^r')/(1-y^p) with C reflecting the compute capacity ratio between current capacity and max capacity. so C tries to compensate changes in (1-y^r') but it can't be accurate. Instead of scaling the contribution value of PELT algo, we should scale the running time. The PELT signal aims to track the amount of computation of tasks and/or rq so it seems more correct to scale the running time to reflect the effective amount of computation done since the last update. In order to be fully invariant, we need to apply the same amount of running time and idle time whatever the current capacity. Because running at lower capacity implies that the task will run longer, we have to ensure that the same amount of idle time will be applied when system becomes idle and no idle time has been "stolen". But reaching the maximum utilization value (SCHED_CAPACITY_SCALE) means that the task is seen as an always-running task whatever the capacity of the CPU (even at max compute capacity). In this case, we can discard this "stolen" idle times which becomes meaningless. In order to achieve this time scaling, a new clock_pelt is created per rq. The increase of this clock scales with current capacity when something is running on rq and synchronizes with clock_task when rq is idle. With this mechanism, we ensure the same running and idle time whatever the current capacity. This also enables to simplify the pelt algorithm by removing all references of uarch and frequency and applying the same contribution to utilization and loads. Furthermore, the scaling is done only once per update of clock (update_rq_clock_task()) instead of during each update of sched_entities and cfs/rt/dl_rq of the rq like the current implementation. This is interesting when cgroup are involved as shown in the results below: On a hikey (octo Arm64 platform). Performance cpufreq governor and only shallowest c-state to remove variance generated by those power features so we only track the impact of pelt algo. each test runs 16 times: ./perf bench sched pipe (higher is better) kernel tip/sched/core + patch ops/seconds ops/seconds diff cgroup root 59652(+/- 0.18%) 59876(+/- 0.24%) +0.38% level1 55608(+/- 0.27%) 55923(+/- 0.24%) +0.57% level2 52115(+/- 0.29%) 52564(+/- 0.22%) +0.86% hackbench -l 1000 (lower is better) kernel tip/sched/core + patch duration(sec) duration(sec) diff cgroup root 4.453(+/- 2.37%) 4.383(+/- 2.88%) -1.57% level1 4.859(+/- 8.50%) 4.830(+/- 7.07%) -0.60% level2 5.063(+/- 9.83%) 4.928(+/- 9.66%) -2.66% Then, the responsiveness of PELT is improved when CPU is not running at max capacity with this new algorithm. I have put below some examples of duration to reach some typical load values according to the capacity of the CPU with current implementation and with this patch. These values has been computed based on the geometric series and the half period value: Util (%) max capacity half capacity(mainline) half capacity(w/ patch) 972 (95%) 138ms not reachable 276ms 486 (47.5%) 30ms 138ms 60ms 256 (25%) 13ms 32ms 26ms On my hikey (octo Arm64 platform) with schedutil governor, the time to reach max OPP when starting from a null utilization, decreases from 223ms with current scale invariance down to 121ms with the new algorithm. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: patrick.bellasi@arm.com Cc: pjt@google.com Cc: pkondeti@codeaurora.org Cc: quentin.perret@arm.com Cc: rjw@rjwysocki.net Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Link: https://lkml.kernel.org/r/1548257214-13745-3-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-01-23 15:26:53 +00:00
__update_load_avg_blocked_se(p_last_update_time, se);
se->avg.last_update_time = n_last_update_time;
}
/*
* When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
* propagate its contribution. The key to this propagation is the invariant
* that for each group:
*
* ge->avg == grq->avg (1)
*
* _IFF_ we look at the pure running and runnable sums. Because they
* represent the very same entity, just at different points in the hierarchy.
*
sched/pelt: Add a new runnable average signal Now that runnable_load_avg has been removed, we can replace it by a new signal that will highlight the runnable pressure on a cfs_rq. This signal track the waiting time of tasks on rq and can help to better define the state of rqs. At now, only util_avg is used to define the state of a rq: A rq with more that around 80% of utilization and more than 1 tasks is considered as overloaded. But the util_avg signal of a rq can become temporaly low after that a task migrated onto another rq which can bias the classification of the rq. When tasks compete for the same rq, their runnable average signal will be higher than util_avg as it will include the waiting time and we can use this signal to better classify cfs_rqs. The new runnable_avg will track the runnable time of a task which simply adds the waiting time to the running time. The runnable _avg of cfs_rq will be the /Sum of se's runnable_avg and the runnable_avg of group entity will follow the one of the rq similarly to util_avg. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: "Dietmar Eggemann <dietmar.eggemann@arm.com>" Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Valentin Schneider <valentin.schneider@arm.com> Cc: Phil Auld <pauld@redhat.com> Cc: Hillf Danton <hdanton@sina.com> Link: https://lore.kernel.org/r/20200224095223.13361-9-mgorman@techsingularity.net
2020-02-24 09:52:18 +00:00
* Per the above update_tg_cfs_util() and update_tg_cfs_runnable() are trivial
* and simply copies the running/runnable sum over (but still wrong, because
* the group entity and group rq do not have their PELT windows aligned).
*
* However, update_tg_cfs_load() is more complex. So we have:
*
* ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
*
* And since, like util, the runnable part should be directly transferable,
* the following would _appear_ to be the straight forward approach:
*
sched/fair: Update and fix the runnable propagation rule Unlike running, the runnable part can't be directly propagated through the hierarchy when we migrate a task. The main reason is that runnable time can be shared with other sched_entities that stay on the rq and this runnable time will also remain on prev cfs_rq and must not be removed. Instead, we can estimate what should be the new runnable of the prev cfs_rq and check that this estimation stay in a possible range. The prop_runnable_sum is a good estimation when adding runnable_sum but fails most often when we remove it. Instead, we could use the formula below instead: gcfs_rq's runnable_sum = gcfs_rq->avg.load_sum / gcfs_rq->load.weight which assumes that tasks are equally runnable which is not true but easy to compute. Beside these estimates, we have several simple rules that help us to filter out wrong ones: - ge->avg.runnable_sum <= than LOAD_AVG_MAX - ge->avg.runnable_sum >= ge->avg.running_sum (ge->avg.util_sum << LOAD_AVG_MAX) - ge->avg.runnable_sum can't increase when we detach a task The effect of these fixes is better cgroups balancing. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Ben Segall <bsegall@google.com> Cc: Chris Mason <clm@fb.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Tejun Heo <tj@kernel.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Yuyang Du <yuyang.du@intel.com> Link: http://lkml.kernel.org/r/1510842112-21028-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-11-16 14:21:52 +00:00
* grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
*
* And per (1) we have:
*
sched/fair: Update and fix the runnable propagation rule Unlike running, the runnable part can't be directly propagated through the hierarchy when we migrate a task. The main reason is that runnable time can be shared with other sched_entities that stay on the rq and this runnable time will also remain on prev cfs_rq and must not be removed. Instead, we can estimate what should be the new runnable of the prev cfs_rq and check that this estimation stay in a possible range. The prop_runnable_sum is a good estimation when adding runnable_sum but fails most often when we remove it. Instead, we could use the formula below instead: gcfs_rq's runnable_sum = gcfs_rq->avg.load_sum / gcfs_rq->load.weight which assumes that tasks are equally runnable which is not true but easy to compute. Beside these estimates, we have several simple rules that help us to filter out wrong ones: - ge->avg.runnable_sum <= than LOAD_AVG_MAX - ge->avg.runnable_sum >= ge->avg.running_sum (ge->avg.util_sum << LOAD_AVG_MAX) - ge->avg.runnable_sum can't increase when we detach a task The effect of these fixes is better cgroups balancing. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Ben Segall <bsegall@google.com> Cc: Chris Mason <clm@fb.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Tejun Heo <tj@kernel.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Yuyang Du <yuyang.du@intel.com> Link: http://lkml.kernel.org/r/1510842112-21028-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-11-16 14:21:52 +00:00
* ge->avg.runnable_avg == grq->avg.runnable_avg
*
* Which gives:
*
* ge->load.weight * grq->avg.load_avg
* ge->avg.load_avg = ----------------------------------- (4)
* grq->load.weight
*
* Except that is wrong!
*
* Because while for entities historical weight is not important and we
* really only care about our future and therefore can consider a pure
* runnable sum, runqueues can NOT do this.
*
* We specifically want runqueues to have a load_avg that includes
* historical weights. Those represent the blocked load, the load we expect
* to (shortly) return to us. This only works by keeping the weights as
* integral part of the sum. We therefore cannot decompose as per (3).
*
sched/fair: Update and fix the runnable propagation rule Unlike running, the runnable part can't be directly propagated through the hierarchy when we migrate a task. The main reason is that runnable time can be shared with other sched_entities that stay on the rq and this runnable time will also remain on prev cfs_rq and must not be removed. Instead, we can estimate what should be the new runnable of the prev cfs_rq and check that this estimation stay in a possible range. The prop_runnable_sum is a good estimation when adding runnable_sum but fails most often when we remove it. Instead, we could use the formula below instead: gcfs_rq's runnable_sum = gcfs_rq->avg.load_sum / gcfs_rq->load.weight which assumes that tasks are equally runnable which is not true but easy to compute. Beside these estimates, we have several simple rules that help us to filter out wrong ones: - ge->avg.runnable_sum <= than LOAD_AVG_MAX - ge->avg.runnable_sum >= ge->avg.running_sum (ge->avg.util_sum << LOAD_AVG_MAX) - ge->avg.runnable_sum can't increase when we detach a task The effect of these fixes is better cgroups balancing. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Ben Segall <bsegall@google.com> Cc: Chris Mason <clm@fb.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Tejun Heo <tj@kernel.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Yuyang Du <yuyang.du@intel.com> Link: http://lkml.kernel.org/r/1510842112-21028-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-11-16 14:21:52 +00:00
* Another reason this doesn't work is that runnable isn't a 0-sum entity.
* Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
* rq itself is runnable anywhere between 2/3 and 1 depending on how the
* runnable section of these tasks overlap (or not). If they were to perfectly
* align the rq as a whole would be runnable 2/3 of the time. If however we
* always have at least 1 runnable task, the rq as a whole is always runnable.
*
sched/fair: Update and fix the runnable propagation rule Unlike running, the runnable part can't be directly propagated through the hierarchy when we migrate a task. The main reason is that runnable time can be shared with other sched_entities that stay on the rq and this runnable time will also remain on prev cfs_rq and must not be removed. Instead, we can estimate what should be the new runnable of the prev cfs_rq and check that this estimation stay in a possible range. The prop_runnable_sum is a good estimation when adding runnable_sum but fails most often when we remove it. Instead, we could use the formula below instead: gcfs_rq's runnable_sum = gcfs_rq->avg.load_sum / gcfs_rq->load.weight which assumes that tasks are equally runnable which is not true but easy to compute. Beside these estimates, we have several simple rules that help us to filter out wrong ones: - ge->avg.runnable_sum <= than LOAD_AVG_MAX - ge->avg.runnable_sum >= ge->avg.running_sum (ge->avg.util_sum << LOAD_AVG_MAX) - ge->avg.runnable_sum can't increase when we detach a task The effect of these fixes is better cgroups balancing. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Ben Segall <bsegall@google.com> Cc: Chris Mason <clm@fb.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Tejun Heo <tj@kernel.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Yuyang Du <yuyang.du@intel.com> Link: http://lkml.kernel.org/r/1510842112-21028-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-11-16 14:21:52 +00:00
* So we'll have to approximate.. :/
*
sched/fair: Update and fix the runnable propagation rule Unlike running, the runnable part can't be directly propagated through the hierarchy when we migrate a task. The main reason is that runnable time can be shared with other sched_entities that stay on the rq and this runnable time will also remain on prev cfs_rq and must not be removed. Instead, we can estimate what should be the new runnable of the prev cfs_rq and check that this estimation stay in a possible range. The prop_runnable_sum is a good estimation when adding runnable_sum but fails most often when we remove it. Instead, we could use the formula below instead: gcfs_rq's runnable_sum = gcfs_rq->avg.load_sum / gcfs_rq->load.weight which assumes that tasks are equally runnable which is not true but easy to compute. Beside these estimates, we have several simple rules that help us to filter out wrong ones: - ge->avg.runnable_sum <= than LOAD_AVG_MAX - ge->avg.runnable_sum >= ge->avg.running_sum (ge->avg.util_sum << LOAD_AVG_MAX) - ge->avg.runnable_sum can't increase when we detach a task The effect of these fixes is better cgroups balancing. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Ben Segall <bsegall@google.com> Cc: Chris Mason <clm@fb.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Tejun Heo <tj@kernel.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Yuyang Du <yuyang.du@intel.com> Link: http://lkml.kernel.org/r/1510842112-21028-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-11-16 14:21:52 +00:00
* Given the constraint:
*
sched/fair: Update and fix the runnable propagation rule Unlike running, the runnable part can't be directly propagated through the hierarchy when we migrate a task. The main reason is that runnable time can be shared with other sched_entities that stay on the rq and this runnable time will also remain on prev cfs_rq and must not be removed. Instead, we can estimate what should be the new runnable of the prev cfs_rq and check that this estimation stay in a possible range. The prop_runnable_sum is a good estimation when adding runnable_sum but fails most often when we remove it. Instead, we could use the formula below instead: gcfs_rq's runnable_sum = gcfs_rq->avg.load_sum / gcfs_rq->load.weight which assumes that tasks are equally runnable which is not true but easy to compute. Beside these estimates, we have several simple rules that help us to filter out wrong ones: - ge->avg.runnable_sum <= than LOAD_AVG_MAX - ge->avg.runnable_sum >= ge->avg.running_sum (ge->avg.util_sum << LOAD_AVG_MAX) - ge->avg.runnable_sum can't increase when we detach a task The effect of these fixes is better cgroups balancing. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Ben Segall <bsegall@google.com> Cc: Chris Mason <clm@fb.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Tejun Heo <tj@kernel.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Yuyang Du <yuyang.du@intel.com> Link: http://lkml.kernel.org/r/1510842112-21028-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-11-16 14:21:52 +00:00
* ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
*
sched/fair: Update and fix the runnable propagation rule Unlike running, the runnable part can't be directly propagated through the hierarchy when we migrate a task. The main reason is that runnable time can be shared with other sched_entities that stay on the rq and this runnable time will also remain on prev cfs_rq and must not be removed. Instead, we can estimate what should be the new runnable of the prev cfs_rq and check that this estimation stay in a possible range. The prop_runnable_sum is a good estimation when adding runnable_sum but fails most often when we remove it. Instead, we could use the formula below instead: gcfs_rq's runnable_sum = gcfs_rq->avg.load_sum / gcfs_rq->load.weight which assumes that tasks are equally runnable which is not true but easy to compute. Beside these estimates, we have several simple rules that help us to filter out wrong ones: - ge->avg.runnable_sum <= than LOAD_AVG_MAX - ge->avg.runnable_sum >= ge->avg.running_sum (ge->avg.util_sum << LOAD_AVG_MAX) - ge->avg.runnable_sum can't increase when we detach a task The effect of these fixes is better cgroups balancing. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Ben Segall <bsegall@google.com> Cc: Chris Mason <clm@fb.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Tejun Heo <tj@kernel.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Yuyang Du <yuyang.du@intel.com> Link: http://lkml.kernel.org/r/1510842112-21028-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-11-16 14:21:52 +00:00
* We can construct a rule that adds runnable to a rq by assuming minimal
* overlap.
*
sched/fair: Update and fix the runnable propagation rule Unlike running, the runnable part can't be directly propagated through the hierarchy when we migrate a task. The main reason is that runnable time can be shared with other sched_entities that stay on the rq and this runnable time will also remain on prev cfs_rq and must not be removed. Instead, we can estimate what should be the new runnable of the prev cfs_rq and check that this estimation stay in a possible range. The prop_runnable_sum is a good estimation when adding runnable_sum but fails most often when we remove it. Instead, we could use the formula below instead: gcfs_rq's runnable_sum = gcfs_rq->avg.load_sum / gcfs_rq->load.weight which assumes that tasks are equally runnable which is not true but easy to compute. Beside these estimates, we have several simple rules that help us to filter out wrong ones: - ge->avg.runnable_sum <= than LOAD_AVG_MAX - ge->avg.runnable_sum >= ge->avg.running_sum (ge->avg.util_sum << LOAD_AVG_MAX) - ge->avg.runnable_sum can't increase when we detach a task The effect of these fixes is better cgroups balancing. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Ben Segall <bsegall@google.com> Cc: Chris Mason <clm@fb.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Tejun Heo <tj@kernel.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Yuyang Du <yuyang.du@intel.com> Link: http://lkml.kernel.org/r/1510842112-21028-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-11-16 14:21:52 +00:00
* On removal, we'll assume each task is equally runnable; which yields:
*
sched/fair: Update and fix the runnable propagation rule Unlike running, the runnable part can't be directly propagated through the hierarchy when we migrate a task. The main reason is that runnable time can be shared with other sched_entities that stay on the rq and this runnable time will also remain on prev cfs_rq and must not be removed. Instead, we can estimate what should be the new runnable of the prev cfs_rq and check that this estimation stay in a possible range. The prop_runnable_sum is a good estimation when adding runnable_sum but fails most often when we remove it. Instead, we could use the formula below instead: gcfs_rq's runnable_sum = gcfs_rq->avg.load_sum / gcfs_rq->load.weight which assumes that tasks are equally runnable which is not true but easy to compute. Beside these estimates, we have several simple rules that help us to filter out wrong ones: - ge->avg.runnable_sum <= than LOAD_AVG_MAX - ge->avg.runnable_sum >= ge->avg.running_sum (ge->avg.util_sum << LOAD_AVG_MAX) - ge->avg.runnable_sum can't increase when we detach a task The effect of these fixes is better cgroups balancing. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Ben Segall <bsegall@google.com> Cc: Chris Mason <clm@fb.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Tejun Heo <tj@kernel.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Yuyang Du <yuyang.du@intel.com> Link: http://lkml.kernel.org/r/1510842112-21028-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-11-16 14:21:52 +00:00
* grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
*
sched/fair: Update and fix the runnable propagation rule Unlike running, the runnable part can't be directly propagated through the hierarchy when we migrate a task. The main reason is that runnable time can be shared with other sched_entities that stay on the rq and this runnable time will also remain on prev cfs_rq and must not be removed. Instead, we can estimate what should be the new runnable of the prev cfs_rq and check that this estimation stay in a possible range. The prop_runnable_sum is a good estimation when adding runnable_sum but fails most often when we remove it. Instead, we could use the formula below instead: gcfs_rq's runnable_sum = gcfs_rq->avg.load_sum / gcfs_rq->load.weight which assumes that tasks are equally runnable which is not true but easy to compute. Beside these estimates, we have several simple rules that help us to filter out wrong ones: - ge->avg.runnable_sum <= than LOAD_AVG_MAX - ge->avg.runnable_sum >= ge->avg.running_sum (ge->avg.util_sum << LOAD_AVG_MAX) - ge->avg.runnable_sum can't increase when we detach a task The effect of these fixes is better cgroups balancing. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Ben Segall <bsegall@google.com> Cc: Chris Mason <clm@fb.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Tejun Heo <tj@kernel.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Yuyang Du <yuyang.du@intel.com> Link: http://lkml.kernel.org/r/1510842112-21028-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-11-16 14:21:52 +00:00
* XXX: only do this for the part of runnable > running ?
*
*/
static inline void
update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
{
long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
/* Nothing to update */
if (!delta)
return;
sched/fair: Update and fix the runnable propagation rule Unlike running, the runnable part can't be directly propagated through the hierarchy when we migrate a task. The main reason is that runnable time can be shared with other sched_entities that stay on the rq and this runnable time will also remain on prev cfs_rq and must not be removed. Instead, we can estimate what should be the new runnable of the prev cfs_rq and check that this estimation stay in a possible range. The prop_runnable_sum is a good estimation when adding runnable_sum but fails most often when we remove it. Instead, we could use the formula below instead: gcfs_rq's runnable_sum = gcfs_rq->avg.load_sum / gcfs_rq->load.weight which assumes that tasks are equally runnable which is not true but easy to compute. Beside these estimates, we have several simple rules that help us to filter out wrong ones: - ge->avg.runnable_sum <= than LOAD_AVG_MAX - ge->avg.runnable_sum >= ge->avg.running_sum (ge->avg.util_sum << LOAD_AVG_MAX) - ge->avg.runnable_sum can't increase when we detach a task The effect of these fixes is better cgroups balancing. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Ben Segall <bsegall@google.com> Cc: Chris Mason <clm@fb.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Tejun Heo <tj@kernel.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Yuyang Du <yuyang.du@intel.com> Link: http://lkml.kernel.org/r/1510842112-21028-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-11-16 14:21:52 +00:00
/*
* The relation between sum and avg is:
*
* LOAD_AVG_MAX - 1024 + sa->period_contrib
*
* however, the PELT windows are not aligned between grq and gse.
*/
/* Set new sched_entity's utilization */
se->avg.util_avg = gcfs_rq->avg.util_avg;
se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
/* Update parent cfs_rq utilization */
add_positive(&cfs_rq->avg.util_avg, delta);
cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
}
sched/pelt: Add a new runnable average signal Now that runnable_load_avg has been removed, we can replace it by a new signal that will highlight the runnable pressure on a cfs_rq. This signal track the waiting time of tasks on rq and can help to better define the state of rqs. At now, only util_avg is used to define the state of a rq: A rq with more that around 80% of utilization and more than 1 tasks is considered as overloaded. But the util_avg signal of a rq can become temporaly low after that a task migrated onto another rq which can bias the classification of the rq. When tasks compete for the same rq, their runnable average signal will be higher than util_avg as it will include the waiting time and we can use this signal to better classify cfs_rqs. The new runnable_avg will track the runnable time of a task which simply adds the waiting time to the running time. The runnable _avg of cfs_rq will be the /Sum of se's runnable_avg and the runnable_avg of group entity will follow the one of the rq similarly to util_avg. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: "Dietmar Eggemann <dietmar.eggemann@arm.com>" Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Valentin Schneider <valentin.schneider@arm.com> Cc: Phil Auld <pauld@redhat.com> Cc: Hillf Danton <hdanton@sina.com> Link: https://lore.kernel.org/r/20200224095223.13361-9-mgorman@techsingularity.net
2020-02-24 09:52:18 +00:00
static inline void
update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
{
long delta = gcfs_rq->avg.runnable_avg - se->avg.runnable_avg;
/* Nothing to update */
if (!delta)
return;
/*
* The relation between sum and avg is:
*
* LOAD_AVG_MAX - 1024 + sa->period_contrib
*
* however, the PELT windows are not aligned between grq and gse.
*/
/* Set new sched_entity's runnable */
se->avg.runnable_avg = gcfs_rq->avg.runnable_avg;
se->avg.runnable_sum = se->avg.runnable_avg * LOAD_AVG_MAX;
/* Update parent cfs_rq runnable */
add_positive(&cfs_rq->avg.runnable_avg, delta);
cfs_rq->avg.runnable_sum = cfs_rq->avg.runnable_avg * LOAD_AVG_MAX;
}
static inline void
update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
{
sched/fair: Update and fix the runnable propagation rule Unlike running, the runnable part can't be directly propagated through the hierarchy when we migrate a task. The main reason is that runnable time can be shared with other sched_entities that stay on the rq and this runnable time will also remain on prev cfs_rq and must not be removed. Instead, we can estimate what should be the new runnable of the prev cfs_rq and check that this estimation stay in a possible range. The prop_runnable_sum is a good estimation when adding runnable_sum but fails most often when we remove it. Instead, we could use the formula below instead: gcfs_rq's runnable_sum = gcfs_rq->avg.load_sum / gcfs_rq->load.weight which assumes that tasks are equally runnable which is not true but easy to compute. Beside these estimates, we have several simple rules that help us to filter out wrong ones: - ge->avg.runnable_sum <= than LOAD_AVG_MAX - ge->avg.runnable_sum >= ge->avg.running_sum (ge->avg.util_sum << LOAD_AVG_MAX) - ge->avg.runnable_sum can't increase when we detach a task The effect of these fixes is better cgroups balancing. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Ben Segall <bsegall@google.com> Cc: Chris Mason <clm@fb.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Tejun Heo <tj@kernel.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Yuyang Du <yuyang.du@intel.com> Link: http://lkml.kernel.org/r/1510842112-21028-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-11-16 14:21:52 +00:00
long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
unsigned long load_avg;
u64 load_sum = 0;
sched/fair: Update and fix the runnable propagation rule Unlike running, the runnable part can't be directly propagated through the hierarchy when we migrate a task. The main reason is that runnable time can be shared with other sched_entities that stay on the rq and this runnable time will also remain on prev cfs_rq and must not be removed. Instead, we can estimate what should be the new runnable of the prev cfs_rq and check that this estimation stay in a possible range. The prop_runnable_sum is a good estimation when adding runnable_sum but fails most often when we remove it. Instead, we could use the formula below instead: gcfs_rq's runnable_sum = gcfs_rq->avg.load_sum / gcfs_rq->load.weight which assumes that tasks are equally runnable which is not true but easy to compute. Beside these estimates, we have several simple rules that help us to filter out wrong ones: - ge->avg.runnable_sum <= than LOAD_AVG_MAX - ge->avg.runnable_sum >= ge->avg.running_sum (ge->avg.util_sum << LOAD_AVG_MAX) - ge->avg.runnable_sum can't increase when we detach a task The effect of these fixes is better cgroups balancing. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Ben Segall <bsegall@google.com> Cc: Chris Mason <clm@fb.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Tejun Heo <tj@kernel.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Yuyang Du <yuyang.du@intel.com> Link: http://lkml.kernel.org/r/1510842112-21028-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-11-16 14:21:52 +00:00
s64 delta_sum;
if (!runnable_sum)
return;
gcfs_rq->prop_runnable_sum = 0;
sched/fair: Update and fix the runnable propagation rule Unlike running, the runnable part can't be directly propagated through the hierarchy when we migrate a task. The main reason is that runnable time can be shared with other sched_entities that stay on the rq and this runnable time will also remain on prev cfs_rq and must not be removed. Instead, we can estimate what should be the new runnable of the prev cfs_rq and check that this estimation stay in a possible range. The prop_runnable_sum is a good estimation when adding runnable_sum but fails most often when we remove it. Instead, we could use the formula below instead: gcfs_rq's runnable_sum = gcfs_rq->avg.load_sum / gcfs_rq->load.weight which assumes that tasks are equally runnable which is not true but easy to compute. Beside these estimates, we have several simple rules that help us to filter out wrong ones: - ge->avg.runnable_sum <= than LOAD_AVG_MAX - ge->avg.runnable_sum >= ge->avg.running_sum (ge->avg.util_sum << LOAD_AVG_MAX) - ge->avg.runnable_sum can't increase when we detach a task The effect of these fixes is better cgroups balancing. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Ben Segall <bsegall@google.com> Cc: Chris Mason <clm@fb.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Tejun Heo <tj@kernel.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Yuyang Du <yuyang.du@intel.com> Link: http://lkml.kernel.org/r/1510842112-21028-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-11-16 14:21:52 +00:00
if (runnable_sum >= 0) {
/*
* Add runnable; clip at LOAD_AVG_MAX. Reflects that until
* the CPU is saturated running == runnable.
*/
runnable_sum += se->avg.load_sum;
runnable_sum = min(runnable_sum, (long)LOAD_AVG_MAX);
} else {
/*
* Estimate the new unweighted runnable_sum of the gcfs_rq by
* assuming all tasks are equally runnable.
*/
if (scale_load_down(gcfs_rq->load.weight)) {
load_sum = div_s64(gcfs_rq->avg.load_sum,
scale_load_down(gcfs_rq->load.weight));
}
/* But make sure to not inflate se's runnable */
runnable_sum = min(se->avg.load_sum, load_sum);
}
/*
* runnable_sum can't be lower than running_sum
sched/fair: Update scale invariance of PELT The current implementation of load tracking invariance scales the contribution with current frequency and uarch performance (only for utilization) of the CPU. One main result of this formula is that the figures are capped by current capacity of CPU. Another one is that the load_avg is not invariant because not scaled with uarch. The util_avg of a periodic task that runs r time slots every p time slots varies in the range : U * (1-y^r)/(1-y^p) * y^i < Utilization < U * (1-y^r)/(1-y^p) with U is the max util_avg value = SCHED_CAPACITY_SCALE At a lower capacity, the range becomes: U * C * (1-y^r')/(1-y^p) * y^i' < Utilization < U * C * (1-y^r')/(1-y^p) with C reflecting the compute capacity ratio between current capacity and max capacity. so C tries to compensate changes in (1-y^r') but it can't be accurate. Instead of scaling the contribution value of PELT algo, we should scale the running time. The PELT signal aims to track the amount of computation of tasks and/or rq so it seems more correct to scale the running time to reflect the effective amount of computation done since the last update. In order to be fully invariant, we need to apply the same amount of running time and idle time whatever the current capacity. Because running at lower capacity implies that the task will run longer, we have to ensure that the same amount of idle time will be applied when system becomes idle and no idle time has been "stolen". But reaching the maximum utilization value (SCHED_CAPACITY_SCALE) means that the task is seen as an always-running task whatever the capacity of the CPU (even at max compute capacity). In this case, we can discard this "stolen" idle times which becomes meaningless. In order to achieve this time scaling, a new clock_pelt is created per rq. The increase of this clock scales with current capacity when something is running on rq and synchronizes with clock_task when rq is idle. With this mechanism, we ensure the same running and idle time whatever the current capacity. This also enables to simplify the pelt algorithm by removing all references of uarch and frequency and applying the same contribution to utilization and loads. Furthermore, the scaling is done only once per update of clock (update_rq_clock_task()) instead of during each update of sched_entities and cfs/rt/dl_rq of the rq like the current implementation. This is interesting when cgroup are involved as shown in the results below: On a hikey (octo Arm64 platform). Performance cpufreq governor and only shallowest c-state to remove variance generated by those power features so we only track the impact of pelt algo. each test runs 16 times: ./perf bench sched pipe (higher is better) kernel tip/sched/core + patch ops/seconds ops/seconds diff cgroup root 59652(+/- 0.18%) 59876(+/- 0.24%) +0.38% level1 55608(+/- 0.27%) 55923(+/- 0.24%) +0.57% level2 52115(+/- 0.29%) 52564(+/- 0.22%) +0.86% hackbench -l 1000 (lower is better) kernel tip/sched/core + patch duration(sec) duration(sec) diff cgroup root 4.453(+/- 2.37%) 4.383(+/- 2.88%) -1.57% level1 4.859(+/- 8.50%) 4.830(+/- 7.07%) -0.60% level2 5.063(+/- 9.83%) 4.928(+/- 9.66%) -2.66% Then, the responsiveness of PELT is improved when CPU is not running at max capacity with this new algorithm. I have put below some examples of duration to reach some typical load values according to the capacity of the CPU with current implementation and with this patch. These values has been computed based on the geometric series and the half period value: Util (%) max capacity half capacity(mainline) half capacity(w/ patch) 972 (95%) 138ms not reachable 276ms 486 (47.5%) 30ms 138ms 60ms 256 (25%) 13ms 32ms 26ms On my hikey (octo Arm64 platform) with schedutil governor, the time to reach max OPP when starting from a null utilization, decreases from 223ms with current scale invariance down to 121ms with the new algorithm. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: patrick.bellasi@arm.com Cc: pjt@google.com Cc: pkondeti@codeaurora.org Cc: quentin.perret@arm.com Cc: rjw@rjwysocki.net Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Link: https://lkml.kernel.org/r/1548257214-13745-3-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-01-23 15:26:53 +00:00
* Rescale running sum to be in the same range as runnable sum
* running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT]
* runnable_sum is in [0 : LOAD_AVG_MAX]
sched/fair: Update and fix the runnable propagation rule Unlike running, the runnable part can't be directly propagated through the hierarchy when we migrate a task. The main reason is that runnable time can be shared with other sched_entities that stay on the rq and this runnable time will also remain on prev cfs_rq and must not be removed. Instead, we can estimate what should be the new runnable of the prev cfs_rq and check that this estimation stay in a possible range. The prop_runnable_sum is a good estimation when adding runnable_sum but fails most often when we remove it. Instead, we could use the formula below instead: gcfs_rq's runnable_sum = gcfs_rq->avg.load_sum / gcfs_rq->load.weight which assumes that tasks are equally runnable which is not true but easy to compute. Beside these estimates, we have several simple rules that help us to filter out wrong ones: - ge->avg.runnable_sum <= than LOAD_AVG_MAX - ge->avg.runnable_sum >= ge->avg.running_sum (ge->avg.util_sum << LOAD_AVG_MAX) - ge->avg.runnable_sum can't increase when we detach a task The effect of these fixes is better cgroups balancing. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Ben Segall <bsegall@google.com> Cc: Chris Mason <clm@fb.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Tejun Heo <tj@kernel.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Yuyang Du <yuyang.du@intel.com> Link: http://lkml.kernel.org/r/1510842112-21028-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-11-16 14:21:52 +00:00
*/
sched/fair: Update scale invariance of PELT The current implementation of load tracking invariance scales the contribution with current frequency and uarch performance (only for utilization) of the CPU. One main result of this formula is that the figures are capped by current capacity of CPU. Another one is that the load_avg is not invariant because not scaled with uarch. The util_avg of a periodic task that runs r time slots every p time slots varies in the range : U * (1-y^r)/(1-y^p) * y^i < Utilization < U * (1-y^r)/(1-y^p) with U is the max util_avg value = SCHED_CAPACITY_SCALE At a lower capacity, the range becomes: U * C * (1-y^r')/(1-y^p) * y^i' < Utilization < U * C * (1-y^r')/(1-y^p) with C reflecting the compute capacity ratio between current capacity and max capacity. so C tries to compensate changes in (1-y^r') but it can't be accurate. Instead of scaling the contribution value of PELT algo, we should scale the running time. The PELT signal aims to track the amount of computation of tasks and/or rq so it seems more correct to scale the running time to reflect the effective amount of computation done since the last update. In order to be fully invariant, we need to apply the same amount of running time and idle time whatever the current capacity. Because running at lower capacity implies that the task will run longer, we have to ensure that the same amount of idle time will be applied when system becomes idle and no idle time has been "stolen". But reaching the maximum utilization value (SCHED_CAPACITY_SCALE) means that the task is seen as an always-running task whatever the capacity of the CPU (even at max compute capacity). In this case, we can discard this "stolen" idle times which becomes meaningless. In order to achieve this time scaling, a new clock_pelt is created per rq. The increase of this clock scales with current capacity when something is running on rq and synchronizes with clock_task when rq is idle. With this mechanism, we ensure the same running and idle time whatever the current capacity. This also enables to simplify the pelt algorithm by removing all references of uarch and frequency and applying the same contribution to utilization and loads. Furthermore, the scaling is done only once per update of clock (update_rq_clock_task()) instead of during each update of sched_entities and cfs/rt/dl_rq of the rq like the current implementation. This is interesting when cgroup are involved as shown in the results below: On a hikey (octo Arm64 platform). Performance cpufreq governor and only shallowest c-state to remove variance generated by those power features so we only track the impact of pelt algo. each test runs 16 times: ./perf bench sched pipe (higher is better) kernel tip/sched/core + patch ops/seconds ops/seconds diff cgroup root 59652(+/- 0.18%) 59876(+/- 0.24%) +0.38% level1 55608(+/- 0.27%) 55923(+/- 0.24%) +0.57% level2 52115(+/- 0.29%) 52564(+/- 0.22%) +0.86% hackbench -l 1000 (lower is better) kernel tip/sched/core + patch duration(sec) duration(sec) diff cgroup root 4.453(+/- 2.37%) 4.383(+/- 2.88%) -1.57% level1 4.859(+/- 8.50%) 4.830(+/- 7.07%) -0.60% level2 5.063(+/- 9.83%) 4.928(+/- 9.66%) -2.66% Then, the responsiveness of PELT is improved when CPU is not running at max capacity with this new algorithm. I have put below some examples of duration to reach some typical load values according to the capacity of the CPU with current implementation and with this patch. These values has been computed based on the geometric series and the half period value: Util (%) max capacity half capacity(mainline) half capacity(w/ patch) 972 (95%) 138ms not reachable 276ms 486 (47.5%) 30ms 138ms 60ms 256 (25%) 13ms 32ms 26ms On my hikey (octo Arm64 platform) with schedutil governor, the time to reach max OPP when starting from a null utilization, decreases from 223ms with current scale invariance down to 121ms with the new algorithm. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: patrick.bellasi@arm.com Cc: pjt@google.com Cc: pkondeti@codeaurora.org Cc: quentin.perret@arm.com Cc: rjw@rjwysocki.net Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Link: https://lkml.kernel.org/r/1548257214-13745-3-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-01-23 15:26:53 +00:00
running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT;
sched/fair: Update and fix the runnable propagation rule Unlike running, the runnable part can't be directly propagated through the hierarchy when we migrate a task. The main reason is that runnable time can be shared with other sched_entities that stay on the rq and this runnable time will also remain on prev cfs_rq and must not be removed. Instead, we can estimate what should be the new runnable of the prev cfs_rq and check that this estimation stay in a possible range. The prop_runnable_sum is a good estimation when adding runnable_sum but fails most often when we remove it. Instead, we could use the formula below instead: gcfs_rq's runnable_sum = gcfs_rq->avg.load_sum / gcfs_rq->load.weight which assumes that tasks are equally runnable which is not true but easy to compute. Beside these estimates, we have several simple rules that help us to filter out wrong ones: - ge->avg.runnable_sum <= than LOAD_AVG_MAX - ge->avg.runnable_sum >= ge->avg.running_sum (ge->avg.util_sum << LOAD_AVG_MAX) - ge->avg.runnable_sum can't increase when we detach a task The effect of these fixes is better cgroups balancing. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Ben Segall <bsegall@google.com> Cc: Chris Mason <clm@fb.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Tejun Heo <tj@kernel.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Yuyang Du <yuyang.du@intel.com> Link: http://lkml.kernel.org/r/1510842112-21028-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-11-16 14:21:52 +00:00
runnable_sum = max(runnable_sum, running_sum);
load_sum = (s64)se_weight(se) * runnable_sum;
load_avg = div_s64(load_sum, LOAD_AVG_MAX);
sched/fair: Update and fix the runnable propagation rule Unlike running, the runnable part can't be directly propagated through the hierarchy when we migrate a task. The main reason is that runnable time can be shared with other sched_entities that stay on the rq and this runnable time will also remain on prev cfs_rq and must not be removed. Instead, we can estimate what should be the new runnable of the prev cfs_rq and check that this estimation stay in a possible range. The prop_runnable_sum is a good estimation when adding runnable_sum but fails most often when we remove it. Instead, we could use the formula below instead: gcfs_rq's runnable_sum = gcfs_rq->avg.load_sum / gcfs_rq->load.weight which assumes that tasks are equally runnable which is not true but easy to compute. Beside these estimates, we have several simple rules that help us to filter out wrong ones: - ge->avg.runnable_sum <= than LOAD_AVG_MAX - ge->avg.runnable_sum >= ge->avg.running_sum (ge->avg.util_sum << LOAD_AVG_MAX) - ge->avg.runnable_sum can't increase when we detach a task The effect of these fixes is better cgroups balancing. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Ben Segall <bsegall@google.com> Cc: Chris Mason <clm@fb.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Tejun Heo <tj@kernel.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Yuyang Du <yuyang.du@intel.com> Link: http://lkml.kernel.org/r/1510842112-21028-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-11-16 14:21:52 +00:00
delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
delta_avg = load_avg - se->avg.load_avg;
sched/fair: Update and fix the runnable propagation rule Unlike running, the runnable part can't be directly propagated through the hierarchy when we migrate a task. The main reason is that runnable time can be shared with other sched_entities that stay on the rq and this runnable time will also remain on prev cfs_rq and must not be removed. Instead, we can estimate what should be the new runnable of the prev cfs_rq and check that this estimation stay in a possible range. The prop_runnable_sum is a good estimation when adding runnable_sum but fails most often when we remove it. Instead, we could use the formula below instead: gcfs_rq's runnable_sum = gcfs_rq->avg.load_sum / gcfs_rq->load.weight which assumes that tasks are equally runnable which is not true but easy to compute. Beside these estimates, we have several simple rules that help us to filter out wrong ones: - ge->avg.runnable_sum <= than LOAD_AVG_MAX - ge->avg.runnable_sum >= ge->avg.running_sum (ge->avg.util_sum << LOAD_AVG_MAX) - ge->avg.runnable_sum can't increase when we detach a task The effect of these fixes is better cgroups balancing. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Ben Segall <bsegall@google.com> Cc: Chris Mason <clm@fb.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Tejun Heo <tj@kernel.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Yuyang Du <yuyang.du@intel.com> Link: http://lkml.kernel.org/r/1510842112-21028-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-11-16 14:21:52 +00:00
se->avg.load_sum = runnable_sum;
se->avg.load_avg = load_avg;
add_positive(&cfs_rq->avg.load_avg, delta_avg);
add_positive(&cfs_rq->avg.load_sum, delta_sum);
}
static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
{
cfs_rq->propagate = 1;
cfs_rq->prop_runnable_sum += runnable_sum;
}
/* Update task and its cfs_rq load average */
static inline int propagate_entity_load_avg(struct sched_entity *se)
{
struct cfs_rq *cfs_rq, *gcfs_rq;
if (entity_is_task(se))
return 0;
gcfs_rq = group_cfs_rq(se);
if (!gcfs_rq->propagate)
return 0;
gcfs_rq->propagate = 0;
cfs_rq = cfs_rq_of(se);
add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
update_tg_cfs_util(cfs_rq, se, gcfs_rq);
sched/pelt: Add a new runnable average signal Now that runnable_load_avg has been removed, we can replace it by a new signal that will highlight the runnable pressure on a cfs_rq. This signal track the waiting time of tasks on rq and can help to better define the state of rqs. At now, only util_avg is used to define the state of a rq: A rq with more that around 80% of utilization and more than 1 tasks is considered as overloaded. But the util_avg signal of a rq can become temporaly low after that a task migrated onto another rq which can bias the classification of the rq. When tasks compete for the same rq, their runnable average signal will be higher than util_avg as it will include the waiting time and we can use this signal to better classify cfs_rqs. The new runnable_avg will track the runnable time of a task which simply adds the waiting time to the running time. The runnable _avg of cfs_rq will be the /Sum of se's runnable_avg and the runnable_avg of group entity will follow the one of the rq similarly to util_avg. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: "Dietmar Eggemann <dietmar.eggemann@arm.com>" Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Valentin Schneider <valentin.schneider@arm.com> Cc: Phil Auld <pauld@redhat.com> Cc: Hillf Danton <hdanton@sina.com> Link: https://lore.kernel.org/r/20200224095223.13361-9-mgorman@techsingularity.net
2020-02-24 09:52:18 +00:00
update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
update_tg_cfs_load(cfs_rq, se, gcfs_rq);
trace_pelt_cfs_tp(cfs_rq);
trace_pelt_se_tp(se);
return 1;
}
/*
* Check if we need to update the load and the utilization of a blocked
* group_entity:
*/
static inline bool skip_blocked_update(struct sched_entity *se)
{
struct cfs_rq *gcfs_rq = group_cfs_rq(se);
/*
* If sched_entity still have not zero load or utilization, we have to
* decay it:
*/
if (se->avg.load_avg || se->avg.util_avg)
return false;
/*
* If there is a pending propagation, we have to update the load and
* the utilization of the sched_entity:
*/
if (gcfs_rq->propagate)
return false;
/*
* Otherwise, the load and the utilization of the sched_entity is
* already zero and there is no pending propagation, so it will be a
* waste of time to try to decay it:
*/
return true;
}
#else /* CONFIG_FAIR_GROUP_SCHED */
sched/fair: Rewrite runnable load and utilization average tracking The idea of runnable load average (let runnable time contribute to weight) was proposed by Paul Turner and Ben Segall, and it is still followed by this rewrite. This rewrite aims to solve the following issues: 1. cfs_rq's load average (namely runnable_load_avg and blocked_load_avg) is updated at the granularity of an entity at a time, which results in the cfs_rq's load average is stale or partially updated: at any time, only one entity is up to date, all other entities are effectively lagging behind. This is undesirable. To illustrate, if we have n runnable entities in the cfs_rq, as time elapses, they certainly become outdated: t0: cfs_rq { e1_old, e2_old, ..., en_old } and when we update: t1: update e1, then we have cfs_rq { e1_new, e2_old, ..., en_old } t2: update e2, then we have cfs_rq { e1_old, e2_new, ..., en_old } ... We solve this by combining all runnable entities' load averages together in cfs_rq's avg, and update the cfs_rq's avg as a whole. This is based on the fact that if we regard the update as a function, then: w * update(e) = update(w * e) and update(e1) + update(e2) = update(e1 + e2), then w1 * update(e1) + w2 * update(e2) = update(w1 * e1 + w2 * e2) therefore, by this rewrite, we have an entirely updated cfs_rq at the time we update it: t1: update cfs_rq { e1_new, e2_new, ..., en_new } t2: update cfs_rq { e1_new, e2_new, ..., en_new } ... 2. cfs_rq's load average is different between top rq->cfs_rq and other task_group's per CPU cfs_rqs in whether or not blocked_load_average contributes to the load. The basic idea behind runnable load average (the same for utilization) is that the blocked state is taken into account as opposed to only accounting for the currently runnable state. Therefore, the average should include both the runnable/running and blocked load averages. This rewrite does that. In addition, we also combine runnable/running and blocked averages of all entities into the cfs_rq's average, and update it together at once. This is based on the fact that: update(runnable) + update(blocked) = update(runnable + blocked) This significantly reduces the code as we don't need to separately maintain/update runnable/running load and blocked load. 3. How task_group entities' share is calculated is complex and imprecise. We reduce the complexity in this rewrite to allow a very simple rule: the task_group's load_avg is aggregated from its per CPU cfs_rqs's load_avgs. Then group entity's weight is simply proportional to its own cfs_rq's load_avg / task_group's load_avg. To illustrate, if a task_group has { cfs_rq1, cfs_rq2, ..., cfs_rqn }, then, task_group_avg = cfs_rq1_avg + cfs_rq2_avg + ... + cfs_rqn_avg, then cfs_rqx's entity's share = cfs_rqx_avg / task_group_avg * task_group's share To sum up, this rewrite in principle is equivalent to the current one, but fixes the issues described above. Turns out, it significantly reduces the code complexity and hence increases clarity and efficiency. In addition, the new averages are more smooth/continuous (no spurious spikes and valleys) and updated more consistently and quickly to reflect the load dynamics. As a result, we have less load tracking overhead, better performance, and especially better power efficiency due to more balanced load. Signed-off-by: Yuyang Du <yuyang.du@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: arjan@linux.intel.com Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: fengguang.wu@intel.com Cc: len.brown@intel.com Cc: morten.rasmussen@arm.com Cc: pjt@google.com Cc: rafael.j.wysocki@intel.com Cc: umgwanakikbuti@gmail.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1436918682-4971-3-git-send-email-yuyang.du@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-07-15 00:04:37 +00:00
static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
static inline int propagate_entity_load_avg(struct sched_entity *se)
{
return 0;
}
static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
#endif /* CONFIG_FAIR_GROUP_SCHED */
/**
* update_cfs_rq_load_avg - update the cfs_rq's load/util averages
sched/fair: Update scale invariance of PELT The current implementation of load tracking invariance scales the contribution with current frequency and uarch performance (only for utilization) of the CPU. One main result of this formula is that the figures are capped by current capacity of CPU. Another one is that the load_avg is not invariant because not scaled with uarch. The util_avg of a periodic task that runs r time slots every p time slots varies in the range : U * (1-y^r)/(1-y^p) * y^i < Utilization < U * (1-y^r)/(1-y^p) with U is the max util_avg value = SCHED_CAPACITY_SCALE At a lower capacity, the range becomes: U * C * (1-y^r')/(1-y^p) * y^i' < Utilization < U * C * (1-y^r')/(1-y^p) with C reflecting the compute capacity ratio between current capacity and max capacity. so C tries to compensate changes in (1-y^r') but it can't be accurate. Instead of scaling the contribution value of PELT algo, we should scale the running time. The PELT signal aims to track the amount of computation of tasks and/or rq so it seems more correct to scale the running time to reflect the effective amount of computation done since the last update. In order to be fully invariant, we need to apply the same amount of running time and idle time whatever the current capacity. Because running at lower capacity implies that the task will run longer, we have to ensure that the same amount of idle time will be applied when system becomes idle and no idle time has been "stolen". But reaching the maximum utilization value (SCHED_CAPACITY_SCALE) means that the task is seen as an always-running task whatever the capacity of the CPU (even at max compute capacity). In this case, we can discard this "stolen" idle times which becomes meaningless. In order to achieve this time scaling, a new clock_pelt is created per rq. The increase of this clock scales with current capacity when something is running on rq and synchronizes with clock_task when rq is idle. With this mechanism, we ensure the same running and idle time whatever the current capacity. This also enables to simplify the pelt algorithm by removing all references of uarch and frequency and applying the same contribution to utilization and loads. Furthermore, the scaling is done only once per update of clock (update_rq_clock_task()) instead of during each update of sched_entities and cfs/rt/dl_rq of the rq like the current implementation. This is interesting when cgroup are involved as shown in the results below: On a hikey (octo Arm64 platform). Performance cpufreq governor and only shallowest c-state to remove variance generated by those power features so we only track the impact of pelt algo. each test runs 16 times: ./perf bench sched pipe (higher is better) kernel tip/sched/core + patch ops/seconds ops/seconds diff cgroup root 59652(+/- 0.18%) 59876(+/- 0.24%) +0.38% level1 55608(+/- 0.27%) 55923(+/- 0.24%) +0.57% level2 52115(+/- 0.29%) 52564(+/- 0.22%) +0.86% hackbench -l 1000 (lower is better) kernel tip/sched/core + patch duration(sec) duration(sec) diff cgroup root 4.453(+/- 2.37%) 4.383(+/- 2.88%) -1.57% level1 4.859(+/- 8.50%) 4.830(+/- 7.07%) -0.60% level2 5.063(+/- 9.83%) 4.928(+/- 9.66%) -2.66% Then, the responsiveness of PELT is improved when CPU is not running at max capacity with this new algorithm. I have put below some examples of duration to reach some typical load values according to the capacity of the CPU with current implementation and with this patch. These values has been computed based on the geometric series and the half period value: Util (%) max capacity half capacity(mainline) half capacity(w/ patch) 972 (95%) 138ms not reachable 276ms 486 (47.5%) 30ms 138ms 60ms 256 (25%) 13ms 32ms 26ms On my hikey (octo Arm64 platform) with schedutil governor, the time to reach max OPP when starting from a null utilization, decreases from 223ms with current scale invariance down to 121ms with the new algorithm. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: patrick.bellasi@arm.com Cc: pjt@google.com Cc: pkondeti@codeaurora.org Cc: quentin.perret@arm.com Cc: rjw@rjwysocki.net Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Link: https://lkml.kernel.org/r/1548257214-13745-3-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-01-23 15:26:53 +00:00
* @now: current time, as per cfs_rq_clock_pelt()
* @cfs_rq: cfs_rq to update
*
* The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
* avg. The immediate corollary is that all (fair) tasks must be attached, see
* post_init_entity_util_avg().
*
* cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
*
* Returns true if the load decayed or we removed load.
*
* Since both these conditions indicate a changed cfs_rq->avg.load we should
* call update_tg_load_avg() when this function returns true.
*/
static inline int
update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
{
sched/pelt: Add a new runnable average signal Now that runnable_load_avg has been removed, we can replace it by a new signal that will highlight the runnable pressure on a cfs_rq. This signal track the waiting time of tasks on rq and can help to better define the state of rqs. At now, only util_avg is used to define the state of a rq: A rq with more that around 80% of utilization and more than 1 tasks is considered as overloaded. But the util_avg signal of a rq can become temporaly low after that a task migrated onto another rq which can bias the classification of the rq. When tasks compete for the same rq, their runnable average signal will be higher than util_avg as it will include the waiting time and we can use this signal to better classify cfs_rqs. The new runnable_avg will track the runnable time of a task which simply adds the waiting time to the running time. The runnable _avg of cfs_rq will be the /Sum of se's runnable_avg and the runnable_avg of group entity will follow the one of the rq similarly to util_avg. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: "Dietmar Eggemann <dietmar.eggemann@arm.com>" Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Valentin Schneider <valentin.schneider@arm.com> Cc: Phil Auld <pauld@redhat.com> Cc: Hillf Danton <hdanton@sina.com> Link: https://lore.kernel.org/r/20200224095223.13361-9-mgorman@techsingularity.net
2020-02-24 09:52:18 +00:00
unsigned long removed_load = 0, removed_util = 0, removed_runnable = 0;
sched/fair: Rewrite runnable load and utilization average tracking The idea of runnable load average (let runnable time contribute to weight) was proposed by Paul Turner and Ben Segall, and it is still followed by this rewrite. This rewrite aims to solve the following issues: 1. cfs_rq's load average (namely runnable_load_avg and blocked_load_avg) is updated at the granularity of an entity at a time, which results in the cfs_rq's load average is stale or partially updated: at any time, only one entity is up to date, all other entities are effectively lagging behind. This is undesirable. To illustrate, if we have n runnable entities in the cfs_rq, as time elapses, they certainly become outdated: t0: cfs_rq { e1_old, e2_old, ..., en_old } and when we update: t1: update e1, then we have cfs_rq { e1_new, e2_old, ..., en_old } t2: update e2, then we have cfs_rq { e1_old, e2_new, ..., en_old } ... We solve this by combining all runnable entities' load averages together in cfs_rq's avg, and update the cfs_rq's avg as a whole. This is based on the fact that if we regard the update as a function, then: w * update(e) = update(w * e) and update(e1) + update(e2) = update(e1 + e2), then w1 * update(e1) + w2 * update(e2) = update(w1 * e1 + w2 * e2) therefore, by this rewrite, we have an entirely updated cfs_rq at the time we update it: t1: update cfs_rq { e1_new, e2_new, ..., en_new } t2: update cfs_rq { e1_new, e2_new, ..., en_new } ... 2. cfs_rq's load average is different between top rq->cfs_rq and other task_group's per CPU cfs_rqs in whether or not blocked_load_average contributes to the load. The basic idea behind runnable load average (the same for utilization) is that the blocked state is taken into account as opposed to only accounting for the currently runnable state. Therefore, the average should include both the runnable/running and blocked load averages. This rewrite does that. In addition, we also combine runnable/running and blocked averages of all entities into the cfs_rq's average, and update it together at once. This is based on the fact that: update(runnable) + update(blocked) = update(runnable + blocked) This significantly reduces the code as we don't need to separately maintain/update runnable/running load and blocked load. 3. How task_group entities' share is calculated is complex and imprecise. We reduce the complexity in this rewrite to allow a very simple rule: the task_group's load_avg is aggregated from its per CPU cfs_rqs's load_avgs. Then group entity's weight is simply proportional to its own cfs_rq's load_avg / task_group's load_avg. To illustrate, if a task_group has { cfs_rq1, cfs_rq2, ..., cfs_rqn }, then, task_group_avg = cfs_rq1_avg + cfs_rq2_avg + ... + cfs_rqn_avg, then cfs_rqx's entity's share = cfs_rqx_avg / task_group_avg * task_group's share To sum up, this rewrite in principle is equivalent to the current one, but fixes the issues described above. Turns out, it significantly reduces the code complexity and hence increases clarity and efficiency. In addition, the new averages are more smooth/continuous (no spurious spikes and valleys) and updated more consistently and quickly to reflect the load dynamics. As a result, we have less load tracking overhead, better performance, and especially better power efficiency due to more balanced load. Signed-off-by: Yuyang Du <yuyang.du@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: arjan@linux.intel.com Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: fengguang.wu@intel.com Cc: len.brown@intel.com Cc: morten.rasmussen@arm.com Cc: pjt@google.com Cc: rafael.j.wysocki@intel.com Cc: umgwanakikbuti@gmail.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1436918682-4971-3-git-send-email-yuyang.du@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-07-15 00:04:37 +00:00
struct sched_avg *sa = &cfs_rq->avg;
int decayed = 0;
if (cfs_rq->removed.nr) {
unsigned long r;
u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib;
raw_spin_lock(&cfs_rq->removed.lock);
swap(cfs_rq->removed.util_avg, removed_util);
swap(cfs_rq->removed.load_avg, removed_load);
sched/pelt: Add a new runnable average signal Now that runnable_load_avg has been removed, we can replace it by a new signal that will highlight the runnable pressure on a cfs_rq. This signal track the waiting time of tasks on rq and can help to better define the state of rqs. At now, only util_avg is used to define the state of a rq: A rq with more that around 80% of utilization and more than 1 tasks is considered as overloaded. But the util_avg signal of a rq can become temporaly low after that a task migrated onto another rq which can bias the classification of the rq. When tasks compete for the same rq, their runnable average signal will be higher than util_avg as it will include the waiting time and we can use this signal to better classify cfs_rqs. The new runnable_avg will track the runnable time of a task which simply adds the waiting time to the running time. The runnable _avg of cfs_rq will be the /Sum of se's runnable_avg and the runnable_avg of group entity will follow the one of the rq similarly to util_avg. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: "Dietmar Eggemann <dietmar.eggemann@arm.com>" Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Valentin Schneider <valentin.schneider@arm.com> Cc: Phil Auld <pauld@redhat.com> Cc: Hillf Danton <hdanton@sina.com> Link: https://lore.kernel.org/r/20200224095223.13361-9-mgorman@techsingularity.net
2020-02-24 09:52:18 +00:00
swap(cfs_rq->removed.runnable_avg, removed_runnable);
cfs_rq->removed.nr = 0;
raw_spin_unlock(&cfs_rq->removed.lock);
r = removed_load;
sched/fair: Fix cfs_rq avg tracking underflow As per commit: b7fa30c9cc48 ("sched/fair: Fix post_init_entity_util_avg() serialization") > the code generated from update_cfs_rq_load_avg(): > > if (atomic_long_read(&cfs_rq->removed_load_avg)) { > s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0); > sa->load_avg = max_t(long, sa->load_avg - r, 0); > sa->load_sum = max_t(s64, sa->load_sum - r * LOAD_AVG_MAX, 0); > removed_load = 1; > } > > turns into: > > ffffffff81087064: 49 8b 85 98 00 00 00 mov 0x98(%r13),%rax > ffffffff8108706b: 48 85 c0 test %rax,%rax > ffffffff8108706e: 74 40 je ffffffff810870b0 <update_blocked_averages+0xc0> > ffffffff81087070: 4c 89 f8 mov %r15,%rax > ffffffff81087073: 49 87 85 98 00 00 00 xchg %rax,0x98(%r13) > ffffffff8108707a: 49 29 45 70 sub %rax,0x70(%r13) > ffffffff8108707e: 4c 89 f9 mov %r15,%rcx > ffffffff81087081: bb 01 00 00 00 mov $0x1,%ebx > ffffffff81087086: 49 83 7d 70 00 cmpq $0x0,0x70(%r13) > ffffffff8108708b: 49 0f 49 4d 70 cmovns 0x70(%r13),%rcx > > Which you'll note ends up with sa->load_avg -= r in memory at > ffffffff8108707a. So I _should_ have looked at other unserialized users of ->load_avg, but alas. Luckily nikbor reported a similar /0 from task_h_load() which instantly triggered recollection of this here problem. Aside from the intermediate value hitting memory and causing problems, there's another problem: the underflow detection relies on the signed bit. This reduces the effective width of the variables, IOW its effectively the same as having these variables be of signed type. This patch changes to a different means of unsigned underflow detection to not rely on the signed bit. This allows the variables to use the 'full' unsigned range. And it does so with explicit LOAD - STORE to ensure any intermediate value will never be visible in memory, allowing these unserialized loads. Note: GCC generates crap code for this, might warrant a look later. Note2: I say 'full' above, if we end up at U*_MAX we'll still explode; maybe we should do clamping on add too. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Chris Wilson <chris@chris-wilson.co.uk> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Yuyang Du <yuyang.du@intel.com> Cc: bsegall@google.com Cc: kernel@kyup.com Cc: morten.rasmussen@arm.com Cc: pjt@google.com Cc: steve.muckle@linaro.org Fixes: 9d89c257dfb9 ("sched/fair: Rewrite runnable load and utilization average tracking") Link: http://lkml.kernel.org/r/20160617091948.GJ30927@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-06-16 08:50:40 +00:00
sub_positive(&sa->load_avg, r);
sub_positive(&sa->load_sum, r * divider);
r = removed_util;
sched/fair: Fix cfs_rq avg tracking underflow As per commit: b7fa30c9cc48 ("sched/fair: Fix post_init_entity_util_avg() serialization") > the code generated from update_cfs_rq_load_avg(): > > if (atomic_long_read(&cfs_rq->removed_load_avg)) { > s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0); > sa->load_avg = max_t(long, sa->load_avg - r, 0); > sa->load_sum = max_t(s64, sa->load_sum - r * LOAD_AVG_MAX, 0); > removed_load = 1; > } > > turns into: > > ffffffff81087064: 49 8b 85 98 00 00 00 mov 0x98(%r13),%rax > ffffffff8108706b: 48 85 c0 test %rax,%rax > ffffffff8108706e: 74 40 je ffffffff810870b0 <update_blocked_averages+0xc0> > ffffffff81087070: 4c 89 f8 mov %r15,%rax > ffffffff81087073: 49 87 85 98 00 00 00 xchg %rax,0x98(%r13) > ffffffff8108707a: 49 29 45 70 sub %rax,0x70(%r13) > ffffffff8108707e: 4c 89 f9 mov %r15,%rcx > ffffffff81087081: bb 01 00 00 00 mov $0x1,%ebx > ffffffff81087086: 49 83 7d 70 00 cmpq $0x0,0x70(%r13) > ffffffff8108708b: 49 0f 49 4d 70 cmovns 0x70(%r13),%rcx > > Which you'll note ends up with sa->load_avg -= r in memory at > ffffffff8108707a. So I _should_ have looked at other unserialized users of ->load_avg, but alas. Luckily nikbor reported a similar /0 from task_h_load() which instantly triggered recollection of this here problem. Aside from the intermediate value hitting memory and causing problems, there's another problem: the underflow detection relies on the signed bit. This reduces the effective width of the variables, IOW its effectively the same as having these variables be of signed type. This patch changes to a different means of unsigned underflow detection to not rely on the signed bit. This allows the variables to use the 'full' unsigned range. And it does so with explicit LOAD - STORE to ensure any intermediate value will never be visible in memory, allowing these unserialized loads. Note: GCC generates crap code for this, might warrant a look later. Note2: I say 'full' above, if we end up at U*_MAX we'll still explode; maybe we should do clamping on add too. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Chris Wilson <chris@chris-wilson.co.uk> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Yuyang Du <yuyang.du@intel.com> Cc: bsegall@google.com Cc: kernel@kyup.com Cc: morten.rasmussen@arm.com Cc: pjt@google.com Cc: steve.muckle@linaro.org Fixes: 9d89c257dfb9 ("sched/fair: Rewrite runnable load and utilization average tracking") Link: http://lkml.kernel.org/r/20160617091948.GJ30927@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-06-16 08:50:40 +00:00
sub_positive(&sa->util_avg, r);
sub_positive(&sa->util_sum, r * divider);
sched/pelt: Add a new runnable average signal Now that runnable_load_avg has been removed, we can replace it by a new signal that will highlight the runnable pressure on a cfs_rq. This signal track the waiting time of tasks on rq and can help to better define the state of rqs. At now, only util_avg is used to define the state of a rq: A rq with more that around 80% of utilization and more than 1 tasks is considered as overloaded. But the util_avg signal of a rq can become temporaly low after that a task migrated onto another rq which can bias the classification of the rq. When tasks compete for the same rq, their runnable average signal will be higher than util_avg as it will include the waiting time and we can use this signal to better classify cfs_rqs. The new runnable_avg will track the runnable time of a task which simply adds the waiting time to the running time. The runnable _avg of cfs_rq will be the /Sum of se's runnable_avg and the runnable_avg of group entity will follow the one of the rq similarly to util_avg. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: "Dietmar Eggemann <dietmar.eggemann@arm.com>" Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Valentin Schneider <valentin.schneider@arm.com> Cc: Phil Auld <pauld@redhat.com> Cc: Hillf Danton <hdanton@sina.com> Link: https://lore.kernel.org/r/20200224095223.13361-9-mgorman@techsingularity.net
2020-02-24 09:52:18 +00:00
r = removed_runnable;
sub_positive(&sa->runnable_avg, r);
sub_positive(&sa->runnable_sum, r * divider);
/*
* removed_runnable is the unweighted version of removed_load so we
* can use it to estimate removed_load_sum.
*/
add_tg_cfs_propagate(cfs_rq,
-(long)(removed_runnable * divider) >> SCHED_CAPACITY_SHIFT);
decayed = 1;
sched/fair: Rewrite runnable load and utilization average tracking The idea of runnable load average (let runnable time contribute to weight) was proposed by Paul Turner and Ben Segall, and it is still followed by this rewrite. This rewrite aims to solve the following issues: 1. cfs_rq's load average (namely runnable_load_avg and blocked_load_avg) is updated at the granularity of an entity at a time, which results in the cfs_rq's load average is stale or partially updated: at any time, only one entity is up to date, all other entities are effectively lagging behind. This is undesirable. To illustrate, if we have n runnable entities in the cfs_rq, as time elapses, they certainly become outdated: t0: cfs_rq { e1_old, e2_old, ..., en_old } and when we update: t1: update e1, then we have cfs_rq { e1_new, e2_old, ..., en_old } t2: update e2, then we have cfs_rq { e1_old, e2_new, ..., en_old } ... We solve this by combining all runnable entities' load averages together in cfs_rq's avg, and update the cfs_rq's avg as a whole. This is based on the fact that if we regard the update as a function, then: w * update(e) = update(w * e) and update(e1) + update(e2) = update(e1 + e2), then w1 * update(e1) + w2 * update(e2) = update(w1 * e1 + w2 * e2) therefore, by this rewrite, we have an entirely updated cfs_rq at the time we update it: t1: update cfs_rq { e1_new, e2_new, ..., en_new } t2: update cfs_rq { e1_new, e2_new, ..., en_new } ... 2. cfs_rq's load average is different between top rq->cfs_rq and other task_group's per CPU cfs_rqs in whether or not blocked_load_average contributes to the load. The basic idea behind runnable load average (the same for utilization) is that the blocked state is taken into account as opposed to only accounting for the currently runnable state. Therefore, the average should include both the runnable/running and blocked load averages. This rewrite does that. In addition, we also combine runnable/running and blocked averages of all entities into the cfs_rq's average, and update it together at once. This is based on the fact that: update(runnable) + update(blocked) = update(runnable + blocked) This significantly reduces the code as we don't need to separately maintain/update runnable/running load and blocked load. 3. How task_group entities' share is calculated is complex and imprecise. We reduce the complexity in this rewrite to allow a very simple rule: the task_group's load_avg is aggregated from its per CPU cfs_rqs's load_avgs. Then group entity's weight is simply proportional to its own cfs_rq's load_avg / task_group's load_avg. To illustrate, if a task_group has { cfs_rq1, cfs_rq2, ..., cfs_rqn }, then, task_group_avg = cfs_rq1_avg + cfs_rq2_avg + ... + cfs_rqn_avg, then cfs_rqx's entity's share = cfs_rqx_avg / task_group_avg * task_group's share To sum up, this rewrite in principle is equivalent to the current one, but fixes the issues described above. Turns out, it significantly reduces the code complexity and hence increases clarity and efficiency. In addition, the new averages are more smooth/continuous (no spurious spikes and valleys) and updated more consistently and quickly to reflect the load dynamics. As a result, we have less load tracking overhead, better performance, and especially better power efficiency due to more balanced load. Signed-off-by: Yuyang Du <yuyang.du@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: arjan@linux.intel.com Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: fengguang.wu@intel.com Cc: len.brown@intel.com Cc: morten.rasmussen@arm.com Cc: pjt@google.com Cc: rafael.j.wysocki@intel.com Cc: umgwanakikbuti@gmail.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1436918682-4971-3-git-send-email-yuyang.du@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-07-15 00:04:37 +00:00
}
sched: Add sched_avg::utilization_avg_contrib Add new statistics which reflect the average time a task is running on the CPU and the sum of these running time of the tasks on a runqueue. The latter is named utilization_load_avg. This patch is based on the usage metric that was proposed in the 1st versions of the per-entity load tracking patchset by Paul Turner <pjt@google.com> but that has be removed afterwards. This version differs from the original one in the sense that it's not linked to task_group. The rq's utilization_load_avg will be used to check if a rq is overloaded or not instead of trying to compute how many tasks a group of CPUs can handle. Rename runnable_avg_period into avg_period as it is now used with both runnable_avg_sum and running_avg_sum. Add some descriptions of the variables to explain their differences. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Ben Segall <bsegall@google.com> Cc: Ben Segall <bsegall@google.com> Cc: Morten.Rasmussen@arm.com Cc: Paul Turner <pjt@google.com> Cc: dietmar.eggemann@arm.com Cc: efault@gmx.de Cc: kamalesh@linux.vnet.ibm.com Cc: linaro-kernel@lists.linaro.org Cc: nicolas.pitre@linaro.org Cc: preeti@linux.vnet.ibm.com Cc: riel@redhat.com Link: http://lkml.kernel.org/r/1425052454-25797-2-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-02-27 15:54:04 +00:00
sched/fair: Update scale invariance of PELT The current implementation of load tracking invariance scales the contribution with current frequency and uarch performance (only for utilization) of the CPU. One main result of this formula is that the figures are capped by current capacity of CPU. Another one is that the load_avg is not invariant because not scaled with uarch. The util_avg of a periodic task that runs r time slots every p time slots varies in the range : U * (1-y^r)/(1-y^p) * y^i < Utilization < U * (1-y^r)/(1-y^p) with U is the max util_avg value = SCHED_CAPACITY_SCALE At a lower capacity, the range becomes: U * C * (1-y^r')/(1-y^p) * y^i' < Utilization < U * C * (1-y^r')/(1-y^p) with C reflecting the compute capacity ratio between current capacity and max capacity. so C tries to compensate changes in (1-y^r') but it can't be accurate. Instead of scaling the contribution value of PELT algo, we should scale the running time. The PELT signal aims to track the amount of computation of tasks and/or rq so it seems more correct to scale the running time to reflect the effective amount of computation done since the last update. In order to be fully invariant, we need to apply the same amount of running time and idle time whatever the current capacity. Because running at lower capacity implies that the task will run longer, we have to ensure that the same amount of idle time will be applied when system becomes idle and no idle time has been "stolen". But reaching the maximum utilization value (SCHED_CAPACITY_SCALE) means that the task is seen as an always-running task whatever the capacity of the CPU (even at max compute capacity). In this case, we can discard this "stolen" idle times which becomes meaningless. In order to achieve this time scaling, a new clock_pelt is created per rq. The increase of this clock scales with current capacity when something is running on rq and synchronizes with clock_task when rq is idle. With this mechanism, we ensure the same running and idle time whatever the current capacity. This also enables to simplify the pelt algorithm by removing all references of uarch and frequency and applying the same contribution to utilization and loads. Furthermore, the scaling is done only once per update of clock (update_rq_clock_task()) instead of during each update of sched_entities and cfs/rt/dl_rq of the rq like the current implementation. This is interesting when cgroup are involved as shown in the results below: On a hikey (octo Arm64 platform). Performance cpufreq governor and only shallowest c-state to remove variance generated by those power features so we only track the impact of pelt algo. each test runs 16 times: ./perf bench sched pipe (higher is better) kernel tip/sched/core + patch ops/seconds ops/seconds diff cgroup root 59652(+/- 0.18%) 59876(+/- 0.24%) +0.38% level1 55608(+/- 0.27%) 55923(+/- 0.24%) +0.57% level2 52115(+/- 0.29%) 52564(+/- 0.22%) +0.86% hackbench -l 1000 (lower is better) kernel tip/sched/core + patch duration(sec) duration(sec) diff cgroup root 4.453(+/- 2.37%) 4.383(+/- 2.88%) -1.57% level1 4.859(+/- 8.50%) 4.830(+/- 7.07%) -0.60% level2 5.063(+/- 9.83%) 4.928(+/- 9.66%) -2.66% Then, the responsiveness of PELT is improved when CPU is not running at max capacity with this new algorithm. I have put below some examples of duration to reach some typical load values according to the capacity of the CPU with current implementation and with this patch. These values has been computed based on the geometric series and the half period value: Util (%) max capacity half capacity(mainline) half capacity(w/ patch) 972 (95%) 138ms not reachable 276ms 486 (47.5%) 30ms 138ms 60ms 256 (25%) 13ms 32ms 26ms On my hikey (octo Arm64 platform) with schedutil governor, the time to reach max OPP when starting from a null utilization, decreases from 223ms with current scale invariance down to 121ms with the new algorithm. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: patrick.bellasi@arm.com Cc: pjt@google.com Cc: pkondeti@codeaurora.org Cc: quentin.perret@arm.com Cc: rjw@rjwysocki.net Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Link: https://lkml.kernel.org/r/1548257214-13745-3-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-01-23 15:26:53 +00:00
decayed |= __update_load_avg_cfs_rq(now, cfs_rq);
sched: Add sched_avg::utilization_avg_contrib Add new statistics which reflect the average time a task is running on the CPU and the sum of these running time of the tasks on a runqueue. The latter is named utilization_load_avg. This patch is based on the usage metric that was proposed in the 1st versions of the per-entity load tracking patchset by Paul Turner <pjt@google.com> but that has be removed afterwards. This version differs from the original one in the sense that it's not linked to task_group. The rq's utilization_load_avg will be used to check if a rq is overloaded or not instead of trying to compute how many tasks a group of CPUs can handle. Rename runnable_avg_period into avg_period as it is now used with both runnable_avg_sum and running_avg_sum. Add some descriptions of the variables to explain their differences. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Ben Segall <bsegall@google.com> Cc: Ben Segall <bsegall@google.com> Cc: Morten.Rasmussen@arm.com Cc: Paul Turner <pjt@google.com> Cc: dietmar.eggemann@arm.com Cc: efault@gmx.de Cc: kamalesh@linux.vnet.ibm.com Cc: linaro-kernel@lists.linaro.org Cc: nicolas.pitre@linaro.org Cc: preeti@linux.vnet.ibm.com Cc: riel@redhat.com Link: http://lkml.kernel.org/r/1425052454-25797-2-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-02-27 15:54:04 +00:00
sched/fair: Rewrite runnable load and utilization average tracking The idea of runnable load average (let runnable time contribute to weight) was proposed by Paul Turner and Ben Segall, and it is still followed by this rewrite. This rewrite aims to solve the following issues: 1. cfs_rq's load average (namely runnable_load_avg and blocked_load_avg) is updated at the granularity of an entity at a time, which results in the cfs_rq's load average is stale or partially updated: at any time, only one entity is up to date, all other entities are effectively lagging behind. This is undesirable. To illustrate, if we have n runnable entities in the cfs_rq, as time elapses, they certainly become outdated: t0: cfs_rq { e1_old, e2_old, ..., en_old } and when we update: t1: update e1, then we have cfs_rq { e1_new, e2_old, ..., en_old } t2: update e2, then we have cfs_rq { e1_old, e2_new, ..., en_old } ... We solve this by combining all runnable entities' load averages together in cfs_rq's avg, and update the cfs_rq's avg as a whole. This is based on the fact that if we regard the update as a function, then: w * update(e) = update(w * e) and update(e1) + update(e2) = update(e1 + e2), then w1 * update(e1) + w2 * update(e2) = update(w1 * e1 + w2 * e2) therefore, by this rewrite, we have an entirely updated cfs_rq at the time we update it: t1: update cfs_rq { e1_new, e2_new, ..., en_new } t2: update cfs_rq { e1_new, e2_new, ..., en_new } ... 2. cfs_rq's load average is different between top rq->cfs_rq and other task_group's per CPU cfs_rqs in whether or not blocked_load_average contributes to the load. The basic idea behind runnable load average (the same for utilization) is that the blocked state is taken into account as opposed to only accounting for the currently runnable state. Therefore, the average should include both the runnable/running and blocked load averages. This rewrite does that. In addition, we also combine runnable/running and blocked averages of all entities into the cfs_rq's average, and update it together at once. This is based on the fact that: update(runnable) + update(blocked) = update(runnable + blocked) This significantly reduces the code as we don't need to separately maintain/update runnable/running load and blocked load. 3. How task_group entities' share is calculated is complex and imprecise. We reduce the complexity in this rewrite to allow a very simple rule: the task_group's load_avg is aggregated from its per CPU cfs_rqs's load_avgs. Then group entity's weight is simply proportional to its own cfs_rq's load_avg / task_group's load_avg. To illustrate, if a task_group has { cfs_rq1, cfs_rq2, ..., cfs_rqn }, then, task_group_avg = cfs_rq1_avg + cfs_rq2_avg + ... + cfs_rqn_avg, then cfs_rqx's entity's share = cfs_rqx_avg / task_group_avg * task_group's share To sum up, this rewrite in principle is equivalent to the current one, but fixes the issues described above. Turns out, it significantly reduces the code complexity and hence increases clarity and efficiency. In addition, the new averages are more smooth/continuous (no spurious spikes and valleys) and updated more consistently and quickly to reflect the load dynamics. As a result, we have less load tracking overhead, better performance, and especially better power efficiency due to more balanced load. Signed-off-by: Yuyang Du <yuyang.du@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: arjan@linux.intel.com Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: fengguang.wu@intel.com Cc: len.brown@intel.com Cc: morten.rasmussen@arm.com Cc: pjt@google.com Cc: rafael.j.wysocki@intel.com Cc: umgwanakikbuti@gmail.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1436918682-4971-3-git-send-email-yuyang.du@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-07-15 00:04:37 +00:00
#ifndef CONFIG_64BIT
smp_wmb();
cfs_rq->load_last_update_time_copy = sa->last_update_time;
#endif
sched: Add sched_avg::utilization_avg_contrib Add new statistics which reflect the average time a task is running on the CPU and the sum of these running time of the tasks on a runqueue. The latter is named utilization_load_avg. This patch is based on the usage metric that was proposed in the 1st versions of the per-entity load tracking patchset by Paul Turner <pjt@google.com> but that has be removed afterwards. This version differs from the original one in the sense that it's not linked to task_group. The rq's utilization_load_avg will be used to check if a rq is overloaded or not instead of trying to compute how many tasks a group of CPUs can handle. Rename runnable_avg_period into avg_period as it is now used with both runnable_avg_sum and running_avg_sum. Add some descriptions of the variables to explain their differences. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Ben Segall <bsegall@google.com> Cc: Ben Segall <bsegall@google.com> Cc: Morten.Rasmussen@arm.com Cc: Paul Turner <pjt@google.com> Cc: dietmar.eggemann@arm.com Cc: efault@gmx.de Cc: kamalesh@linux.vnet.ibm.com Cc: linaro-kernel@lists.linaro.org Cc: nicolas.pitre@linaro.org Cc: preeti@linux.vnet.ibm.com Cc: riel@redhat.com Link: http://lkml.kernel.org/r/1425052454-25797-2-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-02-27 15:54:04 +00:00
return decayed;
}
/**
* attach_entity_load_avg - attach this entity to its cfs_rq load avg
* @cfs_rq: cfs_rq to attach to
* @se: sched_entity to attach
*
* Must call update_cfs_rq_load_avg() before this, since we rely on
* cfs_rq->avg.last_update_time being current.
*/
static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
{
u32 divider = LOAD_AVG_MAX - 1024 + cfs_rq->avg.period_contrib;
/*
* When we attach the @se to the @cfs_rq, we must align the decay
* window because without that, really weird and wonderful things can
* happen.
*
* XXX illustrate
*/
se->avg.last_update_time = cfs_rq->avg.last_update_time;
se->avg.period_contrib = cfs_rq->avg.period_contrib;
/*
* Hell(o) Nasty stuff.. we need to recompute _sum based on the new
* period_contrib. This isn't strictly correct, but since we're
* entirely outside of the PELT hierarchy, nobody cares if we truncate
* _sum a little.
*/
se->avg.util_sum = se->avg.util_avg * divider;
sched/pelt: Add a new runnable average signal Now that runnable_load_avg has been removed, we can replace it by a new signal that will highlight the runnable pressure on a cfs_rq. This signal track the waiting time of tasks on rq and can help to better define the state of rqs. At now, only util_avg is used to define the state of a rq: A rq with more that around 80% of utilization and more than 1 tasks is considered as overloaded. But the util_avg signal of a rq can become temporaly low after that a task migrated onto another rq which can bias the classification of the rq. When tasks compete for the same rq, their runnable average signal will be higher than util_avg as it will include the waiting time and we can use this signal to better classify cfs_rqs. The new runnable_avg will track the runnable time of a task which simply adds the waiting time to the running time. The runnable _avg of cfs_rq will be the /Sum of se's runnable_avg and the runnable_avg of group entity will follow the one of the rq similarly to util_avg. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: "Dietmar Eggemann <dietmar.eggemann@arm.com>" Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Valentin Schneider <valentin.schneider@arm.com> Cc: Phil Auld <pauld@redhat.com> Cc: Hillf Danton <hdanton@sina.com> Link: https://lore.kernel.org/r/20200224095223.13361-9-mgorman@techsingularity.net
2020-02-24 09:52:18 +00:00
se->avg.runnable_sum = se->avg.runnable_avg * divider;
se->avg.load_sum = divider;
if (se_weight(se)) {
se->avg.load_sum =
div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se));
}
enqueue_load_avg(cfs_rq, se);
cfs_rq->avg.util_avg += se->avg.util_avg;
cfs_rq->avg.util_sum += se->avg.util_sum;
sched/pelt: Add a new runnable average signal Now that runnable_load_avg has been removed, we can replace it by a new signal that will highlight the runnable pressure on a cfs_rq. This signal track the waiting time of tasks on rq and can help to better define the state of rqs. At now, only util_avg is used to define the state of a rq: A rq with more that around 80% of utilization and more than 1 tasks is considered as overloaded. But the util_avg signal of a rq can become temporaly low after that a task migrated onto another rq which can bias the classification of the rq. When tasks compete for the same rq, their runnable average signal will be higher than util_avg as it will include the waiting time and we can use this signal to better classify cfs_rqs. The new runnable_avg will track the runnable time of a task which simply adds the waiting time to the running time. The runnable _avg of cfs_rq will be the /Sum of se's runnable_avg and the runnable_avg of group entity will follow the one of the rq similarly to util_avg. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: "Dietmar Eggemann <dietmar.eggemann@arm.com>" Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Valentin Schneider <valentin.schneider@arm.com> Cc: Phil Auld <pauld@redhat.com> Cc: Hillf Danton <hdanton@sina.com> Link: https://lore.kernel.org/r/20200224095223.13361-9-mgorman@techsingularity.net
2020-02-24 09:52:18 +00:00
cfs_rq->avg.runnable_avg += se->avg.runnable_avg;
cfs_rq->avg.runnable_sum += se->avg.runnable_sum;
add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
cfs_rq_util_change(cfs_rq, 0);
trace_pelt_cfs_tp(cfs_rq);
}
/**
* detach_entity_load_avg - detach this entity from its cfs_rq load avg
* @cfs_rq: cfs_rq to detach from
* @se: sched_entity to detach
*
* Must call update_cfs_rq_load_avg() before this, since we rely on
* cfs_rq->avg.last_update_time being current.
*/
static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
{
dequeue_load_avg(cfs_rq, se);
sched/fair: Fix cfs_rq avg tracking underflow As per commit: b7fa30c9cc48 ("sched/fair: Fix post_init_entity_util_avg() serialization") > the code generated from update_cfs_rq_load_avg(): > > if (atomic_long_read(&cfs_rq->removed_load_avg)) { > s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0); > sa->load_avg = max_t(long, sa->load_avg - r, 0); > sa->load_sum = max_t(s64, sa->load_sum - r * LOAD_AVG_MAX, 0); > removed_load = 1; > } > > turns into: > > ffffffff81087064: 49 8b 85 98 00 00 00 mov 0x98(%r13),%rax > ffffffff8108706b: 48 85 c0 test %rax,%rax > ffffffff8108706e: 74 40 je ffffffff810870b0 <update_blocked_averages+0xc0> > ffffffff81087070: 4c 89 f8 mov %r15,%rax > ffffffff81087073: 49 87 85 98 00 00 00 xchg %rax,0x98(%r13) > ffffffff8108707a: 49 29 45 70 sub %rax,0x70(%r13) > ffffffff8108707e: 4c 89 f9 mov %r15,%rcx > ffffffff81087081: bb 01 00 00 00 mov $0x1,%ebx > ffffffff81087086: 49 83 7d 70 00 cmpq $0x0,0x70(%r13) > ffffffff8108708b: 49 0f 49 4d 70 cmovns 0x70(%r13),%rcx > > Which you'll note ends up with sa->load_avg -= r in memory at > ffffffff8108707a. So I _should_ have looked at other unserialized users of ->load_avg, but alas. Luckily nikbor reported a similar /0 from task_h_load() which instantly triggered recollection of this here problem. Aside from the intermediate value hitting memory and causing problems, there's another problem: the underflow detection relies on the signed bit. This reduces the effective width of the variables, IOW its effectively the same as having these variables be of signed type. This patch changes to a different means of unsigned underflow detection to not rely on the signed bit. This allows the variables to use the 'full' unsigned range. And it does so with explicit LOAD - STORE to ensure any intermediate value will never be visible in memory, allowing these unserialized loads. Note: GCC generates crap code for this, might warrant a look later. Note2: I say 'full' above, if we end up at U*_MAX we'll still explode; maybe we should do clamping on add too. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Chris Wilson <chris@chris-wilson.co.uk> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Yuyang Du <yuyang.du@intel.com> Cc: bsegall@google.com Cc: kernel@kyup.com Cc: morten.rasmussen@arm.com Cc: pjt@google.com Cc: steve.muckle@linaro.org Fixes: 9d89c257dfb9 ("sched/fair: Rewrite runnable load and utilization average tracking") Link: http://lkml.kernel.org/r/20160617091948.GJ30927@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-06-16 08:50:40 +00:00
sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
sched/pelt: Add a new runnable average signal Now that runnable_load_avg has been removed, we can replace it by a new signal that will highlight the runnable pressure on a cfs_rq. This signal track the waiting time of tasks on rq and can help to better define the state of rqs. At now, only util_avg is used to define the state of a rq: A rq with more that around 80% of utilization and more than 1 tasks is considered as overloaded. But the util_avg signal of a rq can become temporaly low after that a task migrated onto another rq which can bias the classification of the rq. When tasks compete for the same rq, their runnable average signal will be higher than util_avg as it will include the waiting time and we can use this signal to better classify cfs_rqs. The new runnable_avg will track the runnable time of a task which simply adds the waiting time to the running time. The runnable _avg of cfs_rq will be the /Sum of se's runnable_avg and the runnable_avg of group entity will follow the one of the rq similarly to util_avg. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: "Dietmar Eggemann <dietmar.eggemann@arm.com>" Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Valentin Schneider <valentin.schneider@arm.com> Cc: Phil Auld <pauld@redhat.com> Cc: Hillf Danton <hdanton@sina.com> Link: https://lore.kernel.org/r/20200224095223.13361-9-mgorman@techsingularity.net
2020-02-24 09:52:18 +00:00
sub_positive(&cfs_rq->avg.runnable_avg, se->avg.runnable_avg);
sub_positive(&cfs_rq->avg.runnable_sum, se->avg.runnable_sum);
add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
cfs_rq_util_change(cfs_rq, 0);
trace_pelt_cfs_tp(cfs_rq);
}
/*
* Optional action to be done while updating the load average
*/
#define UPDATE_TG 0x1
#define SKIP_AGE_LOAD 0x2
#define DO_ATTACH 0x4
/* Update task and its cfs_rq load average */
static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
{
sched/fair: Update scale invariance of PELT The current implementation of load tracking invariance scales the contribution with current frequency and uarch performance (only for utilization) of the CPU. One main result of this formula is that the figures are capped by current capacity of CPU. Another one is that the load_avg is not invariant because not scaled with uarch. The util_avg of a periodic task that runs r time slots every p time slots varies in the range : U * (1-y^r)/(1-y^p) * y^i < Utilization < U * (1-y^r)/(1-y^p) with U is the max util_avg value = SCHED_CAPACITY_SCALE At a lower capacity, the range becomes: U * C * (1-y^r')/(1-y^p) * y^i' < Utilization < U * C * (1-y^r')/(1-y^p) with C reflecting the compute capacity ratio between current capacity and max capacity. so C tries to compensate changes in (1-y^r') but it can't be accurate. Instead of scaling the contribution value of PELT algo, we should scale the running time. The PELT signal aims to track the amount of computation of tasks and/or rq so it seems more correct to scale the running time to reflect the effective amount of computation done since the last update. In order to be fully invariant, we need to apply the same amount of running time and idle time whatever the current capacity. Because running at lower capacity implies that the task will run longer, we have to ensure that the same amount of idle time will be applied when system becomes idle and no idle time has been "stolen". But reaching the maximum utilization value (SCHED_CAPACITY_SCALE) means that the task is seen as an always-running task whatever the capacity of the CPU (even at max compute capacity). In this case, we can discard this "stolen" idle times which becomes meaningless. In order to achieve this time scaling, a new clock_pelt is created per rq. The increase of this clock scales with current capacity when something is running on rq and synchronizes with clock_task when rq is idle. With this mechanism, we ensure the same running and idle time whatever the current capacity. This also enables to simplify the pelt algorithm by removing all references of uarch and frequency and applying the same contribution to utilization and loads. Furthermore, the scaling is done only once per update of clock (update_rq_clock_task()) instead of during each update of sched_entities and cfs/rt/dl_rq of the rq like the current implementation. This is interesting when cgroup are involved as shown in the results below: On a hikey (octo Arm64 platform). Performance cpufreq governor and only shallowest c-state to remove variance generated by those power features so we only track the impact of pelt algo. each test runs 16 times: ./perf bench sched pipe (higher is better) kernel tip/sched/core + patch ops/seconds ops/seconds diff cgroup root 59652(+/- 0.18%) 59876(+/- 0.24%) +0.38% level1 55608(+/- 0.27%) 55923(+/- 0.24%) +0.57% level2 52115(+/- 0.29%) 52564(+/- 0.22%) +0.86% hackbench -l 1000 (lower is better) kernel tip/sched/core + patch duration(sec) duration(sec) diff cgroup root 4.453(+/- 2.37%) 4.383(+/- 2.88%) -1.57% level1 4.859(+/- 8.50%) 4.830(+/- 7.07%) -0.60% level2 5.063(+/- 9.83%) 4.928(+/- 9.66%) -2.66% Then, the responsiveness of PELT is improved when CPU is not running at max capacity with this new algorithm. I have put below some examples of duration to reach some typical load values according to the capacity of the CPU with current implementation and with this patch. These values has been computed based on the geometric series and the half period value: Util (%) max capacity half capacity(mainline) half capacity(w/ patch) 972 (95%) 138ms not reachable 276ms 486 (47.5%) 30ms 138ms 60ms 256 (25%) 13ms 32ms 26ms On my hikey (octo Arm64 platform) with schedutil governor, the time to reach max OPP when starting from a null utilization, decreases from 223ms with current scale invariance down to 121ms with the new algorithm. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: patrick.bellasi@arm.com Cc: pjt@google.com Cc: pkondeti@codeaurora.org Cc: quentin.perret@arm.com Cc: rjw@rjwysocki.net Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Link: https://lkml.kernel.org/r/1548257214-13745-3-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-01-23 15:26:53 +00:00
u64 now = cfs_rq_clock_pelt(cfs_rq);
int decayed;
/*
* Track task load average for carrying it to new CPU after migrated, and
* track group sched_entity load average for task_h_load calc in migration
*/
if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
sched/fair: Update scale invariance of PELT The current implementation of load tracking invariance scales the contribution with current frequency and uarch performance (only for utilization) of the CPU. One main result of this formula is that the figures are capped by current capacity of CPU. Another one is that the load_avg is not invariant because not scaled with uarch. The util_avg of a periodic task that runs r time slots every p time slots varies in the range : U * (1-y^r)/(1-y^p) * y^i < Utilization < U * (1-y^r)/(1-y^p) with U is the max util_avg value = SCHED_CAPACITY_SCALE At a lower capacity, the range becomes: U * C * (1-y^r')/(1-y^p) * y^i' < Utilization < U * C * (1-y^r')/(1-y^p) with C reflecting the compute capacity ratio between current capacity and max capacity. so C tries to compensate changes in (1-y^r') but it can't be accurate. Instead of scaling the contribution value of PELT algo, we should scale the running time. The PELT signal aims to track the amount of computation of tasks and/or rq so it seems more correct to scale the running time to reflect the effective amount of computation done since the last update. In order to be fully invariant, we need to apply the same amount of running time and idle time whatever the current capacity. Because running at lower capacity implies that the task will run longer, we have to ensure that the same amount of idle time will be applied when system becomes idle and no idle time has been "stolen". But reaching the maximum utilization value (SCHED_CAPACITY_SCALE) means that the task is seen as an always-running task whatever the capacity of the CPU (even at max compute capacity). In this case, we can discard this "stolen" idle times which becomes meaningless. In order to achieve this time scaling, a new clock_pelt is created per rq. The increase of this clock scales with current capacity when something is running on rq and synchronizes with clock_task when rq is idle. With this mechanism, we ensure the same running and idle time whatever the current capacity. This also enables to simplify the pelt algorithm by removing all references of uarch and frequency and applying the same contribution to utilization and loads. Furthermore, the scaling is done only once per update of clock (update_rq_clock_task()) instead of during each update of sched_entities and cfs/rt/dl_rq of the rq like the current implementation. This is interesting when cgroup are involved as shown in the results below: On a hikey (octo Arm64 platform). Performance cpufreq governor and only shallowest c-state to remove variance generated by those power features so we only track the impact of pelt algo. each test runs 16 times: ./perf bench sched pipe (higher is better) kernel tip/sched/core + patch ops/seconds ops/seconds diff cgroup root 59652(+/- 0.18%) 59876(+/- 0.24%) +0.38% level1 55608(+/- 0.27%) 55923(+/- 0.24%) +0.57% level2 52115(+/- 0.29%) 52564(+/- 0.22%) +0.86% hackbench -l 1000 (lower is better) kernel tip/sched/core + patch duration(sec) duration(sec) diff cgroup root 4.453(+/- 2.37%) 4.383(+/- 2.88%) -1.57% level1 4.859(+/- 8.50%) 4.830(+/- 7.07%) -0.60% level2 5.063(+/- 9.83%) 4.928(+/- 9.66%) -2.66% Then, the responsiveness of PELT is improved when CPU is not running at max capacity with this new algorithm. I have put below some examples of duration to reach some typical load values according to the capacity of the CPU with current implementation and with this patch. These values has been computed based on the geometric series and the half period value: Util (%) max capacity half capacity(mainline) half capacity(w/ patch) 972 (95%) 138ms not reachable 276ms 486 (47.5%) 30ms 138ms 60ms 256 (25%) 13ms 32ms 26ms On my hikey (octo Arm64 platform) with schedutil governor, the time to reach max OPP when starting from a null utilization, decreases from 223ms with current scale invariance down to 121ms with the new algorithm. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: patrick.bellasi@arm.com Cc: pjt@google.com Cc: pkondeti@codeaurora.org Cc: quentin.perret@arm.com Cc: rjw@rjwysocki.net Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Link: https://lkml.kernel.org/r/1548257214-13745-3-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-01-23 15:26:53 +00:00
__update_load_avg_se(now, cfs_rq, se);
decayed = update_cfs_rq_load_avg(now, cfs_rq);
decayed |= propagate_entity_load_avg(se);
if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
/*
* DO_ATTACH means we're here from enqueue_entity().
* !last_update_time means we've passed through
* migrate_task_rq_fair() indicating we migrated.
*
* IOW we're enqueueing a task on a new CPU.
*/
attach_entity_load_avg(cfs_rq, se);
update_tg_load_avg(cfs_rq, 0);
sched/cpufreq: Move the cfs_rq_util_change() call to cpufreq_update_util() update_cfs_rq_load_avg() calls cfs_rq_util_change() every time PELT decays, which might be inefficient when the cpufreq driver has rate limitation. When a task is attached on a CPU, we have this call path: update_load_avg() update_cfs_rq_load_avg() cfs_rq_util_change -- > trig frequency update attach_entity_load_avg() cfs_rq_util_change -- > trig frequency update The 1st frequency update will not take into account the utilization of the newly attached task and the 2nd one might be discarded because of rate limitation of the cpufreq driver. update_cfs_rq_load_avg() is only called by update_blocked_averages() and update_load_avg() so we can move the call to cfs_rq_util_change/cpufreq_update_util() into these two functions. It's also interesting to note that update_load_avg() already calls cfs_rq_util_change() directly for the !SMP case. This change will also ensure that cpufreq_update_util() is called even when there is no more CFS rq in the leaf_cfs_rq_list to update, but only IRQ, RT or DL PELT signals. [ mingo: Minor updates. ] Reported-by: Doug Smythies <dsmythies@telus.net> Tested-by: Doug Smythies <dsmythies@telus.net> Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Acked-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: juri.lelli@redhat.com Cc: linux-pm@vger.kernel.org Cc: mgorman@suse.de Cc: rostedt@goodmis.org Cc: sargun@sargun.me Cc: srinivas.pandruvada@linux.intel.com Cc: tj@kernel.org Cc: xiexiuqi@huawei.com Cc: xiezhipeng1@huawei.com Fixes: 039ae8bcf7a5 ("sched/fair: Fix O(nr_cgroups) in the load balancing path") Link: https://lkml.kernel.org/r/1574083279-799-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-11-18 13:21:19 +00:00
} else if (decayed) {
cfs_rq_util_change(cfs_rq, 0);
if (flags & UPDATE_TG)
update_tg_load_avg(cfs_rq, 0);
}
}
sched/fair: Rewrite runnable load and utilization average tracking The idea of runnable load average (let runnable time contribute to weight) was proposed by Paul Turner and Ben Segall, and it is still followed by this rewrite. This rewrite aims to solve the following issues: 1. cfs_rq's load average (namely runnable_load_avg and blocked_load_avg) is updated at the granularity of an entity at a time, which results in the cfs_rq's load average is stale or partially updated: at any time, only one entity is up to date, all other entities are effectively lagging behind. This is undesirable. To illustrate, if we have n runnable entities in the cfs_rq, as time elapses, they certainly become outdated: t0: cfs_rq { e1_old, e2_old, ..., en_old } and when we update: t1: update e1, then we have cfs_rq { e1_new, e2_old, ..., en_old } t2: update e2, then we have cfs_rq { e1_old, e2_new, ..., en_old } ... We solve this by combining all runnable entities' load averages together in cfs_rq's avg, and update the cfs_rq's avg as a whole. This is based on the fact that if we regard the update as a function, then: w * update(e) = update(w * e) and update(e1) + update(e2) = update(e1 + e2), then w1 * update(e1) + w2 * update(e2) = update(w1 * e1 + w2 * e2) therefore, by this rewrite, we have an entirely updated cfs_rq at the time we update it: t1: update cfs_rq { e1_new, e2_new, ..., en_new } t2: update cfs_rq { e1_new, e2_new, ..., en_new } ... 2. cfs_rq's load average is different between top rq->cfs_rq and other task_group's per CPU cfs_rqs in whether or not blocked_load_average contributes to the load. The basic idea behind runnable load average (the same for utilization) is that the blocked state is taken into account as opposed to only accounting for the currently runnable state. Therefore, the average should include both the runnable/running and blocked load averages. This rewrite does that. In addition, we also combine runnable/running and blocked averages of all entities into the cfs_rq's average, and update it together at once. This is based on the fact that: update(runnable) + update(blocked) = update(runnable + blocked) This significantly reduces the code as we don't need to separately maintain/update runnable/running load and blocked load. 3. How task_group entities' share is calculated is complex and imprecise. We reduce the complexity in this rewrite to allow a very simple rule: the task_group's load_avg is aggregated from its per CPU cfs_rqs's load_avgs. Then group entity's weight is simply proportional to its own cfs_rq's load_avg / task_group's load_avg. To illustrate, if a task_group has { cfs_rq1, cfs_rq2, ..., cfs_rqn }, then, task_group_avg = cfs_rq1_avg + cfs_rq2_avg + ... + cfs_rqn_avg, then cfs_rqx's entity's share = cfs_rqx_avg / task_group_avg * task_group's share To sum up, this rewrite in principle is equivalent to the current one, but fixes the issues described above. Turns out, it significantly reduces the code complexity and hence increases clarity and efficiency. In addition, the new averages are more smooth/continuous (no spurious spikes and valleys) and updated more consistently and quickly to reflect the load dynamics. As a result, we have less load tracking overhead, better performance, and especially better power efficiency due to more balanced load. Signed-off-by: Yuyang Du <yuyang.du@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: arjan@linux.intel.com Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: fengguang.wu@intel.com Cc: len.brown@intel.com Cc: morten.rasmussen@arm.com Cc: pjt@google.com Cc: rafael.j.wysocki@intel.com Cc: umgwanakikbuti@gmail.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1436918682-4971-3-git-send-email-yuyang.du@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-07-15 00:04:37 +00:00
#ifndef CONFIG_64BIT
static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
{
sched/fair: Rewrite runnable load and utilization average tracking The idea of runnable load average (let runnable time contribute to weight) was proposed by Paul Turner and Ben Segall, and it is still followed by this rewrite. This rewrite aims to solve the following issues: 1. cfs_rq's load average (namely runnable_load_avg and blocked_load_avg) is updated at the granularity of an entity at a time, which results in the cfs_rq's load average is stale or partially updated: at any time, only one entity is up to date, all other entities are effectively lagging behind. This is undesirable. To illustrate, if we have n runnable entities in the cfs_rq, as time elapses, they certainly become outdated: t0: cfs_rq { e1_old, e2_old, ..., en_old } and when we update: t1: update e1, then we have cfs_rq { e1_new, e2_old, ..., en_old } t2: update e2, then we have cfs_rq { e1_old, e2_new, ..., en_old } ... We solve this by combining all runnable entities' load averages together in cfs_rq's avg, and update the cfs_rq's avg as a whole. This is based on the fact that if we regard the update as a function, then: w * update(e) = update(w * e) and update(e1) + update(e2) = update(e1 + e2), then w1 * update(e1) + w2 * update(e2) = update(w1 * e1 + w2 * e2) therefore, by this rewrite, we have an entirely updated cfs_rq at the time we update it: t1: update cfs_rq { e1_new, e2_new, ..., en_new } t2: update cfs_rq { e1_new, e2_new, ..., en_new } ... 2. cfs_rq's load average is different between top rq->cfs_rq and other task_group's per CPU cfs_rqs in whether or not blocked_load_average contributes to the load. The basic idea behind runnable load average (the same for utilization) is that the blocked state is taken into account as opposed to only accounting for the currently runnable state. Therefore, the average should include both the runnable/running and blocked load averages. This rewrite does that. In addition, we also combine runnable/running and blocked averages of all entities into the cfs_rq's average, and update it together at once. This is based on the fact that: update(runnable) + update(blocked) = update(runnable + blocked) This significantly reduces the code as we don't need to separately maintain/update runnable/running load and blocked load. 3. How task_group entities' share is calculated is complex and imprecise. We reduce the complexity in this rewrite to allow a very simple rule: the task_group's load_avg is aggregated from its per CPU cfs_rqs's load_avgs. Then group entity's weight is simply proportional to its own cfs_rq's load_avg / task_group's load_avg. To illustrate, if a task_group has { cfs_rq1, cfs_rq2, ..., cfs_rqn }, then, task_group_avg = cfs_rq1_avg + cfs_rq2_avg + ... + cfs_rqn_avg, then cfs_rqx's entity's share = cfs_rqx_avg / task_group_avg * task_group's share To sum up, this rewrite in principle is equivalent to the current one, but fixes the issues described above. Turns out, it significantly reduces the code complexity and hence increases clarity and efficiency. In addition, the new averages are more smooth/continuous (no spurious spikes and valleys) and updated more consistently and quickly to reflect the load dynamics. As a result, we have less load tracking overhead, better performance, and especially better power efficiency due to more balanced load. Signed-off-by: Yuyang Du <yuyang.du@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: arjan@linux.intel.com Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: fengguang.wu@intel.com Cc: len.brown@intel.com Cc: morten.rasmussen@arm.com Cc: pjt@google.com Cc: rafael.j.wysocki@intel.com Cc: umgwanakikbuti@gmail.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1436918682-4971-3-git-send-email-yuyang.du@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-07-15 00:04:37 +00:00
u64 last_update_time_copy;
u64 last_update_time;
sched/fair: Rewrite runnable load and utilization average tracking The idea of runnable load average (let runnable time contribute to weight) was proposed by Paul Turner and Ben Segall, and it is still followed by this rewrite. This rewrite aims to solve the following issues: 1. cfs_rq's load average (namely runnable_load_avg and blocked_load_avg) is updated at the granularity of an entity at a time, which results in the cfs_rq's load average is stale or partially updated: at any time, only one entity is up to date, all other entities are effectively lagging behind. This is undesirable. To illustrate, if we have n runnable entities in the cfs_rq, as time elapses, they certainly become outdated: t0: cfs_rq { e1_old, e2_old, ..., en_old } and when we update: t1: update e1, then we have cfs_rq { e1_new, e2_old, ..., en_old } t2: update e2, then we have cfs_rq { e1_old, e2_new, ..., en_old } ... We solve this by combining all runnable entities' load averages together in cfs_rq's avg, and update the cfs_rq's avg as a whole. This is based on the fact that if we regard the update as a function, then: w * update(e) = update(w * e) and update(e1) + update(e2) = update(e1 + e2), then w1 * update(e1) + w2 * update(e2) = update(w1 * e1 + w2 * e2) therefore, by this rewrite, we have an entirely updated cfs_rq at the time we update it: t1: update cfs_rq { e1_new, e2_new, ..., en_new } t2: update cfs_rq { e1_new, e2_new, ..., en_new } ... 2. cfs_rq's load average is different between top rq->cfs_rq and other task_group's per CPU cfs_rqs in whether or not blocked_load_average contributes to the load. The basic idea behind runnable load average (the same for utilization) is that the blocked state is taken into account as opposed to only accounting for the currently runnable state. Therefore, the average should include both the runnable/running and blocked load averages. This rewrite does that. In addition, we also combine runnable/running and blocked averages of all entities into the cfs_rq's average, and update it together at once. This is based on the fact that: update(runnable) + update(blocked) = update(runnable + blocked) This significantly reduces the code as we don't need to separately maintain/update runnable/running load and blocked load. 3. How task_group entities' share is calculated is complex and imprecise. We reduce the complexity in this rewrite to allow a very simple rule: the task_group's load_avg is aggregated from its per CPU cfs_rqs's load_avgs. Then group entity's weight is simply proportional to its own cfs_rq's load_avg / task_group's load_avg. To illustrate, if a task_group has { cfs_rq1, cfs_rq2, ..., cfs_rqn }, then, task_group_avg = cfs_rq1_avg + cfs_rq2_avg + ... + cfs_rqn_avg, then cfs_rqx's entity's share = cfs_rqx_avg / task_group_avg * task_group's share To sum up, this rewrite in principle is equivalent to the current one, but fixes the issues described above. Turns out, it significantly reduces the code complexity and hence increases clarity and efficiency. In addition, the new averages are more smooth/continuous (no spurious spikes and valleys) and updated more consistently and quickly to reflect the load dynamics. As a result, we have less load tracking overhead, better performance, and especially better power efficiency due to more balanced load. Signed-off-by: Yuyang Du <yuyang.du@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: arjan@linux.intel.com Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: fengguang.wu@intel.com Cc: len.brown@intel.com Cc: morten.rasmussen@arm.com Cc: pjt@google.com Cc: rafael.j.wysocki@intel.com Cc: umgwanakikbuti@gmail.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1436918682-4971-3-git-send-email-yuyang.du@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-07-15 00:04:37 +00:00
do {
last_update_time_copy = cfs_rq->load_last_update_time_copy;
smp_rmb();
last_update_time = cfs_rq->avg.last_update_time;
} while (last_update_time != last_update_time_copy);
return last_update_time;
}
sched/fair: Rewrite runnable load and utilization average tracking The idea of runnable load average (let runnable time contribute to weight) was proposed by Paul Turner and Ben Segall, and it is still followed by this rewrite. This rewrite aims to solve the following issues: 1. cfs_rq's load average (namely runnable_load_avg and blocked_load_avg) is updated at the granularity of an entity at a time, which results in the cfs_rq's load average is stale or partially updated: at any time, only one entity is up to date, all other entities are effectively lagging behind. This is undesirable. To illustrate, if we have n runnable entities in the cfs_rq, as time elapses, they certainly become outdated: t0: cfs_rq { e1_old, e2_old, ..., en_old } and when we update: t1: update e1, then we have cfs_rq { e1_new, e2_old, ..., en_old } t2: update e2, then we have cfs_rq { e1_old, e2_new, ..., en_old } ... We solve this by combining all runnable entities' load averages together in cfs_rq's avg, and update the cfs_rq's avg as a whole. This is based on the fact that if we regard the update as a function, then: w * update(e) = update(w * e) and update(e1) + update(e2) = update(e1 + e2), then w1 * update(e1) + w2 * update(e2) = update(w1 * e1 + w2 * e2) therefore, by this rewrite, we have an entirely updated cfs_rq at the time we update it: t1: update cfs_rq { e1_new, e2_new, ..., en_new } t2: update cfs_rq { e1_new, e2_new, ..., en_new } ... 2. cfs_rq's load average is different between top rq->cfs_rq and other task_group's per CPU cfs_rqs in whether or not blocked_load_average contributes to the load. The basic idea behind runnable load average (the same for utilization) is that the blocked state is taken into account as opposed to only accounting for the currently runnable state. Therefore, the average should include both the runnable/running and blocked load averages. This rewrite does that. In addition, we also combine runnable/running and blocked averages of all entities into the cfs_rq's average, and update it together at once. This is based on the fact that: update(runnable) + update(blocked) = update(runnable + blocked) This significantly reduces the code as we don't need to separately maintain/update runnable/running load and blocked load. 3. How task_group entities' share is calculated is complex and imprecise. We reduce the complexity in this rewrite to allow a very simple rule: the task_group's load_avg is aggregated from its per CPU cfs_rqs's load_avgs. Then group entity's weight is simply proportional to its own cfs_rq's load_avg / task_group's load_avg. To illustrate, if a task_group has { cfs_rq1, cfs_rq2, ..., cfs_rqn }, then, task_group_avg = cfs_rq1_avg + cfs_rq2_avg + ... + cfs_rqn_avg, then cfs_rqx's entity's share = cfs_rqx_avg / task_group_avg * task_group's share To sum up, this rewrite in principle is equivalent to the current one, but fixes the issues described above. Turns out, it significantly reduces the code complexity and hence increases clarity and efficiency. In addition, the new averages are more smooth/continuous (no spurious spikes and valleys) and updated more consistently and quickly to reflect the load dynamics. As a result, we have less load tracking overhead, better performance, and especially better power efficiency due to more balanced load. Signed-off-by: Yuyang Du <yuyang.du@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: arjan@linux.intel.com Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: fengguang.wu@intel.com Cc: len.brown@intel.com Cc: morten.rasmussen@arm.com Cc: pjt@google.com Cc: rafael.j.wysocki@intel.com Cc: umgwanakikbuti@gmail.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1436918682-4971-3-git-send-email-yuyang.du@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-07-15 00:04:37 +00:00
#else
static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
{
return cfs_rq->avg.last_update_time;
}
sched/fair: Rewrite runnable load and utilization average tracking The idea of runnable load average (let runnable time contribute to weight) was proposed by Paul Turner and Ben Segall, and it is still followed by this rewrite. This rewrite aims to solve the following issues: 1. cfs_rq's load average (namely runnable_load_avg and blocked_load_avg) is updated at the granularity of an entity at a time, which results in the cfs_rq's load average is stale or partially updated: at any time, only one entity is up to date, all other entities are effectively lagging behind. This is undesirable. To illustrate, if we have n runnable entities in the cfs_rq, as time elapses, they certainly become outdated: t0: cfs_rq { e1_old, e2_old, ..., en_old } and when we update: t1: update e1, then we have cfs_rq { e1_new, e2_old, ..., en_old } t2: update e2, then we have cfs_rq { e1_old, e2_new, ..., en_old } ... We solve this by combining all runnable entities' load averages together in cfs_rq's avg, and update the cfs_rq's avg as a whole. This is based on the fact that if we regard the update as a function, then: w * update(e) = update(w * e) and update(e1) + update(e2) = update(e1 + e2), then w1 * update(e1) + w2 * update(e2) = update(w1 * e1 + w2 * e2) therefore, by this rewrite, we have an entirely updated cfs_rq at the time we update it: t1: update cfs_rq { e1_new, e2_new, ..., en_new } t2: update cfs_rq { e1_new, e2_new, ..., en_new } ... 2. cfs_rq's load average is different between top rq->cfs_rq and other task_group's per CPU cfs_rqs in whether or not blocked_load_average contributes to the load. The basic idea behind runnable load average (the same for utilization) is that the blocked state is taken into account as opposed to only accounting for the currently runnable state. Therefore, the average should include both the runnable/running and blocked load averages. This rewrite does that. In addition, we also combine runnable/running and blocked averages of all entities into the cfs_rq's average, and update it together at once. This is based on the fact that: update(runnable) + update(blocked) = update(runnable + blocked) This significantly reduces the code as we don't need to separately maintain/update runnable/running load and blocked load. 3. How task_group entities' share is calculated is complex and imprecise. We reduce the complexity in this rewrite to allow a very simple rule: the task_group's load_avg is aggregated from its per CPU cfs_rqs's load_avgs. Then group entity's weight is simply proportional to its own cfs_rq's load_avg / task_group's load_avg. To illustrate, if a task_group has { cfs_rq1, cfs_rq2, ..., cfs_rqn }, then, task_group_avg = cfs_rq1_avg + cfs_rq2_avg + ... + cfs_rqn_avg, then cfs_rqx's entity's share = cfs_rqx_avg / task_group_avg * task_group's share To sum up, this rewrite in principle is equivalent to the current one, but fixes the issues described above. Turns out, it significantly reduces the code complexity and hence increases clarity and efficiency. In addition, the new averages are more smooth/continuous (no spurious spikes and valleys) and updated more consistently and quickly to reflect the load dynamics. As a result, we have less load tracking overhead, better performance, and especially better power efficiency due to more balanced load. Signed-off-by: Yuyang Du <yuyang.du@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: arjan@linux.intel.com Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: fengguang.wu@intel.com Cc: len.brown@intel.com Cc: morten.rasmussen@arm.com Cc: pjt@google.com Cc: rafael.j.wysocki@intel.com Cc: umgwanakikbuti@gmail.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1436918682-4971-3-git-send-email-yuyang.du@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-07-15 00:04:37 +00:00
#endif
sched/fair: Compute task/cpu utilization at wake-up correctly At task wake-up load-tracking isn't updated until the task is enqueued. The task's own view of its utilization contribution may therefore not be aligned with its contribution to the cfs_rq load-tracking which may have been updated in the meantime. Basically, the task's own utilization hasn't yet accounted for the sleep decay, while the cfs_rq may have (partially). Estimating the cfs_rq utilization in case the task is migrated at wake-up as task_rq(p)->cfs.avg.util_avg - p->se.avg.util_avg is therefore incorrect as the two load-tracking signals aren't time synchronized (different last update). To solve this problem, this patch synchronizes the task utilization with its previous rq before the task utilization is used in the wake-up path. Currently the update/synchronization is done _after_ the task has been placed by select_task_rq_fair(). The synchronization is done without having to take the rq lock using the existing mechanism used in remove_entity_load_avg(). Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: freedom.tan@mediatek.com Cc: keita.kobayashi.ym@renesas.com Cc: mgalbraith@suse.de Cc: sgurrappadi@nvidia.com Cc: vincent.guittot@linaro.org Cc: yuyang.du@intel.com Link: http://lkml.kernel.org/r/1476452472-24740-2-git-send-email-morten.rasmussen@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-10-14 13:41:07 +00:00
/*
* Synchronize entity load avg of dequeued entity without locking
* the previous rq.
*/
static void sync_entity_load_avg(struct sched_entity *se)
sched/fair: Compute task/cpu utilization at wake-up correctly At task wake-up load-tracking isn't updated until the task is enqueued. The task's own view of its utilization contribution may therefore not be aligned with its contribution to the cfs_rq load-tracking which may have been updated in the meantime. Basically, the task's own utilization hasn't yet accounted for the sleep decay, while the cfs_rq may have (partially). Estimating the cfs_rq utilization in case the task is migrated at wake-up as task_rq(p)->cfs.avg.util_avg - p->se.avg.util_avg is therefore incorrect as the two load-tracking signals aren't time synchronized (different last update). To solve this problem, this patch synchronizes the task utilization with its previous rq before the task utilization is used in the wake-up path. Currently the update/synchronization is done _after_ the task has been placed by select_task_rq_fair(). The synchronization is done without having to take the rq lock using the existing mechanism used in remove_entity_load_avg(). Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: freedom.tan@mediatek.com Cc: keita.kobayashi.ym@renesas.com Cc: mgalbraith@suse.de Cc: sgurrappadi@nvidia.com Cc: vincent.guittot@linaro.org Cc: yuyang.du@intel.com Link: http://lkml.kernel.org/r/1476452472-24740-2-git-send-email-morten.rasmussen@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-10-14 13:41:07 +00:00
{
struct cfs_rq *cfs_rq = cfs_rq_of(se);
u64 last_update_time;
last_update_time = cfs_rq_last_update_time(cfs_rq);
sched/fair: Update scale invariance of PELT The current implementation of load tracking invariance scales the contribution with current frequency and uarch performance (only for utilization) of the CPU. One main result of this formula is that the figures are capped by current capacity of CPU. Another one is that the load_avg is not invariant because not scaled with uarch. The util_avg of a periodic task that runs r time slots every p time slots varies in the range : U * (1-y^r)/(1-y^p) * y^i < Utilization < U * (1-y^r)/(1-y^p) with U is the max util_avg value = SCHED_CAPACITY_SCALE At a lower capacity, the range becomes: U * C * (1-y^r')/(1-y^p) * y^i' < Utilization < U * C * (1-y^r')/(1-y^p) with C reflecting the compute capacity ratio between current capacity and max capacity. so C tries to compensate changes in (1-y^r') but it can't be accurate. Instead of scaling the contribution value of PELT algo, we should scale the running time. The PELT signal aims to track the amount of computation of tasks and/or rq so it seems more correct to scale the running time to reflect the effective amount of computation done since the last update. In order to be fully invariant, we need to apply the same amount of running time and idle time whatever the current capacity. Because running at lower capacity implies that the task will run longer, we have to ensure that the same amount of idle time will be applied when system becomes idle and no idle time has been "stolen". But reaching the maximum utilization value (SCHED_CAPACITY_SCALE) means that the task is seen as an always-running task whatever the capacity of the CPU (even at max compute capacity). In this case, we can discard this "stolen" idle times which becomes meaningless. In order to achieve this time scaling, a new clock_pelt is created per rq. The increase of this clock scales with current capacity when something is running on rq and synchronizes with clock_task when rq is idle. With this mechanism, we ensure the same running and idle time whatever the current capacity. This also enables to simplify the pelt algorithm by removing all references of uarch and frequency and applying the same contribution to utilization and loads. Furthermore, the scaling is done only once per update of clock (update_rq_clock_task()) instead of during each update of sched_entities and cfs/rt/dl_rq of the rq like the current implementation. This is interesting when cgroup are involved as shown in the results below: On a hikey (octo Arm64 platform). Performance cpufreq governor and only shallowest c-state to remove variance generated by those power features so we only track the impact of pelt algo. each test runs 16 times: ./perf bench sched pipe (higher is better) kernel tip/sched/core + patch ops/seconds ops/seconds diff cgroup root 59652(+/- 0.18%) 59876(+/- 0.24%) +0.38% level1 55608(+/- 0.27%) 55923(+/- 0.24%) +0.57% level2 52115(+/- 0.29%) 52564(+/- 0.22%) +0.86% hackbench -l 1000 (lower is better) kernel tip/sched/core + patch duration(sec) duration(sec) diff cgroup root 4.453(+/- 2.37%) 4.383(+/- 2.88%) -1.57% level1 4.859(+/- 8.50%) 4.830(+/- 7.07%) -0.60% level2 5.063(+/- 9.83%) 4.928(+/- 9.66%) -2.66% Then, the responsiveness of PELT is improved when CPU is not running at max capacity with this new algorithm. I have put below some examples of duration to reach some typical load values according to the capacity of the CPU with current implementation and with this patch. These values has been computed based on the geometric series and the half period value: Util (%) max capacity half capacity(mainline) half capacity(w/ patch) 972 (95%) 138ms not reachable 276ms 486 (47.5%) 30ms 138ms 60ms 256 (25%) 13ms 32ms 26ms On my hikey (octo Arm64 platform) with schedutil governor, the time to reach max OPP when starting from a null utilization, decreases from 223ms with current scale invariance down to 121ms with the new algorithm. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: patrick.bellasi@arm.com Cc: pjt@google.com Cc: pkondeti@codeaurora.org Cc: quentin.perret@arm.com Cc: rjw@rjwysocki.net Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Link: https://lkml.kernel.org/r/1548257214-13745-3-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-01-23 15:26:53 +00:00
__update_load_avg_blocked_se(last_update_time, se);
sched/fair: Compute task/cpu utilization at wake-up correctly At task wake-up load-tracking isn't updated until the task is enqueued. The task's own view of its utilization contribution may therefore not be aligned with its contribution to the cfs_rq load-tracking which may have been updated in the meantime. Basically, the task's own utilization hasn't yet accounted for the sleep decay, while the cfs_rq may have (partially). Estimating the cfs_rq utilization in case the task is migrated at wake-up as task_rq(p)->cfs.avg.util_avg - p->se.avg.util_avg is therefore incorrect as the two load-tracking signals aren't time synchronized (different last update). To solve this problem, this patch synchronizes the task utilization with its previous rq before the task utilization is used in the wake-up path. Currently the update/synchronization is done _after_ the task has been placed by select_task_rq_fair(). The synchronization is done without having to take the rq lock using the existing mechanism used in remove_entity_load_avg(). Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: freedom.tan@mediatek.com Cc: keita.kobayashi.ym@renesas.com Cc: mgalbraith@suse.de Cc: sgurrappadi@nvidia.com Cc: vincent.guittot@linaro.org Cc: yuyang.du@intel.com Link: http://lkml.kernel.org/r/1476452472-24740-2-git-send-email-morten.rasmussen@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-10-14 13:41:07 +00:00
}
/*
* Task first catches up with cfs_rq, and then subtract
* itself from the cfs_rq (task must be off the queue now).
*/
static void remove_entity_load_avg(struct sched_entity *se)
{
struct cfs_rq *cfs_rq = cfs_rq_of(se);
unsigned long flags;
/*
sched/fair: Fix PELT integrity for new tasks Vincent and Yuyang found another few scenarios in which entity tracking goes wobbly. The scenarios are basically due to the fact that new tasks are not immediately attached and thereby differ from the normal situation -- a task is always attached to a cfs_rq load average (such that it includes its blocked contribution) and are explicitly detached/attached on migration to another cfs_rq. Scenario 1: switch to fair class p->sched_class = fair_class; if (queued) enqueue_task(p); ... enqueue_entity() enqueue_entity_load_avg() migrated = !sa->last_update_time (true) if (migrated) attach_entity_load_avg() check_class_changed() switched_from() (!fair) switched_to() (fair) switched_to_fair() attach_entity_load_avg() If @p is a new task that hasn't been fair before, it will have !last_update_time and, per the above, end up in attach_entity_load_avg() _twice_. Scenario 2: change between cgroups sched_move_group(p) if (queued) dequeue_task() task_move_group_fair() detach_task_cfs_rq() detach_entity_load_avg() set_task_rq() attach_task_cfs_rq() attach_entity_load_avg() if (queued) enqueue_task(); ... enqueue_entity() enqueue_entity_load_avg() migrated = !sa->last_update_time (true) if (migrated) attach_entity_load_avg() Similar as with scenario 1, if @p is a new task, it will have !load_update_time and we'll end up in attach_entity_load_avg() _twice_. Furthermore, notice how we do a detach_entity_load_avg() on something that wasn't attached to begin with. As stated above; the problem is that the new task isn't yet attached to the load tracking and thereby violates the invariant assumption. This patch remedies this by ensuring a new task is indeed properly attached to the load tracking on creation, through post_init_entity_util_avg(). Of course, this isn't entirely as straightforward as one might think, since the task is hashed before we call wake_up_new_task() and thus can be poked at. We avoid this by adding TASK_NEW and teaching cpu_cgroup_can_attach() to refuse such tasks. Reported-by: Yuyang Du <yuyang.du@intel.com> Reported-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-06-16 11:29:28 +00:00
* tasks cannot exit without having gone through wake_up_new_task() ->
* post_init_entity_util_avg() which will have added things to the
* cfs_rq, so we can remove unconditionally.
*/
sched/fair: Compute task/cpu utilization at wake-up correctly At task wake-up load-tracking isn't updated until the task is enqueued. The task's own view of its utilization contribution may therefore not be aligned with its contribution to the cfs_rq load-tracking which may have been updated in the meantime. Basically, the task's own utilization hasn't yet accounted for the sleep decay, while the cfs_rq may have (partially). Estimating the cfs_rq utilization in case the task is migrated at wake-up as task_rq(p)->cfs.avg.util_avg - p->se.avg.util_avg is therefore incorrect as the two load-tracking signals aren't time synchronized (different last update). To solve this problem, this patch synchronizes the task utilization with its previous rq before the task utilization is used in the wake-up path. Currently the update/synchronization is done _after_ the task has been placed by select_task_rq_fair(). The synchronization is done without having to take the rq lock using the existing mechanism used in remove_entity_load_avg(). Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: freedom.tan@mediatek.com Cc: keita.kobayashi.ym@renesas.com Cc: mgalbraith@suse.de Cc: sgurrappadi@nvidia.com Cc: vincent.guittot@linaro.org Cc: yuyang.du@intel.com Link: http://lkml.kernel.org/r/1476452472-24740-2-git-send-email-morten.rasmussen@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-10-14 13:41:07 +00:00
sync_entity_load_avg(se);
raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
++cfs_rq->removed.nr;
cfs_rq->removed.util_avg += se->avg.util_avg;
cfs_rq->removed.load_avg += se->avg.load_avg;
sched/pelt: Add a new runnable average signal Now that runnable_load_avg has been removed, we can replace it by a new signal that will highlight the runnable pressure on a cfs_rq. This signal track the waiting time of tasks on rq and can help to better define the state of rqs. At now, only util_avg is used to define the state of a rq: A rq with more that around 80% of utilization and more than 1 tasks is considered as overloaded. But the util_avg signal of a rq can become temporaly low after that a task migrated onto another rq which can bias the classification of the rq. When tasks compete for the same rq, their runnable average signal will be higher than util_avg as it will include the waiting time and we can use this signal to better classify cfs_rqs. The new runnable_avg will track the runnable time of a task which simply adds the waiting time to the running time. The runnable _avg of cfs_rq will be the /Sum of se's runnable_avg and the runnable_avg of group entity will follow the one of the rq similarly to util_avg. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: "Dietmar Eggemann <dietmar.eggemann@arm.com>" Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Valentin Schneider <valentin.schneider@arm.com> Cc: Phil Auld <pauld@redhat.com> Cc: Hillf Danton <hdanton@sina.com> Link: https://lore.kernel.org/r/20200224095223.13361-9-mgorman@techsingularity.net
2020-02-24 09:52:18 +00:00
cfs_rq->removed.runnable_avg += se->avg.runnable_avg;
raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
}
sched: Fix wrong rq's runnable_avg update with rt tasks The current update of the rq's load can be erroneous when RT tasks are involved. The update of the load of a rq that becomes idle, is done only if the avg_idle is less than sysctl_sched_migration_cost. If RT tasks and short idle duration alternate, the runnable_avg will not be updated correctly and the time will be accounted as idle time when a CFS task wakes up. A new idle_enter function is called when the next task is the idle function so the elapsed time will be accounted as run time in the load of the rq, whatever the average idle time is. The function update_rq_runnable_avg is removed from idle_balance. When a RT task is scheduled on an idle CPU, the update of the rq's load is not done when the rq exit idle state because CFS's functions are not called. Then, the idle_balance, which is called just before entering the idle function, updates the rq's load and makes the assumption that the elapsed time since the last update, was only running time. As a consequence, the rq's load of a CPU that only runs a periodic RT task, is close to LOAD_AVG_MAX whatever the running duration of the RT task is. A new idle_exit function is called when the prev task is the idle function so the elapsed time will be accounted as idle time in the rq's load. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Acked-by: Steven Rostedt <rostedt@goodmis.org> Cc: linaro-kernel@lists.linaro.org Cc: peterz@infradead.org Cc: pjt@google.com Cc: fweisbec@gmail.com Cc: efault@gmx.de Link: http://lkml.kernel.org/r/1366302867-5055-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-04-18 16:34:26 +00:00
sched/pelt: Add a new runnable average signal Now that runnable_load_avg has been removed, we can replace it by a new signal that will highlight the runnable pressure on a cfs_rq. This signal track the waiting time of tasks on rq and can help to better define the state of rqs. At now, only util_avg is used to define the state of a rq: A rq with more that around 80% of utilization and more than 1 tasks is considered as overloaded. But the util_avg signal of a rq can become temporaly low after that a task migrated onto another rq which can bias the classification of the rq. When tasks compete for the same rq, their runnable average signal will be higher than util_avg as it will include the waiting time and we can use this signal to better classify cfs_rqs. The new runnable_avg will track the runnable time of a task which simply adds the waiting time to the running time. The runnable _avg of cfs_rq will be the /Sum of se's runnable_avg and the runnable_avg of group entity will follow the one of the rq similarly to util_avg. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: "Dietmar Eggemann <dietmar.eggemann@arm.com>" Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Valentin Schneider <valentin.schneider@arm.com> Cc: Phil Auld <pauld@redhat.com> Cc: Hillf Danton <hdanton@sina.com> Link: https://lore.kernel.org/r/20200224095223.13361-9-mgorman@techsingularity.net
2020-02-24 09:52:18 +00:00
static inline unsigned long cfs_rq_runnable_avg(struct cfs_rq *cfs_rq)
{
return cfs_rq->avg.runnable_avg;
}
static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
{
return cfs_rq->avg.load_avg;
}
sched/fair: Add util_est on top of PELT The util_avg signal computed by PELT is too variable for some use-cases. For example, a big task waking up after a long sleep period will have its utilization almost completely decayed. This introduces some latency before schedutil will be able to pick the best frequency to run a task. The same issue can affect task placement. Indeed, since the task utilization is already decayed at wakeup, when the task is enqueued in a CPU, this can result in a CPU running a big task as being temporarily represented as being almost empty. This leads to a race condition where other tasks can be potentially allocated on a CPU which just started to run a big task which slept for a relatively long period. Moreover, the PELT utilization of a task can be updated every [ms], thus making it a continuously changing value for certain longer running tasks. This means that the instantaneous PELT utilization of a RUNNING task is not really meaningful to properly support scheduler decisions. For all these reasons, a more stable signal can do a better job of representing the expected/estimated utilization of a task/cfs_rq. Such a signal can be easily created on top of PELT by still using it as an estimator which produces values to be aggregated on meaningful events. This patch adds a simple implementation of util_est, a new signal built on top of PELT's util_avg where: util_est(task) = max(task::util_avg, f(task::util_avg@dequeue)) This allows to remember how big a task has been reported by PELT in its previous activations via f(task::util_avg@dequeue), which is the new _task_util_est(struct task_struct*) function added by this patch. If a task should change its behavior and it runs longer in a new activation, after a certain time its util_est will just track the original PELT signal (i.e. task::util_avg). The estimated utilization of cfs_rq is defined only for root ones. That's because the only sensible consumer of this signal are the scheduler and schedutil when looking for the overall CPU utilization due to FAIR tasks. For this reason, the estimated utilization of a root cfs_rq is simply defined as: util_est(cfs_rq) = max(cfs_rq::util_avg, cfs_rq::util_est::enqueued) where: cfs_rq::util_est::enqueued = sum(_task_util_est(task)) for each RUNNABLE task on that root cfs_rq It's worth noting that the estimated utilization is tracked only for objects of interests, specifically: - Tasks: to better support tasks placement decisions - root cfs_rqs: to better support both tasks placement decisions as well as frequencies selection Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Joel Fernandes <joelaf@google.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Rafael J . Wysocki <rafael.j.wysocki@intel.com> Cc: Steve Muckle <smuckle@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Todd Kjos <tkjos@android.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Link: http://lkml.kernel.org/r/20180309095245.11071-2-patrick.bellasi@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-03-09 09:52:42 +00:00
static inline unsigned long task_util(struct task_struct *p)
{
return READ_ONCE(p->se.avg.util_avg);
}
static inline unsigned long _task_util_est(struct task_struct *p)
{
struct util_est ue = READ_ONCE(p->se.avg.util_est);
2018-11-05 14:53:59 +00:00
return (max(ue.ewma, ue.enqueued) | UTIL_AVG_UNCHANGED);
sched/fair: Add util_est on top of PELT The util_avg signal computed by PELT is too variable for some use-cases. For example, a big task waking up after a long sleep period will have its utilization almost completely decayed. This introduces some latency before schedutil will be able to pick the best frequency to run a task. The same issue can affect task placement. Indeed, since the task utilization is already decayed at wakeup, when the task is enqueued in a CPU, this can result in a CPU running a big task as being temporarily represented as being almost empty. This leads to a race condition where other tasks can be potentially allocated on a CPU which just started to run a big task which slept for a relatively long period. Moreover, the PELT utilization of a task can be updated every [ms], thus making it a continuously changing value for certain longer running tasks. This means that the instantaneous PELT utilization of a RUNNING task is not really meaningful to properly support scheduler decisions. For all these reasons, a more stable signal can do a better job of representing the expected/estimated utilization of a task/cfs_rq. Such a signal can be easily created on top of PELT by still using it as an estimator which produces values to be aggregated on meaningful events. This patch adds a simple implementation of util_est, a new signal built on top of PELT's util_avg where: util_est(task) = max(task::util_avg, f(task::util_avg@dequeue)) This allows to remember how big a task has been reported by PELT in its previous activations via f(task::util_avg@dequeue), which is the new _task_util_est(struct task_struct*) function added by this patch. If a task should change its behavior and it runs longer in a new activation, after a certain time its util_est will just track the original PELT signal (i.e. task::util_avg). The estimated utilization of cfs_rq is defined only for root ones. That's because the only sensible consumer of this signal are the scheduler and schedutil when looking for the overall CPU utilization due to FAIR tasks. For this reason, the estimated utilization of a root cfs_rq is simply defined as: util_est(cfs_rq) = max(cfs_rq::util_avg, cfs_rq::util_est::enqueued) where: cfs_rq::util_est::enqueued = sum(_task_util_est(task)) for each RUNNABLE task on that root cfs_rq It's worth noting that the estimated utilization is tracked only for objects of interests, specifically: - Tasks: to better support tasks placement decisions - root cfs_rqs: to better support both tasks placement decisions as well as frequencies selection Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Joel Fernandes <joelaf@google.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Rafael J . Wysocki <rafael.j.wysocki@intel.com> Cc: Steve Muckle <smuckle@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Todd Kjos <tkjos@android.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Link: http://lkml.kernel.org/r/20180309095245.11071-2-patrick.bellasi@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-03-09 09:52:42 +00:00
}
static inline unsigned long task_util_est(struct task_struct *p)
{
return max(task_util(p), _task_util_est(p));
}
#ifdef CONFIG_UCLAMP_TASK
static inline unsigned long uclamp_task_util(struct task_struct *p)
{
return clamp(task_util_est(p),
uclamp_eff_value(p, UCLAMP_MIN),
uclamp_eff_value(p, UCLAMP_MAX));
}
#else
static inline unsigned long uclamp_task_util(struct task_struct *p)
{
return task_util_est(p);
}
#endif
sched/fair: Add util_est on top of PELT The util_avg signal computed by PELT is too variable for some use-cases. For example, a big task waking up after a long sleep period will have its utilization almost completely decayed. This introduces some latency before schedutil will be able to pick the best frequency to run a task. The same issue can affect task placement. Indeed, since the task utilization is already decayed at wakeup, when the task is enqueued in a CPU, this can result in a CPU running a big task as being temporarily represented as being almost empty. This leads to a race condition where other tasks can be potentially allocated on a CPU which just started to run a big task which slept for a relatively long period. Moreover, the PELT utilization of a task can be updated every [ms], thus making it a continuously changing value for certain longer running tasks. This means that the instantaneous PELT utilization of a RUNNING task is not really meaningful to properly support scheduler decisions. For all these reasons, a more stable signal can do a better job of representing the expected/estimated utilization of a task/cfs_rq. Such a signal can be easily created on top of PELT by still using it as an estimator which produces values to be aggregated on meaningful events. This patch adds a simple implementation of util_est, a new signal built on top of PELT's util_avg where: util_est(task) = max(task::util_avg, f(task::util_avg@dequeue)) This allows to remember how big a task has been reported by PELT in its previous activations via f(task::util_avg@dequeue), which is the new _task_util_est(struct task_struct*) function added by this patch. If a task should change its behavior and it runs longer in a new activation, after a certain time its util_est will just track the original PELT signal (i.e. task::util_avg). The estimated utilization of cfs_rq is defined only for root ones. That's because the only sensible consumer of this signal are the scheduler and schedutil when looking for the overall CPU utilization due to FAIR tasks. For this reason, the estimated utilization of a root cfs_rq is simply defined as: util_est(cfs_rq) = max(cfs_rq::util_avg, cfs_rq::util_est::enqueued) where: cfs_rq::util_est::enqueued = sum(_task_util_est(task)) for each RUNNABLE task on that root cfs_rq It's worth noting that the estimated utilization is tracked only for objects of interests, specifically: - Tasks: to better support tasks placement decisions - root cfs_rqs: to better support both tasks placement decisions as well as frequencies selection Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Joel Fernandes <joelaf@google.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Rafael J . Wysocki <rafael.j.wysocki@intel.com> Cc: Steve Muckle <smuckle@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Todd Kjos <tkjos@android.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Link: http://lkml.kernel.org/r/20180309095245.11071-2-patrick.bellasi@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-03-09 09:52:42 +00:00
static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
struct task_struct *p)
{
unsigned int enqueued;
if (!sched_feat(UTIL_EST))
return;
/* Update root cfs_rq's estimated utilization */
enqueued = cfs_rq->avg.util_est.enqueued;
2018-11-05 14:53:59 +00:00
enqueued += _task_util_est(p);
sched/fair: Add util_est on top of PELT The util_avg signal computed by PELT is too variable for some use-cases. For example, a big task waking up after a long sleep period will have its utilization almost completely decayed. This introduces some latency before schedutil will be able to pick the best frequency to run a task. The same issue can affect task placement. Indeed, since the task utilization is already decayed at wakeup, when the task is enqueued in a CPU, this can result in a CPU running a big task as being temporarily represented as being almost empty. This leads to a race condition where other tasks can be potentially allocated on a CPU which just started to run a big task which slept for a relatively long period. Moreover, the PELT utilization of a task can be updated every [ms], thus making it a continuously changing value for certain longer running tasks. This means that the instantaneous PELT utilization of a RUNNING task is not really meaningful to properly support scheduler decisions. For all these reasons, a more stable signal can do a better job of representing the expected/estimated utilization of a task/cfs_rq. Such a signal can be easily created on top of PELT by still using it as an estimator which produces values to be aggregated on meaningful events. This patch adds a simple implementation of util_est, a new signal built on top of PELT's util_avg where: util_est(task) = max(task::util_avg, f(task::util_avg@dequeue)) This allows to remember how big a task has been reported by PELT in its previous activations via f(task::util_avg@dequeue), which is the new _task_util_est(struct task_struct*) function added by this patch. If a task should change its behavior and it runs longer in a new activation, after a certain time its util_est will just track the original PELT signal (i.e. task::util_avg). The estimated utilization of cfs_rq is defined only for root ones. That's because the only sensible consumer of this signal are the scheduler and schedutil when looking for the overall CPU utilization due to FAIR tasks. For this reason, the estimated utilization of a root cfs_rq is simply defined as: util_est(cfs_rq) = max(cfs_rq::util_avg, cfs_rq::util_est::enqueued) where: cfs_rq::util_est::enqueued = sum(_task_util_est(task)) for each RUNNABLE task on that root cfs_rq It's worth noting that the estimated utilization is tracked only for objects of interests, specifically: - Tasks: to better support tasks placement decisions - root cfs_rqs: to better support both tasks placement decisions as well as frequencies selection Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Joel Fernandes <joelaf@google.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Rafael J . Wysocki <rafael.j.wysocki@intel.com> Cc: Steve Muckle <smuckle@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Todd Kjos <tkjos@android.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Link: http://lkml.kernel.org/r/20180309095245.11071-2-patrick.bellasi@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-03-09 09:52:42 +00:00
WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
}
/*
* Check if a (signed) value is within a specified (unsigned) margin,
* based on the observation that:
*
* abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
*
* NOTE: this only works when value + maring < INT_MAX.
*/
static inline bool within_margin(int value, int margin)
{
return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
}
static void
util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p, bool task_sleep)
{
long last_ewma_diff;
struct util_est ue;
sched/pelt: Skip updating util_est when utilization is higher than CPU's capacity util_est is mainly meant to be a lower-bound for tasks utilization. That's why task_util_est() returns the actual util_avg when it's higher than the estimated utilization. With new invaraince signal and without any special check on samples collection, if a task is limited because of thermal capping for example, we could end up overestimating its utilization and thus perhaps generating an unwanted frequency spike when the capping is relaxed... and (even worst) it will take some more activations for the estimated utilization to converge back to the actual utilization. Since we cannot easily know if there is idle time in a CPU when a task completes an activation with a utilization higher then the CPU capacity, we skip the sampling when utilization is higher than CPU's capacity. Suggested-by: Patrick Bellasi <patrick.bellasi@arm.com> Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: pjt@google.com Cc: pkondeti@codeaurora.org Cc: quentin.perret@arm.com Cc: rjw@rjwysocki.net Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Link: https://lkml.kernel.org/r/1548257214-13745-4-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-01-23 15:26:54 +00:00
int cpu;
sched/fair: Add util_est on top of PELT The util_avg signal computed by PELT is too variable for some use-cases. For example, a big task waking up after a long sleep period will have its utilization almost completely decayed. This introduces some latency before schedutil will be able to pick the best frequency to run a task. The same issue can affect task placement. Indeed, since the task utilization is already decayed at wakeup, when the task is enqueued in a CPU, this can result in a CPU running a big task as being temporarily represented as being almost empty. This leads to a race condition where other tasks can be potentially allocated on a CPU which just started to run a big task which slept for a relatively long period. Moreover, the PELT utilization of a task can be updated every [ms], thus making it a continuously changing value for certain longer running tasks. This means that the instantaneous PELT utilization of a RUNNING task is not really meaningful to properly support scheduler decisions. For all these reasons, a more stable signal can do a better job of representing the expected/estimated utilization of a task/cfs_rq. Such a signal can be easily created on top of PELT by still using it as an estimator which produces values to be aggregated on meaningful events. This patch adds a simple implementation of util_est, a new signal built on top of PELT's util_avg where: util_est(task) = max(task::util_avg, f(task::util_avg@dequeue)) This allows to remember how big a task has been reported by PELT in its previous activations via f(task::util_avg@dequeue), which is the new _task_util_est(struct task_struct*) function added by this patch. If a task should change its behavior and it runs longer in a new activation, after a certain time its util_est will just track the original PELT signal (i.e. task::util_avg). The estimated utilization of cfs_rq is defined only for root ones. That's because the only sensible consumer of this signal are the scheduler and schedutil when looking for the overall CPU utilization due to FAIR tasks. For this reason, the estimated utilization of a root cfs_rq is simply defined as: util_est(cfs_rq) = max(cfs_rq::util_avg, cfs_rq::util_est::enqueued) where: cfs_rq::util_est::enqueued = sum(_task_util_est(task)) for each RUNNABLE task on that root cfs_rq It's worth noting that the estimated utilization is tracked only for objects of interests, specifically: - Tasks: to better support tasks placement decisions - root cfs_rqs: to better support both tasks placement decisions as well as frequencies selection Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Joel Fernandes <joelaf@google.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Rafael J . Wysocki <rafael.j.wysocki@intel.com> Cc: Steve Muckle <smuckle@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Todd Kjos <tkjos@android.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Link: http://lkml.kernel.org/r/20180309095245.11071-2-patrick.bellasi@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-03-09 09:52:42 +00:00
if (!sched_feat(UTIL_EST))
return;
sched/util_est: Fix util_est_dequeue() for throttled cfs_rq When a cfs_rq is throttled, parent cfs_rq->nr_running is decreased and everything happens at cfs_rq level. Currently util_est stays unchanged in such case and it keeps accounting the utilization of throttled tasks. This can somewhat make sense as we don't dequeue tasks but only throttled cfs_rq. If a task of another group is enqueued/dequeued and root cfs_rq becomes idle during the dequeue, util_est will be cleared whereas it was accounting util_est of throttled tasks before. So the behavior of util_est is not always the same regarding throttled tasks and depends of side activity. Furthermore, util_est will not be updated when the cfs_rq is unthrottled as everything happens at cfs_rq level. Main results is that util_est will stay null whereas we now have running tasks. We have to wait for the next dequeue/enqueue of the previously throttled tasks to get an up to date util_est. Remove the assumption that cfs_rq's estimated utilization of a CPU is 0 if there is no running task so the util_est of a task remains until the latter is dequeued even if its cfs_rq has been throttled. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Patrick Bellasi <patrick.bellasi@arm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Fixes: 7f65ea42eb00 ("sched/fair: Add util_est on top of PELT") Link: http://lkml.kernel.org/r/1528972380-16268-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-14 10:33:00 +00:00
/* Update root cfs_rq's estimated utilization */
ue.enqueued = cfs_rq->avg.util_est.enqueued;
2018-11-05 14:53:59 +00:00
ue.enqueued -= min_t(unsigned int, ue.enqueued, _task_util_est(p));
sched/fair: Add util_est on top of PELT The util_avg signal computed by PELT is too variable for some use-cases. For example, a big task waking up after a long sleep period will have its utilization almost completely decayed. This introduces some latency before schedutil will be able to pick the best frequency to run a task. The same issue can affect task placement. Indeed, since the task utilization is already decayed at wakeup, when the task is enqueued in a CPU, this can result in a CPU running a big task as being temporarily represented as being almost empty. This leads to a race condition where other tasks can be potentially allocated on a CPU which just started to run a big task which slept for a relatively long period. Moreover, the PELT utilization of a task can be updated every [ms], thus making it a continuously changing value for certain longer running tasks. This means that the instantaneous PELT utilization of a RUNNING task is not really meaningful to properly support scheduler decisions. For all these reasons, a more stable signal can do a better job of representing the expected/estimated utilization of a task/cfs_rq. Such a signal can be easily created on top of PELT by still using it as an estimator which produces values to be aggregated on meaningful events. This patch adds a simple implementation of util_est, a new signal built on top of PELT's util_avg where: util_est(task) = max(task::util_avg, f(task::util_avg@dequeue)) This allows to remember how big a task has been reported by PELT in its previous activations via f(task::util_avg@dequeue), which is the new _task_util_est(struct task_struct*) function added by this patch. If a task should change its behavior and it runs longer in a new activation, after a certain time its util_est will just track the original PELT signal (i.e. task::util_avg). The estimated utilization of cfs_rq is defined only for root ones. That's because the only sensible consumer of this signal are the scheduler and schedutil when looking for the overall CPU utilization due to FAIR tasks. For this reason, the estimated utilization of a root cfs_rq is simply defined as: util_est(cfs_rq) = max(cfs_rq::util_avg, cfs_rq::util_est::enqueued) where: cfs_rq::util_est::enqueued = sum(_task_util_est(task)) for each RUNNABLE task on that root cfs_rq It's worth noting that the estimated utilization is tracked only for objects of interests, specifically: - Tasks: to better support tasks placement decisions - root cfs_rqs: to better support both tasks placement decisions as well as frequencies selection Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Joel Fernandes <joelaf@google.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Rafael J . Wysocki <rafael.j.wysocki@intel.com> Cc: Steve Muckle <smuckle@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Todd Kjos <tkjos@android.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Link: http://lkml.kernel.org/r/20180309095245.11071-2-patrick.bellasi@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-03-09 09:52:42 +00:00
WRITE_ONCE(cfs_rq->avg.util_est.enqueued, ue.enqueued);
/*
* Skip update of task's estimated utilization when the task has not
* yet completed an activation, e.g. being migrated.
*/
if (!task_sleep)
return;
sched/fair: Update util_est only on util_avg updates The estimated utilization of a task is currently updated every time the task is dequeued. However, to keep overheads under control, PELT signals are effectively updated at maximum once every 1ms. Thus, for really short running tasks, it can happen that their util_avg value has not been updates since their last enqueue. If such tasks are also frequently running tasks (e.g. the kind of workload generated by hackbench) it can also happen that their util_avg is updated only every few activations. This means that updating util_est at every dequeue potentially introduces not necessary overheads and it's also conceptually wrong if the util_avg signal has never been updated during a task activation. Let's introduce a throttling mechanism on task's util_est updates to sync them with util_avg updates. To make the solution memory efficient, both in terms of space and load/store operations, we encode a synchronization flag into the LSB of util_est.enqueued. This makes util_est an even values only metric, which is still considered good enough for its purpose. The synchronization bit is (re)set by __update_load_avg_se() once the PELT signal of a task has been updated during its last activation. Such a throttling mechanism allows to keep under control util_est overheads in the wakeup hot path, thus making it a suitable mechanism which can be enabled also on high-intensity workload systems. Thus, this now switches on by default the estimation utilization scheduler feature. Suggested-by: Chris Redpath <chris.redpath@arm.com> Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Joel Fernandes <joelaf@google.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rafael J . Wysocki <rafael.j.wysocki@intel.com> Cc: Steve Muckle <smuckle@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Todd Kjos <tkjos@android.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Link: http://lkml.kernel.org/r/20180309095245.11071-5-patrick.bellasi@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-03-09 09:52:45 +00:00
/*
* If the PELT values haven't changed since enqueue time,
* skip the util_est update.
*/
ue = p->se.avg.util_est;
if (ue.enqueued & UTIL_AVG_UNCHANGED)
return;
sched/fair/util_est: Implement faster ramp-up EWMA on utilization increases The estimated utilization for a task: util_est = max(util_avg, est.enqueue, est.ewma) is defined based on: - util_avg: the PELT defined utilization - est.enqueued: the util_avg at the end of the last activation - est.ewma: a exponential moving average on the est.enqueued samples According to this definition, when a task suddenly changes its bandwidth requirements from small to big, the EWMA will need to collect multiple samples before converging up to track the new big utilization. This slow convergence towards bigger utilization values is not aligned to the default scheduler behavior, which is to optimize for performance. Moreover, the est.ewma component fails to compensate for temporarely utilization drops which spans just few est.enqueued samples. To let util_est do a better job in the scenario depicted above, change its definition by making util_est directly follow upward motion and only decay the est.ewma on downward. Signed-off-by: Patrick Bellasi <patrick.bellasi@matbug.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Douglas Raillard <douglas.raillard@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Quentin Perret <qperret@google.com> Cc: Rafael J . Wysocki <rafael.j.wysocki@intel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: https://lkml.kernel.org/r/20191023205630.14469-1-patrick.bellasi@matbug.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-23 20:56:30 +00:00
/*
* Reset EWMA on utilization increases, the moving average is used only
* to smooth utilization decreases.
*/
ue.enqueued = (task_util(p) | UTIL_AVG_UNCHANGED);
if (sched_feat(UTIL_EST_FASTUP)) {
if (ue.ewma < ue.enqueued) {
ue.ewma = ue.enqueued;
goto done;
}
}
sched/fair: Add util_est on top of PELT The util_avg signal computed by PELT is too variable for some use-cases. For example, a big task waking up after a long sleep period will have its utilization almost completely decayed. This introduces some latency before schedutil will be able to pick the best frequency to run a task. The same issue can affect task placement. Indeed, since the task utilization is already decayed at wakeup, when the task is enqueued in a CPU, this can result in a CPU running a big task as being temporarily represented as being almost empty. This leads to a race condition where other tasks can be potentially allocated on a CPU which just started to run a big task which slept for a relatively long period. Moreover, the PELT utilization of a task can be updated every [ms], thus making it a continuously changing value for certain longer running tasks. This means that the instantaneous PELT utilization of a RUNNING task is not really meaningful to properly support scheduler decisions. For all these reasons, a more stable signal can do a better job of representing the expected/estimated utilization of a task/cfs_rq. Such a signal can be easily created on top of PELT by still using it as an estimator which produces values to be aggregated on meaningful events. This patch adds a simple implementation of util_est, a new signal built on top of PELT's util_avg where: util_est(task) = max(task::util_avg, f(task::util_avg@dequeue)) This allows to remember how big a task has been reported by PELT in its previous activations via f(task::util_avg@dequeue), which is the new _task_util_est(struct task_struct*) function added by this patch. If a task should change its behavior and it runs longer in a new activation, after a certain time its util_est will just track the original PELT signal (i.e. task::util_avg). The estimated utilization of cfs_rq is defined only for root ones. That's because the only sensible consumer of this signal are the scheduler and schedutil when looking for the overall CPU utilization due to FAIR tasks. For this reason, the estimated utilization of a root cfs_rq is simply defined as: util_est(cfs_rq) = max(cfs_rq::util_avg, cfs_rq::util_est::enqueued) where: cfs_rq::util_est::enqueued = sum(_task_util_est(task)) for each RUNNABLE task on that root cfs_rq It's worth noting that the estimated utilization is tracked only for objects of interests, specifically: - Tasks: to better support tasks placement decisions - root cfs_rqs: to better support both tasks placement decisions as well as frequencies selection Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Joel Fernandes <joelaf@google.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Rafael J . Wysocki <rafael.j.wysocki@intel.com> Cc: Steve Muckle <smuckle@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Todd Kjos <tkjos@android.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Link: http://lkml.kernel.org/r/20180309095245.11071-2-patrick.bellasi@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-03-09 09:52:42 +00:00
/*
* Skip update of task's estimated utilization when its EWMA is
* already ~1% close to its last activation value.
*/
last_ewma_diff = ue.enqueued - ue.ewma;
if (within_margin(last_ewma_diff, (SCHED_CAPACITY_SCALE / 100)))
return;
sched/pelt: Skip updating util_est when utilization is higher than CPU's capacity util_est is mainly meant to be a lower-bound for tasks utilization. That's why task_util_est() returns the actual util_avg when it's higher than the estimated utilization. With new invaraince signal and without any special check on samples collection, if a task is limited because of thermal capping for example, we could end up overestimating its utilization and thus perhaps generating an unwanted frequency spike when the capping is relaxed... and (even worst) it will take some more activations for the estimated utilization to converge back to the actual utilization. Since we cannot easily know if there is idle time in a CPU when a task completes an activation with a utilization higher then the CPU capacity, we skip the sampling when utilization is higher than CPU's capacity. Suggested-by: Patrick Bellasi <patrick.bellasi@arm.com> Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: pjt@google.com Cc: pkondeti@codeaurora.org Cc: quentin.perret@arm.com Cc: rjw@rjwysocki.net Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Link: https://lkml.kernel.org/r/1548257214-13745-4-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-01-23 15:26:54 +00:00
/*
* To avoid overestimation of actual task utilization, skip updates if
* we cannot grant there is idle time in this CPU.
*/
cpu = cpu_of(rq_of(cfs_rq));
if (task_util(p) > capacity_orig_of(cpu))
return;
sched/fair: Add util_est on top of PELT The util_avg signal computed by PELT is too variable for some use-cases. For example, a big task waking up after a long sleep period will have its utilization almost completely decayed. This introduces some latency before schedutil will be able to pick the best frequency to run a task. The same issue can affect task placement. Indeed, since the task utilization is already decayed at wakeup, when the task is enqueued in a CPU, this can result in a CPU running a big task as being temporarily represented as being almost empty. This leads to a race condition where other tasks can be potentially allocated on a CPU which just started to run a big task which slept for a relatively long period. Moreover, the PELT utilization of a task can be updated every [ms], thus making it a continuously changing value for certain longer running tasks. This means that the instantaneous PELT utilization of a RUNNING task is not really meaningful to properly support scheduler decisions. For all these reasons, a more stable signal can do a better job of representing the expected/estimated utilization of a task/cfs_rq. Such a signal can be easily created on top of PELT by still using it as an estimator which produces values to be aggregated on meaningful events. This patch adds a simple implementation of util_est, a new signal built on top of PELT's util_avg where: util_est(task) = max(task::util_avg, f(task::util_avg@dequeue)) This allows to remember how big a task has been reported by PELT in its previous activations via f(task::util_avg@dequeue), which is the new _task_util_est(struct task_struct*) function added by this patch. If a task should change its behavior and it runs longer in a new activation, after a certain time its util_est will just track the original PELT signal (i.e. task::util_avg). The estimated utilization of cfs_rq is defined only for root ones. That's because the only sensible consumer of this signal are the scheduler and schedutil when looking for the overall CPU utilization due to FAIR tasks. For this reason, the estimated utilization of a root cfs_rq is simply defined as: util_est(cfs_rq) = max(cfs_rq::util_avg, cfs_rq::util_est::enqueued) where: cfs_rq::util_est::enqueued = sum(_task_util_est(task)) for each RUNNABLE task on that root cfs_rq It's worth noting that the estimated utilization is tracked only for objects of interests, specifically: - Tasks: to better support tasks placement decisions - root cfs_rqs: to better support both tasks placement decisions as well as frequencies selection Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Joel Fernandes <joelaf@google.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Rafael J . Wysocki <rafael.j.wysocki@intel.com> Cc: Steve Muckle <smuckle@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Todd Kjos <tkjos@android.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Link: http://lkml.kernel.org/r/20180309095245.11071-2-patrick.bellasi@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-03-09 09:52:42 +00:00
/*
* Update Task's estimated utilization
*
* When *p completes an activation we can consolidate another sample
* of the task size. This is done by storing the current PELT value
* as ue.enqueued and by using this value to update the Exponential
* Weighted Moving Average (EWMA):
*
* ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
* = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
* = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
* = w * ( last_ewma_diff ) + ewma(t-1)
* = w * (last_ewma_diff + ewma(t-1) / w)
*
* Where 'w' is the weight of new samples, which is configured to be
* 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
*/
ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
ue.ewma += last_ewma_diff;
ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
sched/fair/util_est: Implement faster ramp-up EWMA on utilization increases The estimated utilization for a task: util_est = max(util_avg, est.enqueue, est.ewma) is defined based on: - util_avg: the PELT defined utilization - est.enqueued: the util_avg at the end of the last activation - est.ewma: a exponential moving average on the est.enqueued samples According to this definition, when a task suddenly changes its bandwidth requirements from small to big, the EWMA will need to collect multiple samples before converging up to track the new big utilization. This slow convergence towards bigger utilization values is not aligned to the default scheduler behavior, which is to optimize for performance. Moreover, the est.ewma component fails to compensate for temporarely utilization drops which spans just few est.enqueued samples. To let util_est do a better job in the scenario depicted above, change its definition by making util_est directly follow upward motion and only decay the est.ewma on downward. Signed-off-by: Patrick Bellasi <patrick.bellasi@matbug.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Douglas Raillard <douglas.raillard@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Quentin Perret <qperret@google.com> Cc: Rafael J . Wysocki <rafael.j.wysocki@intel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: https://lkml.kernel.org/r/20191023205630.14469-1-patrick.bellasi@matbug.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-23 20:56:30 +00:00
done:
sched/fair: Add util_est on top of PELT The util_avg signal computed by PELT is too variable for some use-cases. For example, a big task waking up after a long sleep period will have its utilization almost completely decayed. This introduces some latency before schedutil will be able to pick the best frequency to run a task. The same issue can affect task placement. Indeed, since the task utilization is already decayed at wakeup, when the task is enqueued in a CPU, this can result in a CPU running a big task as being temporarily represented as being almost empty. This leads to a race condition where other tasks can be potentially allocated on a CPU which just started to run a big task which slept for a relatively long period. Moreover, the PELT utilization of a task can be updated every [ms], thus making it a continuously changing value for certain longer running tasks. This means that the instantaneous PELT utilization of a RUNNING task is not really meaningful to properly support scheduler decisions. For all these reasons, a more stable signal can do a better job of representing the expected/estimated utilization of a task/cfs_rq. Such a signal can be easily created on top of PELT by still using it as an estimator which produces values to be aggregated on meaningful events. This patch adds a simple implementation of util_est, a new signal built on top of PELT's util_avg where: util_est(task) = max(task::util_avg, f(task::util_avg@dequeue)) This allows to remember how big a task has been reported by PELT in its previous activations via f(task::util_avg@dequeue), which is the new _task_util_est(struct task_struct*) function added by this patch. If a task should change its behavior and it runs longer in a new activation, after a certain time its util_est will just track the original PELT signal (i.e. task::util_avg). The estimated utilization of cfs_rq is defined only for root ones. That's because the only sensible consumer of this signal are the scheduler and schedutil when looking for the overall CPU utilization due to FAIR tasks. For this reason, the estimated utilization of a root cfs_rq is simply defined as: util_est(cfs_rq) = max(cfs_rq::util_avg, cfs_rq::util_est::enqueued) where: cfs_rq::util_est::enqueued = sum(_task_util_est(task)) for each RUNNABLE task on that root cfs_rq It's worth noting that the estimated utilization is tracked only for objects of interests, specifically: - Tasks: to better support tasks placement decisions - root cfs_rqs: to better support both tasks placement decisions as well as frequencies selection Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Joel Fernandes <joelaf@google.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Rafael J . Wysocki <rafael.j.wysocki@intel.com> Cc: Steve Muckle <smuckle@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Todd Kjos <tkjos@android.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Link: http://lkml.kernel.org/r/20180309095245.11071-2-patrick.bellasi@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-03-09 09:52:42 +00:00
WRITE_ONCE(p->se.avg.util_est, ue);
}
sched/fair: Add 'group_misfit_task' load-balance type To maximize throughput in systems with asymmetric CPU capacities (e.g. ARM big.LITTLE) load-balancing has to consider task and CPU utilization as well as per-CPU compute capacity when load-balancing in addition to the current average load based load-balancing policy. Tasks with high utilization that are scheduled on a lower capacity CPU need to be identified and migrated to a higher capacity CPU if possible to maximize throughput. To implement this additional policy an additional group_type (load-balance scenario) is added: 'group_misfit_task'. This represents scenarios where a sched_group has one or more tasks that are not suitable for its per-CPU capacity. 'group_misfit_task' is only considered if the system is not overloaded or imbalanced ('group_imbalanced' or 'group_overloaded'). Identifying misfit tasks requires the rq lock to be held. To avoid taking remote rq locks to examine source sched_groups for misfit tasks, each CPU is responsible for tracking misfit tasks themselves and update the rq->misfit_task flag. This means checking task utilization when tasks are scheduled and on sched_tick. Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: gaku.inami.xh@renesas.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1530699470-29808-3-git-send-email-morten.rasmussen@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-07-04 10:17:40 +00:00
static inline int task_fits_capacity(struct task_struct *p, long capacity)
{
return fits_capacity(uclamp_task_util(p), capacity);
sched/fair: Add 'group_misfit_task' load-balance type To maximize throughput in systems with asymmetric CPU capacities (e.g. ARM big.LITTLE) load-balancing has to consider task and CPU utilization as well as per-CPU compute capacity when load-balancing in addition to the current average load based load-balancing policy. Tasks with high utilization that are scheduled on a lower capacity CPU need to be identified and migrated to a higher capacity CPU if possible to maximize throughput. To implement this additional policy an additional group_type (load-balance scenario) is added: 'group_misfit_task'. This represents scenarios where a sched_group has one or more tasks that are not suitable for its per-CPU capacity. 'group_misfit_task' is only considered if the system is not overloaded or imbalanced ('group_imbalanced' or 'group_overloaded'). Identifying misfit tasks requires the rq lock to be held. To avoid taking remote rq locks to examine source sched_groups for misfit tasks, each CPU is responsible for tracking misfit tasks themselves and update the rq->misfit_task flag. This means checking task utilization when tasks are scheduled and on sched_tick. Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: gaku.inami.xh@renesas.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1530699470-29808-3-git-send-email-morten.rasmussen@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-07-04 10:17:40 +00:00
}
static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
{
if (!static_branch_unlikely(&sched_asym_cpucapacity))
return;
if (!p) {
rq->misfit_task_load = 0;
return;
}
if (task_fits_capacity(p, capacity_of(cpu_of(rq)))) {
rq->misfit_task_load = 0;
return;
}
rq->misfit_task_load = task_h_load(p);
}
#else /* CONFIG_SMP */
#define UPDATE_TG 0x0
#define SKIP_AGE_LOAD 0x0
#define DO_ATTACH 0x0
static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
{
cfs_rq_util_change(cfs_rq, 0);
}
sched/fair: Rewrite runnable load and utilization average tracking The idea of runnable load average (let runnable time contribute to weight) was proposed by Paul Turner and Ben Segall, and it is still followed by this rewrite. This rewrite aims to solve the following issues: 1. cfs_rq's load average (namely runnable_load_avg and blocked_load_avg) is updated at the granularity of an entity at a time, which results in the cfs_rq's load average is stale or partially updated: at any time, only one entity is up to date, all other entities are effectively lagging behind. This is undesirable. To illustrate, if we have n runnable entities in the cfs_rq, as time elapses, they certainly become outdated: t0: cfs_rq { e1_old, e2_old, ..., en_old } and when we update: t1: update e1, then we have cfs_rq { e1_new, e2_old, ..., en_old } t2: update e2, then we have cfs_rq { e1_old, e2_new, ..., en_old } ... We solve this by combining all runnable entities' load averages together in cfs_rq's avg, and update the cfs_rq's avg as a whole. This is based on the fact that if we regard the update as a function, then: w * update(e) = update(w * e) and update(e1) + update(e2) = update(e1 + e2), then w1 * update(e1) + w2 * update(e2) = update(w1 * e1 + w2 * e2) therefore, by this rewrite, we have an entirely updated cfs_rq at the time we update it: t1: update cfs_rq { e1_new, e2_new, ..., en_new } t2: update cfs_rq { e1_new, e2_new, ..., en_new } ... 2. cfs_rq's load average is different between top rq->cfs_rq and other task_group's per CPU cfs_rqs in whether or not blocked_load_average contributes to the load. The basic idea behind runnable load average (the same for utilization) is that the blocked state is taken into account as opposed to only accounting for the currently runnable state. Therefore, the average should include both the runnable/running and blocked load averages. This rewrite does that. In addition, we also combine runnable/running and blocked averages of all entities into the cfs_rq's average, and update it together at once. This is based on the fact that: update(runnable) + update(blocked) = update(runnable + blocked) This significantly reduces the code as we don't need to separately maintain/update runnable/running load and blocked load. 3. How task_group entities' share is calculated is complex and imprecise. We reduce the complexity in this rewrite to allow a very simple rule: the task_group's load_avg is aggregated from its per CPU cfs_rqs's load_avgs. Then group entity's weight is simply proportional to its own cfs_rq's load_avg / task_group's load_avg. To illustrate, if a task_group has { cfs_rq1, cfs_rq2, ..., cfs_rqn }, then, task_group_avg = cfs_rq1_avg + cfs_rq2_avg + ... + cfs_rqn_avg, then cfs_rqx's entity's share = cfs_rqx_avg / task_group_avg * task_group's share To sum up, this rewrite in principle is equivalent to the current one, but fixes the issues described above. Turns out, it significantly reduces the code complexity and hence increases clarity and efficiency. In addition, the new averages are more smooth/continuous (no spurious spikes and valleys) and updated more consistently and quickly to reflect the load dynamics. As a result, we have less load tracking overhead, better performance, and especially better power efficiency due to more balanced load. Signed-off-by: Yuyang Du <yuyang.du@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: arjan@linux.intel.com Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: fengguang.wu@intel.com Cc: len.brown@intel.com Cc: morten.rasmussen@arm.com Cc: pjt@google.com Cc: rafael.j.wysocki@intel.com Cc: umgwanakikbuti@gmail.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1436918682-4971-3-git-send-email-yuyang.du@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-07-15 00:04:37 +00:00
static inline void remove_entity_load_avg(struct sched_entity *se) {}
static inline void
attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
static inline void
detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
static inline int idle_balance(struct rq *rq, struct rq_flags *rf)
{
return 0;
}
sched/fair: Add util_est on top of PELT The util_avg signal computed by PELT is too variable for some use-cases. For example, a big task waking up after a long sleep period will have its utilization almost completely decayed. This introduces some latency before schedutil will be able to pick the best frequency to run a task. The same issue can affect task placement. Indeed, since the task utilization is already decayed at wakeup, when the task is enqueued in a CPU, this can result in a CPU running a big task as being temporarily represented as being almost empty. This leads to a race condition where other tasks can be potentially allocated on a CPU which just started to run a big task which slept for a relatively long period. Moreover, the PELT utilization of a task can be updated every [ms], thus making it a continuously changing value for certain longer running tasks. This means that the instantaneous PELT utilization of a RUNNING task is not really meaningful to properly support scheduler decisions. For all these reasons, a more stable signal can do a better job of representing the expected/estimated utilization of a task/cfs_rq. Such a signal can be easily created on top of PELT by still using it as an estimator which produces values to be aggregated on meaningful events. This patch adds a simple implementation of util_est, a new signal built on top of PELT's util_avg where: util_est(task) = max(task::util_avg, f(task::util_avg@dequeue)) This allows to remember how big a task has been reported by PELT in its previous activations via f(task::util_avg@dequeue), which is the new _task_util_est(struct task_struct*) function added by this patch. If a task should change its behavior and it runs longer in a new activation, after a certain time its util_est will just track the original PELT signal (i.e. task::util_avg). The estimated utilization of cfs_rq is defined only for root ones. That's because the only sensible consumer of this signal are the scheduler and schedutil when looking for the overall CPU utilization due to FAIR tasks. For this reason, the estimated utilization of a root cfs_rq is simply defined as: util_est(cfs_rq) = max(cfs_rq::util_avg, cfs_rq::util_est::enqueued) where: cfs_rq::util_est::enqueued = sum(_task_util_est(task)) for each RUNNABLE task on that root cfs_rq It's worth noting that the estimated utilization is tracked only for objects of interests, specifically: - Tasks: to better support tasks placement decisions - root cfs_rqs: to better support both tasks placement decisions as well as frequencies selection Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Joel Fernandes <joelaf@google.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Rafael J . Wysocki <rafael.j.wysocki@intel.com> Cc: Steve Muckle <smuckle@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Todd Kjos <tkjos@android.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Link: http://lkml.kernel.org/r/20180309095245.11071-2-patrick.bellasi@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-03-09 09:52:42 +00:00
static inline void
util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
static inline void
util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p,
bool task_sleep) {}
sched/fair: Add 'group_misfit_task' load-balance type To maximize throughput in systems with asymmetric CPU capacities (e.g. ARM big.LITTLE) load-balancing has to consider task and CPU utilization as well as per-CPU compute capacity when load-balancing in addition to the current average load based load-balancing policy. Tasks with high utilization that are scheduled on a lower capacity CPU need to be identified and migrated to a higher capacity CPU if possible to maximize throughput. To implement this additional policy an additional group_type (load-balance scenario) is added: 'group_misfit_task'. This represents scenarios where a sched_group has one or more tasks that are not suitable for its per-CPU capacity. 'group_misfit_task' is only considered if the system is not overloaded or imbalanced ('group_imbalanced' or 'group_overloaded'). Identifying misfit tasks requires the rq lock to be held. To avoid taking remote rq locks to examine source sched_groups for misfit tasks, each CPU is responsible for tracking misfit tasks themselves and update the rq->misfit_task flag. This means checking task utilization when tasks are scheduled and on sched_tick. Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: gaku.inami.xh@renesas.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1530699470-29808-3-git-send-email-morten.rasmussen@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-07-04 10:17:40 +00:00
static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {}
sched/fair: Add util_est on top of PELT The util_avg signal computed by PELT is too variable for some use-cases. For example, a big task waking up after a long sleep period will have its utilization almost completely decayed. This introduces some latency before schedutil will be able to pick the best frequency to run a task. The same issue can affect task placement. Indeed, since the task utilization is already decayed at wakeup, when the task is enqueued in a CPU, this can result in a CPU running a big task as being temporarily represented as being almost empty. This leads to a race condition where other tasks can be potentially allocated on a CPU which just started to run a big task which slept for a relatively long period. Moreover, the PELT utilization of a task can be updated every [ms], thus making it a continuously changing value for certain longer running tasks. This means that the instantaneous PELT utilization of a RUNNING task is not really meaningful to properly support scheduler decisions. For all these reasons, a more stable signal can do a better job of representing the expected/estimated utilization of a task/cfs_rq. Such a signal can be easily created on top of PELT by still using it as an estimator which produces values to be aggregated on meaningful events. This patch adds a simple implementation of util_est, a new signal built on top of PELT's util_avg where: util_est(task) = max(task::util_avg, f(task::util_avg@dequeue)) This allows to remember how big a task has been reported by PELT in its previous activations via f(task::util_avg@dequeue), which is the new _task_util_est(struct task_struct*) function added by this patch. If a task should change its behavior and it runs longer in a new activation, after a certain time its util_est will just track the original PELT signal (i.e. task::util_avg). The estimated utilization of cfs_rq is defined only for root ones. That's because the only sensible consumer of this signal are the scheduler and schedutil when looking for the overall CPU utilization due to FAIR tasks. For this reason, the estimated utilization of a root cfs_rq is simply defined as: util_est(cfs_rq) = max(cfs_rq::util_avg, cfs_rq::util_est::enqueued) where: cfs_rq::util_est::enqueued = sum(_task_util_est(task)) for each RUNNABLE task on that root cfs_rq It's worth noting that the estimated utilization is tracked only for objects of interests, specifically: - Tasks: to better support tasks placement decisions - root cfs_rqs: to better support both tasks placement decisions as well as frequencies selection Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Joel Fernandes <joelaf@google.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Rafael J . Wysocki <rafael.j.wysocki@intel.com> Cc: Steve Muckle <smuckle@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Todd Kjos <tkjos@android.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Link: http://lkml.kernel.org/r/20180309095245.11071-2-patrick.bellasi@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-03-09 09:52:42 +00:00
#endif /* CONFIG_SMP */
static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
{
#ifdef CONFIG_SCHED_DEBUG
s64 d = se->vruntime - cfs_rq->min_vruntime;
if (d < 0)
d = -d;
if (d > 3*sysctl_sched_latency)
schedstat_inc(cfs_rq->nr_spread_over);
#endif
}
static void
place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
{
u64 vruntime = cfs_rq->min_vruntime;
/*
* The 'current' period is already promised to the current tasks,
* however the extra weight of the new task will slow them down a
* little, place the new task so that it fits in the slot that
* stays open at the end.
*/
if (initial && sched_feat(START_DEBIT))
vruntime += sched_vslice(cfs_rq, se);
/* sleeps up to a single latency don't count. */
if (!initial) {
unsigned long thresh = sysctl_sched_latency;
/*
* Halve their sleep time's effect, to allow
* for a gentler effect of sleepers:
*/
if (sched_feat(GENTLE_FAIR_SLEEPERS))
thresh >>= 1;
vruntime -= thresh;
}
/* ensure we never gain time by being placed backwards. */
se->vruntime = max_vruntime(se->vruntime, vruntime);
}
static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
sched/debug: Make schedstats a runtime tunable that is disabled by default schedstats is very useful during debugging and performance tuning but it incurs overhead to calculate the stats. As such, even though it can be disabled at build time, it is often enabled as the information is useful. This patch adds a kernel command-line and sysctl tunable to enable or disable schedstats on demand (when it's built in). It is disabled by default as someone who knows they need it can also learn to enable it when necessary. The benefits are dependent on how scheduler-intensive the workload is. If it is then the patch reduces the number of cycles spent calculating the stats with a small benefit from reducing the cache footprint of the scheduler. These measurements were taken from a 48-core 2-socket machine with Xeon(R) E5-2670 v3 cpus although they were also tested on a single socket machine 8-core machine with Intel i7-3770 processors. netperf-tcp 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v3r1 Hmean 64 560.45 ( 0.00%) 575.98 ( 2.77%) Hmean 128 766.66 ( 0.00%) 795.79 ( 3.80%) Hmean 256 950.51 ( 0.00%) 981.50 ( 3.26%) Hmean 1024 1433.25 ( 0.00%) 1466.51 ( 2.32%) Hmean 2048 2810.54 ( 0.00%) 2879.75 ( 2.46%) Hmean 3312 4618.18 ( 0.00%) 4682.09 ( 1.38%) Hmean 4096 5306.42 ( 0.00%) 5346.39 ( 0.75%) Hmean 8192 10581.44 ( 0.00%) 10698.15 ( 1.10%) Hmean 16384 18857.70 ( 0.00%) 18937.61 ( 0.42%) Small gains here, UDP_STREAM showed nothing intresting and neither did the TCP_RR tests. The gains on the 8-core machine were very similar. tbench4 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v3r1 Hmean mb/sec-1 500.85 ( 0.00%) 522.43 ( 4.31%) Hmean mb/sec-2 984.66 ( 0.00%) 1018.19 ( 3.41%) Hmean mb/sec-4 1827.91 ( 0.00%) 1847.78 ( 1.09%) Hmean mb/sec-8 3561.36 ( 0.00%) 3611.28 ( 1.40%) Hmean mb/sec-16 5824.52 ( 0.00%) 5929.03 ( 1.79%) Hmean mb/sec-32 10943.10 ( 0.00%) 10802.83 ( -1.28%) Hmean mb/sec-64 15950.81 ( 0.00%) 16211.31 ( 1.63%) Hmean mb/sec-128 15302.17 ( 0.00%) 15445.11 ( 0.93%) Hmean mb/sec-256 14866.18 ( 0.00%) 15088.73 ( 1.50%) Hmean mb/sec-512 15223.31 ( 0.00%) 15373.69 ( 0.99%) Hmean mb/sec-1024 14574.25 ( 0.00%) 14598.02 ( 0.16%) Hmean mb/sec-2048 13569.02 ( 0.00%) 13733.86 ( 1.21%) Hmean mb/sec-3072 12865.98 ( 0.00%) 13209.23 ( 2.67%) Small gains of 2-4% at low thread counts and otherwise flat. The gains on the 8-core machine were slightly different tbench4 on 8-core i7-3770 single socket machine Hmean mb/sec-1 442.59 ( 0.00%) 448.73 ( 1.39%) Hmean mb/sec-2 796.68 ( 0.00%) 794.39 ( -0.29%) Hmean mb/sec-4 1322.52 ( 0.00%) 1343.66 ( 1.60%) Hmean mb/sec-8 2611.65 ( 0.00%) 2694.86 ( 3.19%) Hmean mb/sec-16 2537.07 ( 0.00%) 2609.34 ( 2.85%) Hmean mb/sec-32 2506.02 ( 0.00%) 2578.18 ( 2.88%) Hmean mb/sec-64 2511.06 ( 0.00%) 2569.16 ( 2.31%) Hmean mb/sec-128 2313.38 ( 0.00%) 2395.50 ( 3.55%) Hmean mb/sec-256 2110.04 ( 0.00%) 2177.45 ( 3.19%) Hmean mb/sec-512 2072.51 ( 0.00%) 2053.97 ( -0.89%) In constract, this shows a relatively steady 2-3% gain at higher thread counts. Due to the nature of the patch and the type of workload, it's not a surprise that the result will depend on the CPU used. hackbench-pipes 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v3r1 Amean 1 0.0637 ( 0.00%) 0.0660 ( -3.59%) Amean 4 0.1229 ( 0.00%) 0.1181 ( 3.84%) Amean 7 0.1921 ( 0.00%) 0.1911 ( 0.52%) Amean 12 0.3117 ( 0.00%) 0.2923 ( 6.23%) Amean 21 0.4050 ( 0.00%) 0.3899 ( 3.74%) Amean 30 0.4586 ( 0.00%) 0.4433 ( 3.33%) Amean 48 0.5910 ( 0.00%) 0.5694 ( 3.65%) Amean 79 0.8663 ( 0.00%) 0.8626 ( 0.43%) Amean 110 1.1543 ( 0.00%) 1.1517 ( 0.22%) Amean 141 1.4457 ( 0.00%) 1.4290 ( 1.16%) Amean 172 1.7090 ( 0.00%) 1.6924 ( 0.97%) Amean 192 1.9126 ( 0.00%) 1.9089 ( 0.19%) Some small gains and losses and while the variance data is not included, it's close to the noise. The UMA machine did not show anything particularly different pipetest 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v2r2 Min Time 4.13 ( 0.00%) 3.99 ( 3.39%) 1st-qrtle Time 4.38 ( 0.00%) 4.27 ( 2.51%) 2nd-qrtle Time 4.46 ( 0.00%) 4.39 ( 1.57%) 3rd-qrtle Time 4.56 ( 0.00%) 4.51 ( 1.10%) Max-90% Time 4.67 ( 0.00%) 4.60 ( 1.50%) Max-93% Time 4.71 ( 0.00%) 4.65 ( 1.27%) Max-95% Time 4.74 ( 0.00%) 4.71 ( 0.63%) Max-99% Time 4.88 ( 0.00%) 4.79 ( 1.84%) Max Time 4.93 ( 0.00%) 4.83 ( 2.03%) Mean Time 4.48 ( 0.00%) 4.39 ( 1.91%) Best99%Mean Time 4.47 ( 0.00%) 4.39 ( 1.91%) Best95%Mean Time 4.46 ( 0.00%) 4.38 ( 1.93%) Best90%Mean Time 4.45 ( 0.00%) 4.36 ( 1.98%) Best50%Mean Time 4.36 ( 0.00%) 4.25 ( 2.49%) Best10%Mean Time 4.23 ( 0.00%) 4.10 ( 3.13%) Best5%Mean Time 4.19 ( 0.00%) 4.06 ( 3.20%) Best1%Mean Time 4.13 ( 0.00%) 4.00 ( 3.39%) Small improvement and similar gains were seen on the UMA machine. The gain is small but it stands to reason that doing less work in the scheduler is a good thing. The downside is that the lack of schedstats and tracepoints may be surprising to experts doing performance analysis until they find the existence of the schedstats= parameter or schedstats sysctl. It will be automatically activated for latencytop and sleep profiling to alleviate the problem. For tracepoints, there is a simple warning as it's not safe to activate schedstats in the context when it's known the tracepoint may be wanted but is unavailable. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Reviewed-by: Matt Fleming <matt@codeblueprint.co.uk> Reviewed-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <mgalbraith@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1454663316-22048-1-git-send-email-mgorman@techsingularity.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-02-05 09:08:36 +00:00
static inline void check_schedstat_required(void)
{
#ifdef CONFIG_SCHEDSTATS
if (schedstat_enabled())
return;
/* Force schedstat enabled if a dependent tracepoint is active */
if (trace_sched_stat_wait_enabled() ||
trace_sched_stat_sleep_enabled() ||
trace_sched_stat_iowait_enabled() ||
trace_sched_stat_blocked_enabled() ||
trace_sched_stat_runtime_enabled()) {
sched/debug: Fix deadlock when enabling sched events I see a hang when enabling sched events: echo 1 > /sys/kernel/debug/tracing/events/sched/enable The printk buffer shows: BUG: spinlock recursion on CPU#1, swapper/1/0 lock: 0xffff88007d5d8c00, .magic: dead4ead, .owner: swapper/1/0, .owner_cpu: 1 CPU: 1 PID: 0 Comm: swapper/1 Not tainted 4.7.0-rc2+ #1 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.8.1-20150318_183358- 04/01/2014 ... Call Trace: <IRQ> [<ffffffff8143d663>] dump_stack+0x85/0xc2 [<ffffffff81115948>] spin_dump+0x78/0xc0 [<ffffffff81115aea>] do_raw_spin_lock+0x11a/0x150 [<ffffffff81891471>] _raw_spin_lock+0x61/0x80 [<ffffffff810e5466>] ? try_to_wake_up+0x256/0x4e0 [<ffffffff810e5466>] try_to_wake_up+0x256/0x4e0 [<ffffffff81891a0a>] ? _raw_spin_unlock_irqrestore+0x4a/0x80 [<ffffffff810e5705>] wake_up_process+0x15/0x20 [<ffffffff810cebb4>] insert_work+0x84/0xc0 [<ffffffff810ced7f>] __queue_work+0x18f/0x660 [<ffffffff810cf9a6>] queue_work_on+0x46/0x90 [<ffffffffa00cd95b>] drm_fb_helper_dirty.isra.11+0xcb/0xe0 [drm_kms_helper] [<ffffffffa00cdac0>] drm_fb_helper_sys_imageblit+0x30/0x40 [drm_kms_helper] [<ffffffff814babcd>] soft_cursor+0x1ad/0x230 [<ffffffff814ba379>] bit_cursor+0x649/0x680 [<ffffffff814b9d30>] ? update_attr.isra.2+0x90/0x90 [<ffffffff814b5e6a>] fbcon_cursor+0x14a/0x1c0 [<ffffffff81555ef8>] hide_cursor+0x28/0x90 [<ffffffff81558b6f>] vt_console_print+0x3bf/0x3f0 [<ffffffff81122c63>] call_console_drivers.constprop.24+0x183/0x200 [<ffffffff811241f4>] console_unlock+0x3d4/0x610 [<ffffffff811247f5>] vprintk_emit+0x3c5/0x610 [<ffffffff81124bc9>] vprintk_default+0x29/0x40 [<ffffffff811e965b>] printk+0x57/0x73 [<ffffffff810f7a9e>] enqueue_entity+0xc2e/0xc70 [<ffffffff810f7b39>] enqueue_task_fair+0x59/0xab0 [<ffffffff8106dcd9>] ? kvm_sched_clock_read+0x9/0x20 [<ffffffff8103fb39>] ? sched_clock+0x9/0x10 [<ffffffff810e3fcc>] activate_task+0x5c/0xa0 [<ffffffff810e4514>] ttwu_do_activate+0x54/0xb0 [<ffffffff810e5cea>] sched_ttwu_pending+0x7a/0xb0 [<ffffffff810e5e51>] scheduler_ipi+0x61/0x170 [<ffffffff81059e7f>] smp_trace_reschedule_interrupt+0x4f/0x2a0 [<ffffffff81893ba6>] trace_reschedule_interrupt+0x96/0xa0 <EOI> [<ffffffff8106e0d6>] ? native_safe_halt+0x6/0x10 [<ffffffff8110fb1d>] ? trace_hardirqs_on+0xd/0x10 [<ffffffff81040ac0>] default_idle+0x20/0x1a0 [<ffffffff8104147f>] arch_cpu_idle+0xf/0x20 [<ffffffff81102f8f>] default_idle_call+0x2f/0x50 [<ffffffff8110332e>] cpu_startup_entry+0x37e/0x450 [<ffffffff8105af70>] start_secondary+0x160/0x1a0 Note the hang only occurs when echoing the above from a physical serial console, not from an ssh session. The bug is caused by a deadlock where the task is trying to grab the rq lock twice because printk()'s aren't safe in sched code. Signed-off-by: Josh Poimboeuf <jpoimboe@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Matt Fleming <matt@codeblueprint.co.uk> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: stable@vger.kernel.org Fixes: cb2517653fcc ("sched/debug: Make schedstats a runtime tunable that is disabled by default") Link: http://lkml.kernel.org/r/20160613073209.gdvdybiruljbkn3p@treble Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-06-13 07:32:09 +00:00
printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
sched/debug: Make schedstats a runtime tunable that is disabled by default schedstats is very useful during debugging and performance tuning but it incurs overhead to calculate the stats. As such, even though it can be disabled at build time, it is often enabled as the information is useful. This patch adds a kernel command-line and sysctl tunable to enable or disable schedstats on demand (when it's built in). It is disabled by default as someone who knows they need it can also learn to enable it when necessary. The benefits are dependent on how scheduler-intensive the workload is. If it is then the patch reduces the number of cycles spent calculating the stats with a small benefit from reducing the cache footprint of the scheduler. These measurements were taken from a 48-core 2-socket machine with Xeon(R) E5-2670 v3 cpus although they were also tested on a single socket machine 8-core machine with Intel i7-3770 processors. netperf-tcp 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v3r1 Hmean 64 560.45 ( 0.00%) 575.98 ( 2.77%) Hmean 128 766.66 ( 0.00%) 795.79 ( 3.80%) Hmean 256 950.51 ( 0.00%) 981.50 ( 3.26%) Hmean 1024 1433.25 ( 0.00%) 1466.51 ( 2.32%) Hmean 2048 2810.54 ( 0.00%) 2879.75 ( 2.46%) Hmean 3312 4618.18 ( 0.00%) 4682.09 ( 1.38%) Hmean 4096 5306.42 ( 0.00%) 5346.39 ( 0.75%) Hmean 8192 10581.44 ( 0.00%) 10698.15 ( 1.10%) Hmean 16384 18857.70 ( 0.00%) 18937.61 ( 0.42%) Small gains here, UDP_STREAM showed nothing intresting and neither did the TCP_RR tests. The gains on the 8-core machine were very similar. tbench4 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v3r1 Hmean mb/sec-1 500.85 ( 0.00%) 522.43 ( 4.31%) Hmean mb/sec-2 984.66 ( 0.00%) 1018.19 ( 3.41%) Hmean mb/sec-4 1827.91 ( 0.00%) 1847.78 ( 1.09%) Hmean mb/sec-8 3561.36 ( 0.00%) 3611.28 ( 1.40%) Hmean mb/sec-16 5824.52 ( 0.00%) 5929.03 ( 1.79%) Hmean mb/sec-32 10943.10 ( 0.00%) 10802.83 ( -1.28%) Hmean mb/sec-64 15950.81 ( 0.00%) 16211.31 ( 1.63%) Hmean mb/sec-128 15302.17 ( 0.00%) 15445.11 ( 0.93%) Hmean mb/sec-256 14866.18 ( 0.00%) 15088.73 ( 1.50%) Hmean mb/sec-512 15223.31 ( 0.00%) 15373.69 ( 0.99%) Hmean mb/sec-1024 14574.25 ( 0.00%) 14598.02 ( 0.16%) Hmean mb/sec-2048 13569.02 ( 0.00%) 13733.86 ( 1.21%) Hmean mb/sec-3072 12865.98 ( 0.00%) 13209.23 ( 2.67%) Small gains of 2-4% at low thread counts and otherwise flat. The gains on the 8-core machine were slightly different tbench4 on 8-core i7-3770 single socket machine Hmean mb/sec-1 442.59 ( 0.00%) 448.73 ( 1.39%) Hmean mb/sec-2 796.68 ( 0.00%) 794.39 ( -0.29%) Hmean mb/sec-4 1322.52 ( 0.00%) 1343.66 ( 1.60%) Hmean mb/sec-8 2611.65 ( 0.00%) 2694.86 ( 3.19%) Hmean mb/sec-16 2537.07 ( 0.00%) 2609.34 ( 2.85%) Hmean mb/sec-32 2506.02 ( 0.00%) 2578.18 ( 2.88%) Hmean mb/sec-64 2511.06 ( 0.00%) 2569.16 ( 2.31%) Hmean mb/sec-128 2313.38 ( 0.00%) 2395.50 ( 3.55%) Hmean mb/sec-256 2110.04 ( 0.00%) 2177.45 ( 3.19%) Hmean mb/sec-512 2072.51 ( 0.00%) 2053.97 ( -0.89%) In constract, this shows a relatively steady 2-3% gain at higher thread counts. Due to the nature of the patch and the type of workload, it's not a surprise that the result will depend on the CPU used. hackbench-pipes 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v3r1 Amean 1 0.0637 ( 0.00%) 0.0660 ( -3.59%) Amean 4 0.1229 ( 0.00%) 0.1181 ( 3.84%) Amean 7 0.1921 ( 0.00%) 0.1911 ( 0.52%) Amean 12 0.3117 ( 0.00%) 0.2923 ( 6.23%) Amean 21 0.4050 ( 0.00%) 0.3899 ( 3.74%) Amean 30 0.4586 ( 0.00%) 0.4433 ( 3.33%) Amean 48 0.5910 ( 0.00%) 0.5694 ( 3.65%) Amean 79 0.8663 ( 0.00%) 0.8626 ( 0.43%) Amean 110 1.1543 ( 0.00%) 1.1517 ( 0.22%) Amean 141 1.4457 ( 0.00%) 1.4290 ( 1.16%) Amean 172 1.7090 ( 0.00%) 1.6924 ( 0.97%) Amean 192 1.9126 ( 0.00%) 1.9089 ( 0.19%) Some small gains and losses and while the variance data is not included, it's close to the noise. The UMA machine did not show anything particularly different pipetest 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v2r2 Min Time 4.13 ( 0.00%) 3.99 ( 3.39%) 1st-qrtle Time 4.38 ( 0.00%) 4.27 ( 2.51%) 2nd-qrtle Time 4.46 ( 0.00%) 4.39 ( 1.57%) 3rd-qrtle Time 4.56 ( 0.00%) 4.51 ( 1.10%) Max-90% Time 4.67 ( 0.00%) 4.60 ( 1.50%) Max-93% Time 4.71 ( 0.00%) 4.65 ( 1.27%) Max-95% Time 4.74 ( 0.00%) 4.71 ( 0.63%) Max-99% Time 4.88 ( 0.00%) 4.79 ( 1.84%) Max Time 4.93 ( 0.00%) 4.83 ( 2.03%) Mean Time 4.48 ( 0.00%) 4.39 ( 1.91%) Best99%Mean Time 4.47 ( 0.00%) 4.39 ( 1.91%) Best95%Mean Time 4.46 ( 0.00%) 4.38 ( 1.93%) Best90%Mean Time 4.45 ( 0.00%) 4.36 ( 1.98%) Best50%Mean Time 4.36 ( 0.00%) 4.25 ( 2.49%) Best10%Mean Time 4.23 ( 0.00%) 4.10 ( 3.13%) Best5%Mean Time 4.19 ( 0.00%) 4.06 ( 3.20%) Best1%Mean Time 4.13 ( 0.00%) 4.00 ( 3.39%) Small improvement and similar gains were seen on the UMA machine. The gain is small but it stands to reason that doing less work in the scheduler is a good thing. The downside is that the lack of schedstats and tracepoints may be surprising to experts doing performance analysis until they find the existence of the schedstats= parameter or schedstats sysctl. It will be automatically activated for latencytop and sleep profiling to alleviate the problem. For tracepoints, there is a simple warning as it's not safe to activate schedstats in the context when it's known the tracepoint may be wanted but is unavailable. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Reviewed-by: Matt Fleming <matt@codeblueprint.co.uk> Reviewed-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <mgalbraith@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1454663316-22048-1-git-send-email-mgorman@techsingularity.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-02-05 09:08:36 +00:00
"stat_blocked and stat_runtime require the "
"kernel parameter schedstats=enable or "
sched/debug: Make schedstats a runtime tunable that is disabled by default schedstats is very useful during debugging and performance tuning but it incurs overhead to calculate the stats. As such, even though it can be disabled at build time, it is often enabled as the information is useful. This patch adds a kernel command-line and sysctl tunable to enable or disable schedstats on demand (when it's built in). It is disabled by default as someone who knows they need it can also learn to enable it when necessary. The benefits are dependent on how scheduler-intensive the workload is. If it is then the patch reduces the number of cycles spent calculating the stats with a small benefit from reducing the cache footprint of the scheduler. These measurements were taken from a 48-core 2-socket machine with Xeon(R) E5-2670 v3 cpus although they were also tested on a single socket machine 8-core machine with Intel i7-3770 processors. netperf-tcp 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v3r1 Hmean 64 560.45 ( 0.00%) 575.98 ( 2.77%) Hmean 128 766.66 ( 0.00%) 795.79 ( 3.80%) Hmean 256 950.51 ( 0.00%) 981.50 ( 3.26%) Hmean 1024 1433.25 ( 0.00%) 1466.51 ( 2.32%) Hmean 2048 2810.54 ( 0.00%) 2879.75 ( 2.46%) Hmean 3312 4618.18 ( 0.00%) 4682.09 ( 1.38%) Hmean 4096 5306.42 ( 0.00%) 5346.39 ( 0.75%) Hmean 8192 10581.44 ( 0.00%) 10698.15 ( 1.10%) Hmean 16384 18857.70 ( 0.00%) 18937.61 ( 0.42%) Small gains here, UDP_STREAM showed nothing intresting and neither did the TCP_RR tests. The gains on the 8-core machine were very similar. tbench4 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v3r1 Hmean mb/sec-1 500.85 ( 0.00%) 522.43 ( 4.31%) Hmean mb/sec-2 984.66 ( 0.00%) 1018.19 ( 3.41%) Hmean mb/sec-4 1827.91 ( 0.00%) 1847.78 ( 1.09%) Hmean mb/sec-8 3561.36 ( 0.00%) 3611.28 ( 1.40%) Hmean mb/sec-16 5824.52 ( 0.00%) 5929.03 ( 1.79%) Hmean mb/sec-32 10943.10 ( 0.00%) 10802.83 ( -1.28%) Hmean mb/sec-64 15950.81 ( 0.00%) 16211.31 ( 1.63%) Hmean mb/sec-128 15302.17 ( 0.00%) 15445.11 ( 0.93%) Hmean mb/sec-256 14866.18 ( 0.00%) 15088.73 ( 1.50%) Hmean mb/sec-512 15223.31 ( 0.00%) 15373.69 ( 0.99%) Hmean mb/sec-1024 14574.25 ( 0.00%) 14598.02 ( 0.16%) Hmean mb/sec-2048 13569.02 ( 0.00%) 13733.86 ( 1.21%) Hmean mb/sec-3072 12865.98 ( 0.00%) 13209.23 ( 2.67%) Small gains of 2-4% at low thread counts and otherwise flat. The gains on the 8-core machine were slightly different tbench4 on 8-core i7-3770 single socket machine Hmean mb/sec-1 442.59 ( 0.00%) 448.73 ( 1.39%) Hmean mb/sec-2 796.68 ( 0.00%) 794.39 ( -0.29%) Hmean mb/sec-4 1322.52 ( 0.00%) 1343.66 ( 1.60%) Hmean mb/sec-8 2611.65 ( 0.00%) 2694.86 ( 3.19%) Hmean mb/sec-16 2537.07 ( 0.00%) 2609.34 ( 2.85%) Hmean mb/sec-32 2506.02 ( 0.00%) 2578.18 ( 2.88%) Hmean mb/sec-64 2511.06 ( 0.00%) 2569.16 ( 2.31%) Hmean mb/sec-128 2313.38 ( 0.00%) 2395.50 ( 3.55%) Hmean mb/sec-256 2110.04 ( 0.00%) 2177.45 ( 3.19%) Hmean mb/sec-512 2072.51 ( 0.00%) 2053.97 ( -0.89%) In constract, this shows a relatively steady 2-3% gain at higher thread counts. Due to the nature of the patch and the type of workload, it's not a surprise that the result will depend on the CPU used. hackbench-pipes 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v3r1 Amean 1 0.0637 ( 0.00%) 0.0660 ( -3.59%) Amean 4 0.1229 ( 0.00%) 0.1181 ( 3.84%) Amean 7 0.1921 ( 0.00%) 0.1911 ( 0.52%) Amean 12 0.3117 ( 0.00%) 0.2923 ( 6.23%) Amean 21 0.4050 ( 0.00%) 0.3899 ( 3.74%) Amean 30 0.4586 ( 0.00%) 0.4433 ( 3.33%) Amean 48 0.5910 ( 0.00%) 0.5694 ( 3.65%) Amean 79 0.8663 ( 0.00%) 0.8626 ( 0.43%) Amean 110 1.1543 ( 0.00%) 1.1517 ( 0.22%) Amean 141 1.4457 ( 0.00%) 1.4290 ( 1.16%) Amean 172 1.7090 ( 0.00%) 1.6924 ( 0.97%) Amean 192 1.9126 ( 0.00%) 1.9089 ( 0.19%) Some small gains and losses and while the variance data is not included, it's close to the noise. The UMA machine did not show anything particularly different pipetest 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v2r2 Min Time 4.13 ( 0.00%) 3.99 ( 3.39%) 1st-qrtle Time 4.38 ( 0.00%) 4.27 ( 2.51%) 2nd-qrtle Time 4.46 ( 0.00%) 4.39 ( 1.57%) 3rd-qrtle Time 4.56 ( 0.00%) 4.51 ( 1.10%) Max-90% Time 4.67 ( 0.00%) 4.60 ( 1.50%) Max-93% Time 4.71 ( 0.00%) 4.65 ( 1.27%) Max-95% Time 4.74 ( 0.00%) 4.71 ( 0.63%) Max-99% Time 4.88 ( 0.00%) 4.79 ( 1.84%) Max Time 4.93 ( 0.00%) 4.83 ( 2.03%) Mean Time 4.48 ( 0.00%) 4.39 ( 1.91%) Best99%Mean Time 4.47 ( 0.00%) 4.39 ( 1.91%) Best95%Mean Time 4.46 ( 0.00%) 4.38 ( 1.93%) Best90%Mean Time 4.45 ( 0.00%) 4.36 ( 1.98%) Best50%Mean Time 4.36 ( 0.00%) 4.25 ( 2.49%) Best10%Mean Time 4.23 ( 0.00%) 4.10 ( 3.13%) Best5%Mean Time 4.19 ( 0.00%) 4.06 ( 3.20%) Best1%Mean Time 4.13 ( 0.00%) 4.00 ( 3.39%) Small improvement and similar gains were seen on the UMA machine. The gain is small but it stands to reason that doing less work in the scheduler is a good thing. The downside is that the lack of schedstats and tracepoints may be surprising to experts doing performance analysis until they find the existence of the schedstats= parameter or schedstats sysctl. It will be automatically activated for latencytop and sleep profiling to alleviate the problem. For tracepoints, there is a simple warning as it's not safe to activate schedstats in the context when it's known the tracepoint may be wanted but is unavailable. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Reviewed-by: Matt Fleming <matt@codeblueprint.co.uk> Reviewed-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <mgalbraith@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1454663316-22048-1-git-send-email-mgorman@techsingularity.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-02-05 09:08:36 +00:00
"kernel.sched_schedstats=1\n");
}
#endif
}
static inline bool cfs_bandwidth_used(void);
/*
* MIGRATION
*
* dequeue
* update_curr()
* update_min_vruntime()
* vruntime -= min_vruntime
*
* enqueue
* update_curr()
* update_min_vruntime()
* vruntime += min_vruntime
*
* this way the vruntime transition between RQs is done when both
* min_vruntime are up-to-date.
*
* WAKEUP (remote)
*
* ->migrate_task_rq_fair() (p->state == TASK_WAKING)
* vruntime -= min_vruntime
*
* enqueue
* update_curr()
* update_min_vruntime()
* vruntime += min_vruntime
*
* this way we don't have the most up-to-date min_vruntime on the originating
* CPU and an up-to-date min_vruntime on the destination CPU.
*/
static void
sched: Remove the cfs_rq dependency from set_task_cpu() In order to remove the cfs_rq dependency from set_task_cpu() we need to ensure the task is cfs_rq invariant for all callsites. The simple approach is to substract cfs_rq->min_vruntime from se->vruntime on dequeue, and add cfs_rq->min_vruntime on enqueue. However, this has the downside of breaking FAIR_SLEEPERS since we loose the old vruntime as we only maintain the relative position. To solve this, we observe that we only migrate runnable tasks, we do this using deactivate_task(.sleep=0) and activate_task(.wakeup=0), therefore we can restrain the min_vruntime invariance to that state. The only other case is wakeup balancing, since we want to maintain the old vruntime we cannot make it relative on dequeue, but since we don't migrate inactive tasks, we can do so right before we activate it again. This is where we need the new pre-wakeup hook, we need to call this while still holding the old rq->lock. We could fold it into ->select_task_rq(), but since that has multiple callsites and would obfuscate the locking requirements, that seems like a fudge. This leaves the fork() case, simply make sure that ->task_fork() leaves the ->vruntime in a relative state. This covers all cases where set_task_cpu() gets called, and ensures it sees a relative vruntime. Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> LKML-Reference: <20091216170518.191697025@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-12-16 17:04:41 +00:00
enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
{
sched/fair: Fix fairness issue on migration Pavan reported that in the presence of very light tasks (or cgroups) the placement of migrated tasks can cause severe fairness issues. The problem is that enqueue_entity() places the task before it updates time, thereby it can place the task far in the past (remember that light tasks will shoot virtual time forward at a high speed, so in relation to the pre-existing light task, we can land far in the past). This is done because update_curr() needs the current task, and we might be placing the current task. The obvious solution is to differentiate between the current and any other task; placing the current before we update time, and placing any other task after, such that !curr tasks end up at the current moment in time, and not in the past. This commit re-introduces the previously reverted commit: 3a47d5124a95 ("sched/fair: Fix fairness issue on migration") ... which is now safe to do, after we've also fixed another underlying bug first, in: sched/fair: Prepare to fix fairness problems on migration and cleaned up other details in the migration code: sched/core: Kill sched_class::task_waking Reported-by: Pavan Kondeti <pkondeti@codeaurora.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-05-11 17:27:56 +00:00
bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
bool curr = cfs_rq->curr == se;
sched: Remove the cfs_rq dependency from set_task_cpu() In order to remove the cfs_rq dependency from set_task_cpu() we need to ensure the task is cfs_rq invariant for all callsites. The simple approach is to substract cfs_rq->min_vruntime from se->vruntime on dequeue, and add cfs_rq->min_vruntime on enqueue. However, this has the downside of breaking FAIR_SLEEPERS since we loose the old vruntime as we only maintain the relative position. To solve this, we observe that we only migrate runnable tasks, we do this using deactivate_task(.sleep=0) and activate_task(.wakeup=0), therefore we can restrain the min_vruntime invariance to that state. The only other case is wakeup balancing, since we want to maintain the old vruntime we cannot make it relative on dequeue, but since we don't migrate inactive tasks, we can do so right before we activate it again. This is where we need the new pre-wakeup hook, we need to call this while still holding the old rq->lock. We could fold it into ->select_task_rq(), but since that has multiple callsites and would obfuscate the locking requirements, that seems like a fudge. This leaves the fork() case, simply make sure that ->task_fork() leaves the ->vruntime in a relative state. This covers all cases where set_task_cpu() gets called, and ensures it sees a relative vruntime. Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> LKML-Reference: <20091216170518.191697025@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-12-16 17:04:41 +00:00
/*
sched/fair: Fix fairness issue on migration Pavan reported that in the presence of very light tasks (or cgroups) the placement of migrated tasks can cause severe fairness issues. The problem is that enqueue_entity() places the task before it updates time, thereby it can place the task far in the past (remember that light tasks will shoot virtual time forward at a high speed, so in relation to the pre-existing light task, we can land far in the past). This is done because update_curr() needs the current task, and we might be placing the current task. The obvious solution is to differentiate between the current and any other task; placing the current before we update time, and placing any other task after, such that !curr tasks end up at the current moment in time, and not in the past. This commit re-introduces the previously reverted commit: 3a47d5124a95 ("sched/fair: Fix fairness issue on migration") ... which is now safe to do, after we've also fixed another underlying bug first, in: sched/fair: Prepare to fix fairness problems on migration and cleaned up other details in the migration code: sched/core: Kill sched_class::task_waking Reported-by: Pavan Kondeti <pkondeti@codeaurora.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-05-11 17:27:56 +00:00
* If we're the current task, we must renormalise before calling
* update_curr().
sched: Remove the cfs_rq dependency from set_task_cpu() In order to remove the cfs_rq dependency from set_task_cpu() we need to ensure the task is cfs_rq invariant for all callsites. The simple approach is to substract cfs_rq->min_vruntime from se->vruntime on dequeue, and add cfs_rq->min_vruntime on enqueue. However, this has the downside of breaking FAIR_SLEEPERS since we loose the old vruntime as we only maintain the relative position. To solve this, we observe that we only migrate runnable tasks, we do this using deactivate_task(.sleep=0) and activate_task(.wakeup=0), therefore we can restrain the min_vruntime invariance to that state. The only other case is wakeup balancing, since we want to maintain the old vruntime we cannot make it relative on dequeue, but since we don't migrate inactive tasks, we can do so right before we activate it again. This is where we need the new pre-wakeup hook, we need to call this while still holding the old rq->lock. We could fold it into ->select_task_rq(), but since that has multiple callsites and would obfuscate the locking requirements, that seems like a fudge. This leaves the fork() case, simply make sure that ->task_fork() leaves the ->vruntime in a relative state. This covers all cases where set_task_cpu() gets called, and ensures it sees a relative vruntime. Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> LKML-Reference: <20091216170518.191697025@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-12-16 17:04:41 +00:00
*/
sched/fair: Fix fairness issue on migration Pavan reported that in the presence of very light tasks (or cgroups) the placement of migrated tasks can cause severe fairness issues. The problem is that enqueue_entity() places the task before it updates time, thereby it can place the task far in the past (remember that light tasks will shoot virtual time forward at a high speed, so in relation to the pre-existing light task, we can land far in the past). This is done because update_curr() needs the current task, and we might be placing the current task. The obvious solution is to differentiate between the current and any other task; placing the current before we update time, and placing any other task after, such that !curr tasks end up at the current moment in time, and not in the past. This commit re-introduces the previously reverted commit: 3a47d5124a95 ("sched/fair: Fix fairness issue on migration") ... which is now safe to do, after we've also fixed another underlying bug first, in: sched/fair: Prepare to fix fairness problems on migration and cleaned up other details in the migration code: sched/core: Kill sched_class::task_waking Reported-by: Pavan Kondeti <pkondeti@codeaurora.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-05-11 17:27:56 +00:00
if (renorm && curr)
sched: Remove the cfs_rq dependency from set_task_cpu() In order to remove the cfs_rq dependency from set_task_cpu() we need to ensure the task is cfs_rq invariant for all callsites. The simple approach is to substract cfs_rq->min_vruntime from se->vruntime on dequeue, and add cfs_rq->min_vruntime on enqueue. However, this has the downside of breaking FAIR_SLEEPERS since we loose the old vruntime as we only maintain the relative position. To solve this, we observe that we only migrate runnable tasks, we do this using deactivate_task(.sleep=0) and activate_task(.wakeup=0), therefore we can restrain the min_vruntime invariance to that state. The only other case is wakeup balancing, since we want to maintain the old vruntime we cannot make it relative on dequeue, but since we don't migrate inactive tasks, we can do so right before we activate it again. This is where we need the new pre-wakeup hook, we need to call this while still holding the old rq->lock. We could fold it into ->select_task_rq(), but since that has multiple callsites and would obfuscate the locking requirements, that seems like a fudge. This leaves the fork() case, simply make sure that ->task_fork() leaves the ->vruntime in a relative state. This covers all cases where set_task_cpu() gets called, and ensures it sees a relative vruntime. Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> LKML-Reference: <20091216170518.191697025@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-12-16 17:04:41 +00:00
se->vruntime += cfs_rq->min_vruntime;
sched/fair: Fix fairness issue on migration Pavan reported that in the presence of very light tasks (or cgroups) the placement of migrated tasks can cause severe fairness issues. The problem is that enqueue_entity() places the task before it updates time, thereby it can place the task far in the past (remember that light tasks will shoot virtual time forward at a high speed, so in relation to the pre-existing light task, we can land far in the past). This is done because update_curr() needs the current task, and we might be placing the current task. The obvious solution is to differentiate between the current and any other task; placing the current before we update time, and placing any other task after, such that !curr tasks end up at the current moment in time, and not in the past. This commit re-introduces the previously reverted commit: 3a47d5124a95 ("sched/fair: Fix fairness issue on migration") ... which is now safe to do, after we've also fixed another underlying bug first, in: sched/fair: Prepare to fix fairness problems on migration and cleaned up other details in the migration code: sched/core: Kill sched_class::task_waking Reported-by: Pavan Kondeti <pkondeti@codeaurora.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-05-11 17:27:56 +00:00
update_curr(cfs_rq);
/*
sched/fair: Fix fairness issue on migration Pavan reported that in the presence of very light tasks (or cgroups) the placement of migrated tasks can cause severe fairness issues. The problem is that enqueue_entity() places the task before it updates time, thereby it can place the task far in the past (remember that light tasks will shoot virtual time forward at a high speed, so in relation to the pre-existing light task, we can land far in the past). This is done because update_curr() needs the current task, and we might be placing the current task. The obvious solution is to differentiate between the current and any other task; placing the current before we update time, and placing any other task after, such that !curr tasks end up at the current moment in time, and not in the past. This commit re-introduces the previously reverted commit: 3a47d5124a95 ("sched/fair: Fix fairness issue on migration") ... which is now safe to do, after we've also fixed another underlying bug first, in: sched/fair: Prepare to fix fairness problems on migration and cleaned up other details in the migration code: sched/core: Kill sched_class::task_waking Reported-by: Pavan Kondeti <pkondeti@codeaurora.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-05-11 17:27:56 +00:00
* Otherwise, renormalise after, such that we're placed at the current
* moment in time, instead of some random moment in the past. Being
* placed in the past could significantly boost this task to the
* fairness detriment of existing tasks.
*/
sched/fair: Fix fairness issue on migration Pavan reported that in the presence of very light tasks (or cgroups) the placement of migrated tasks can cause severe fairness issues. The problem is that enqueue_entity() places the task before it updates time, thereby it can place the task far in the past (remember that light tasks will shoot virtual time forward at a high speed, so in relation to the pre-existing light task, we can land far in the past). This is done because update_curr() needs the current task, and we might be placing the current task. The obvious solution is to differentiate between the current and any other task; placing the current before we update time, and placing any other task after, such that !curr tasks end up at the current moment in time, and not in the past. This commit re-introduces the previously reverted commit: 3a47d5124a95 ("sched/fair: Fix fairness issue on migration") ... which is now safe to do, after we've also fixed another underlying bug first, in: sched/fair: Prepare to fix fairness problems on migration and cleaned up other details in the migration code: sched/core: Kill sched_class::task_waking Reported-by: Pavan Kondeti <pkondeti@codeaurora.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-05-11 17:27:56 +00:00
if (renorm && !curr)
se->vruntime += cfs_rq->min_vruntime;
sched/core: Fix group_entity's share update The update of the share of a cfs_rq is done when its load_avg is updated but before the group_entity's load_avg has been updated for the past time slot. This generates wrong load_avg accounting which can be significant when small tasks are involved in the scheduling. Let take the example of a task a that is dequeued of its task group A: root (cfs_rq) \ (se) A (cfs_rq) \ (se) a Task "a" was the only task in task group A which becomes idle when a is dequeued. We have the sequence: - dequeue_entity a->se - update_load_avg(a->se) - dequeue_entity_load_avg(A->cfs_rq, a->se) - update_cfs_shares(A->cfs_rq) A->cfs_rq->load.weight == 0 A->se->load.weight is updated with the new share (0 in this case) - dequeue_entity A->se - update_load_avg(A->se) but its weight is now null so the last time slot (up to a tick) will be accounted with a weight of 0 instead of its real weight during the time slot. The last time slot will be accounted as an idle one whereas it was a running one. If the running time of task a is short enough that no tick happens when it runs, all running time of group entity A->se will be accounted as idle time. Instead, we should update the share of a cfs_rq (in fact the weight of its group entity) only after having updated the load_avg of the group_entity. update_cfs_shares() now takes the sched_entity as a parameter instead of the cfs_rq, and the weight of the group_entity is updated only once its load_avg has been synced with current time. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: pjt@google.com Link: http://lkml.kernel.org/r/1482335426-7664-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-12-21 15:50:26 +00:00
/*
* When enqueuing a sched_entity, we must:
* - Update loads to have both entity and cfs_rq synced with now.
sched/pelt: Add a new runnable average signal Now that runnable_load_avg has been removed, we can replace it by a new signal that will highlight the runnable pressure on a cfs_rq. This signal track the waiting time of tasks on rq and can help to better define the state of rqs. At now, only util_avg is used to define the state of a rq: A rq with more that around 80% of utilization and more than 1 tasks is considered as overloaded. But the util_avg signal of a rq can become temporaly low after that a task migrated onto another rq which can bias the classification of the rq. When tasks compete for the same rq, their runnable average signal will be higher than util_avg as it will include the waiting time and we can use this signal to better classify cfs_rqs. The new runnable_avg will track the runnable time of a task which simply adds the waiting time to the running time. The runnable _avg of cfs_rq will be the /Sum of se's runnable_avg and the runnable_avg of group entity will follow the one of the rq similarly to util_avg. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: "Dietmar Eggemann <dietmar.eggemann@arm.com>" Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Valentin Schneider <valentin.schneider@arm.com> Cc: Phil Auld <pauld@redhat.com> Cc: Hillf Danton <hdanton@sina.com> Link: https://lore.kernel.org/r/20200224095223.13361-9-mgorman@techsingularity.net
2020-02-24 09:52:18 +00:00
* - Add its load to cfs_rq->runnable_avg
sched/core: Fix group_entity's share update The update of the share of a cfs_rq is done when its load_avg is updated but before the group_entity's load_avg has been updated for the past time slot. This generates wrong load_avg accounting which can be significant when small tasks are involved in the scheduling. Let take the example of a task a that is dequeued of its task group A: root (cfs_rq) \ (se) A (cfs_rq) \ (se) a Task "a" was the only task in task group A which becomes idle when a is dequeued. We have the sequence: - dequeue_entity a->se - update_load_avg(a->se) - dequeue_entity_load_avg(A->cfs_rq, a->se) - update_cfs_shares(A->cfs_rq) A->cfs_rq->load.weight == 0 A->se->load.weight is updated with the new share (0 in this case) - dequeue_entity A->se - update_load_avg(A->se) but its weight is now null so the last time slot (up to a tick) will be accounted with a weight of 0 instead of its real weight during the time slot. The last time slot will be accounted as an idle one whereas it was a running one. If the running time of task a is short enough that no tick happens when it runs, all running time of group entity A->se will be accounted as idle time. Instead, we should update the share of a cfs_rq (in fact the weight of its group entity) only after having updated the load_avg of the group_entity. update_cfs_shares() now takes the sched_entity as a parameter instead of the cfs_rq, and the weight of the group_entity is updated only once its load_avg has been synced with current time. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: pjt@google.com Link: http://lkml.kernel.org/r/1482335426-7664-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-12-21 15:50:26 +00:00
* - For group_entity, update its weight to reflect the new share of
* its group cfs_rq
* - Add its new weight to cfs_rq->load.weight
*/
update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
sched/pelt: Add a new runnable average signal Now that runnable_load_avg has been removed, we can replace it by a new signal that will highlight the runnable pressure on a cfs_rq. This signal track the waiting time of tasks on rq and can help to better define the state of rqs. At now, only util_avg is used to define the state of a rq: A rq with more that around 80% of utilization and more than 1 tasks is considered as overloaded. But the util_avg signal of a rq can become temporaly low after that a task migrated onto another rq which can bias the classification of the rq. When tasks compete for the same rq, their runnable average signal will be higher than util_avg as it will include the waiting time and we can use this signal to better classify cfs_rqs. The new runnable_avg will track the runnable time of a task which simply adds the waiting time to the running time. The runnable _avg of cfs_rq will be the /Sum of se's runnable_avg and the runnable_avg of group entity will follow the one of the rq similarly to util_avg. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: "Dietmar Eggemann <dietmar.eggemann@arm.com>" Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Valentin Schneider <valentin.schneider@arm.com> Cc: Phil Auld <pauld@redhat.com> Cc: Hillf Danton <hdanton@sina.com> Link: https://lore.kernel.org/r/20200224095223.13361-9-mgorman@techsingularity.net
2020-02-24 09:52:18 +00:00
se_update_runnable(se);
update_cfs_group(se);
account_entity_enqueue(cfs_rq, se);
if (flags & ENQUEUE_WAKEUP)
place_entity(cfs_rq, se, 0);
sched/debug: Make schedstats a runtime tunable that is disabled by default schedstats is very useful during debugging and performance tuning but it incurs overhead to calculate the stats. As such, even though it can be disabled at build time, it is often enabled as the information is useful. This patch adds a kernel command-line and sysctl tunable to enable or disable schedstats on demand (when it's built in). It is disabled by default as someone who knows they need it can also learn to enable it when necessary. The benefits are dependent on how scheduler-intensive the workload is. If it is then the patch reduces the number of cycles spent calculating the stats with a small benefit from reducing the cache footprint of the scheduler. These measurements were taken from a 48-core 2-socket machine with Xeon(R) E5-2670 v3 cpus although they were also tested on a single socket machine 8-core machine with Intel i7-3770 processors. netperf-tcp 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v3r1 Hmean 64 560.45 ( 0.00%) 575.98 ( 2.77%) Hmean 128 766.66 ( 0.00%) 795.79 ( 3.80%) Hmean 256 950.51 ( 0.00%) 981.50 ( 3.26%) Hmean 1024 1433.25 ( 0.00%) 1466.51 ( 2.32%) Hmean 2048 2810.54 ( 0.00%) 2879.75 ( 2.46%) Hmean 3312 4618.18 ( 0.00%) 4682.09 ( 1.38%) Hmean 4096 5306.42 ( 0.00%) 5346.39 ( 0.75%) Hmean 8192 10581.44 ( 0.00%) 10698.15 ( 1.10%) Hmean 16384 18857.70 ( 0.00%) 18937.61 ( 0.42%) Small gains here, UDP_STREAM showed nothing intresting and neither did the TCP_RR tests. The gains on the 8-core machine were very similar. tbench4 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v3r1 Hmean mb/sec-1 500.85 ( 0.00%) 522.43 ( 4.31%) Hmean mb/sec-2 984.66 ( 0.00%) 1018.19 ( 3.41%) Hmean mb/sec-4 1827.91 ( 0.00%) 1847.78 ( 1.09%) Hmean mb/sec-8 3561.36 ( 0.00%) 3611.28 ( 1.40%) Hmean mb/sec-16 5824.52 ( 0.00%) 5929.03 ( 1.79%) Hmean mb/sec-32 10943.10 ( 0.00%) 10802.83 ( -1.28%) Hmean mb/sec-64 15950.81 ( 0.00%) 16211.31 ( 1.63%) Hmean mb/sec-128 15302.17 ( 0.00%) 15445.11 ( 0.93%) Hmean mb/sec-256 14866.18 ( 0.00%) 15088.73 ( 1.50%) Hmean mb/sec-512 15223.31 ( 0.00%) 15373.69 ( 0.99%) Hmean mb/sec-1024 14574.25 ( 0.00%) 14598.02 ( 0.16%) Hmean mb/sec-2048 13569.02 ( 0.00%) 13733.86 ( 1.21%) Hmean mb/sec-3072 12865.98 ( 0.00%) 13209.23 ( 2.67%) Small gains of 2-4% at low thread counts and otherwise flat. The gains on the 8-core machine were slightly different tbench4 on 8-core i7-3770 single socket machine Hmean mb/sec-1 442.59 ( 0.00%) 448.73 ( 1.39%) Hmean mb/sec-2 796.68 ( 0.00%) 794.39 ( -0.29%) Hmean mb/sec-4 1322.52 ( 0.00%) 1343.66 ( 1.60%) Hmean mb/sec-8 2611.65 ( 0.00%) 2694.86 ( 3.19%) Hmean mb/sec-16 2537.07 ( 0.00%) 2609.34 ( 2.85%) Hmean mb/sec-32 2506.02 ( 0.00%) 2578.18 ( 2.88%) Hmean mb/sec-64 2511.06 ( 0.00%) 2569.16 ( 2.31%) Hmean mb/sec-128 2313.38 ( 0.00%) 2395.50 ( 3.55%) Hmean mb/sec-256 2110.04 ( 0.00%) 2177.45 ( 3.19%) Hmean mb/sec-512 2072.51 ( 0.00%) 2053.97 ( -0.89%) In constract, this shows a relatively steady 2-3% gain at higher thread counts. Due to the nature of the patch and the type of workload, it's not a surprise that the result will depend on the CPU used. hackbench-pipes 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v3r1 Amean 1 0.0637 ( 0.00%) 0.0660 ( -3.59%) Amean 4 0.1229 ( 0.00%) 0.1181 ( 3.84%) Amean 7 0.1921 ( 0.00%) 0.1911 ( 0.52%) Amean 12 0.3117 ( 0.00%) 0.2923 ( 6.23%) Amean 21 0.4050 ( 0.00%) 0.3899 ( 3.74%) Amean 30 0.4586 ( 0.00%) 0.4433 ( 3.33%) Amean 48 0.5910 ( 0.00%) 0.5694 ( 3.65%) Amean 79 0.8663 ( 0.00%) 0.8626 ( 0.43%) Amean 110 1.1543 ( 0.00%) 1.1517 ( 0.22%) Amean 141 1.4457 ( 0.00%) 1.4290 ( 1.16%) Amean 172 1.7090 ( 0.00%) 1.6924 ( 0.97%) Amean 192 1.9126 ( 0.00%) 1.9089 ( 0.19%) Some small gains and losses and while the variance data is not included, it's close to the noise. The UMA machine did not show anything particularly different pipetest 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v2r2 Min Time 4.13 ( 0.00%) 3.99 ( 3.39%) 1st-qrtle Time 4.38 ( 0.00%) 4.27 ( 2.51%) 2nd-qrtle Time 4.46 ( 0.00%) 4.39 ( 1.57%) 3rd-qrtle Time 4.56 ( 0.00%) 4.51 ( 1.10%) Max-90% Time 4.67 ( 0.00%) 4.60 ( 1.50%) Max-93% Time 4.71 ( 0.00%) 4.65 ( 1.27%) Max-95% Time 4.74 ( 0.00%) 4.71 ( 0.63%) Max-99% Time 4.88 ( 0.00%) 4.79 ( 1.84%) Max Time 4.93 ( 0.00%) 4.83 ( 2.03%) Mean Time 4.48 ( 0.00%) 4.39 ( 1.91%) Best99%Mean Time 4.47 ( 0.00%) 4.39 ( 1.91%) Best95%Mean Time 4.46 ( 0.00%) 4.38 ( 1.93%) Best90%Mean Time 4.45 ( 0.00%) 4.36 ( 1.98%) Best50%Mean Time 4.36 ( 0.00%) 4.25 ( 2.49%) Best10%Mean Time 4.23 ( 0.00%) 4.10 ( 3.13%) Best5%Mean Time 4.19 ( 0.00%) 4.06 ( 3.20%) Best1%Mean Time 4.13 ( 0.00%) 4.00 ( 3.39%) Small improvement and similar gains were seen on the UMA machine. The gain is small but it stands to reason that doing less work in the scheduler is a good thing. The downside is that the lack of schedstats and tracepoints may be surprising to experts doing performance analysis until they find the existence of the schedstats= parameter or schedstats sysctl. It will be automatically activated for latencytop and sleep profiling to alleviate the problem. For tracepoints, there is a simple warning as it's not safe to activate schedstats in the context when it's known the tracepoint may be wanted but is unavailable. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Reviewed-by: Matt Fleming <matt@codeblueprint.co.uk> Reviewed-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <mgalbraith@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1454663316-22048-1-git-send-email-mgorman@techsingularity.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-02-05 09:08:36 +00:00
check_schedstat_required();
update_stats_enqueue(cfs_rq, se, flags);
check_spread(cfs_rq, se);
sched/fair: Fix fairness issue on migration Pavan reported that in the presence of very light tasks (or cgroups) the placement of migrated tasks can cause severe fairness issues. The problem is that enqueue_entity() places the task before it updates time, thereby it can place the task far in the past (remember that light tasks will shoot virtual time forward at a high speed, so in relation to the pre-existing light task, we can land far in the past). This is done because update_curr() needs the current task, and we might be placing the current task. The obvious solution is to differentiate between the current and any other task; placing the current before we update time, and placing any other task after, such that !curr tasks end up at the current moment in time, and not in the past. This commit re-introduces the previously reverted commit: 3a47d5124a95 ("sched/fair: Fix fairness issue on migration") ... which is now safe to do, after we've also fixed another underlying bug first, in: sched/fair: Prepare to fix fairness problems on migration and cleaned up other details in the migration code: sched/core: Kill sched_class::task_waking Reported-by: Pavan Kondeti <pkondeti@codeaurora.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-05-11 17:27:56 +00:00
if (!curr)
__enqueue_entity(cfs_rq, se);
se->on_rq = 1;
/*
* When bandwidth control is enabled, cfs might have been removed
* because of a parent been throttled but cfs->nr_running > 1. Try to
* add it unconditionnally.
*/
if (cfs_rq->nr_running == 1 || cfs_bandwidth_used())
list_add_leaf_cfs_rq(cfs_rq);
if (cfs_rq->nr_running == 1)
check_enqueue_throttle(cfs_rq);
}
static void __clear_buddies_last(struct sched_entity *se)
{
for_each_sched_entity(se) {
struct cfs_rq *cfs_rq = cfs_rq_of(se);
if (cfs_rq->last != se)
break;
cfs_rq->last = NULL;
}
}
static void __clear_buddies_next(struct sched_entity *se)
{
for_each_sched_entity(se) {
struct cfs_rq *cfs_rq = cfs_rq_of(se);
if (cfs_rq->next != se)
break;
cfs_rq->next = NULL;
}
}
static void __clear_buddies_skip(struct sched_entity *se)
{
for_each_sched_entity(se) {
struct cfs_rq *cfs_rq = cfs_rq_of(se);
if (cfs_rq->skip != se)
break;
cfs_rq->skip = NULL;
}
}
static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
{
if (cfs_rq->last == se)
__clear_buddies_last(se);
if (cfs_rq->next == se)
__clear_buddies_next(se);
if (cfs_rq->skip == se)
__clear_buddies_skip(se);
}
static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
sched: Return unused runtime on group dequeue When a local cfs_rq blocks we return the majority of its remaining quota to the global bandwidth pool for use by other runqueues. We do this only when the quota is current and there is more than min_cfs_rq_quota [1ms by default] of runtime remaining on the rq. In the case where there are throttled runqueues and we have sufficient bandwidth to meter out a slice, a second timer is kicked off to handle this delivery, unthrottling where appropriate. Using a 'worst case' antagonist which executes on each cpu for 1ms before moving onto the next on a fairly large machine: no quota generations: 197.47 ms /cgroup/a/cpuacct.usage 199.46 ms /cgroup/a/cpuacct.usage 205.46 ms /cgroup/a/cpuacct.usage 198.46 ms /cgroup/a/cpuacct.usage 208.39 ms /cgroup/a/cpuacct.usage Since we are allowed to use "stale" quota our usage is effectively bounded by the rate of input into the global pool and performance is relatively stable. with quota generations [1s increments]: 119.58 ms /cgroup/a/cpuacct.usage 119.65 ms /cgroup/a/cpuacct.usage 119.64 ms /cgroup/a/cpuacct.usage 119.63 ms /cgroup/a/cpuacct.usage 119.60 ms /cgroup/a/cpuacct.usage The large deficit here is due to quota generations (/intentionally/) preventing us from now using previously stranded slack quota. The cost is that this quota becomes unavailable. with quota generations and quota return: 200.09 ms /cgroup/a/cpuacct.usage 200.09 ms /cgroup/a/cpuacct.usage 198.09 ms /cgroup/a/cpuacct.usage 200.09 ms /cgroup/a/cpuacct.usage 200.06 ms /cgroup/a/cpuacct.usage By returning unused quota we're able to both stably consume our desired quota and prevent unintentional overages due to the abuse of slack quota from previous quota periods (especially on a large machine). Signed-off-by: Paul Turner <pjt@google.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/20110721184758.306848658@google.com Signed-off-by: Ingo Molnar <mingo@elte.hu>
2011-07-21 16:43:41 +00:00
static void
dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
{
/*
* Update run-time statistics of the 'current'.
*/
update_curr(cfs_rq);
sched/core: Fix group_entity's share update The update of the share of a cfs_rq is done when its load_avg is updated but before the group_entity's load_avg has been updated for the past time slot. This generates wrong load_avg accounting which can be significant when small tasks are involved in the scheduling. Let take the example of a task a that is dequeued of its task group A: root (cfs_rq) \ (se) A (cfs_rq) \ (se) a Task "a" was the only task in task group A which becomes idle when a is dequeued. We have the sequence: - dequeue_entity a->se - update_load_avg(a->se) - dequeue_entity_load_avg(A->cfs_rq, a->se) - update_cfs_shares(A->cfs_rq) A->cfs_rq->load.weight == 0 A->se->load.weight is updated with the new share (0 in this case) - dequeue_entity A->se - update_load_avg(A->se) but its weight is now null so the last time slot (up to a tick) will be accounted with a weight of 0 instead of its real weight during the time slot. The last time slot will be accounted as an idle one whereas it was a running one. If the running time of task a is short enough that no tick happens when it runs, all running time of group entity A->se will be accounted as idle time. Instead, we should update the share of a cfs_rq (in fact the weight of its group entity) only after having updated the load_avg of the group_entity. update_cfs_shares() now takes the sched_entity as a parameter instead of the cfs_rq, and the weight of the group_entity is updated only once its load_avg has been synced with current time. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: pjt@google.com Link: http://lkml.kernel.org/r/1482335426-7664-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-12-21 15:50:26 +00:00
/*
* When dequeuing a sched_entity, we must:
* - Update loads to have both entity and cfs_rq synced with now.
sched/pelt: Add a new runnable average signal Now that runnable_load_avg has been removed, we can replace it by a new signal that will highlight the runnable pressure on a cfs_rq. This signal track the waiting time of tasks on rq and can help to better define the state of rqs. At now, only util_avg is used to define the state of a rq: A rq with more that around 80% of utilization and more than 1 tasks is considered as overloaded. But the util_avg signal of a rq can become temporaly low after that a task migrated onto another rq which can bias the classification of the rq. When tasks compete for the same rq, their runnable average signal will be higher than util_avg as it will include the waiting time and we can use this signal to better classify cfs_rqs. The new runnable_avg will track the runnable time of a task which simply adds the waiting time to the running time. The runnable _avg of cfs_rq will be the /Sum of se's runnable_avg and the runnable_avg of group entity will follow the one of the rq similarly to util_avg. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: "Dietmar Eggemann <dietmar.eggemann@arm.com>" Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Valentin Schneider <valentin.schneider@arm.com> Cc: Phil Auld <pauld@redhat.com> Cc: Hillf Danton <hdanton@sina.com> Link: https://lore.kernel.org/r/20200224095223.13361-9-mgorman@techsingularity.net
2020-02-24 09:52:18 +00:00
* - Subtract its load from the cfs_rq->runnable_avg.
* - Subtract its previous weight from cfs_rq->load.weight.
sched/core: Fix group_entity's share update The update of the share of a cfs_rq is done when its load_avg is updated but before the group_entity's load_avg has been updated for the past time slot. This generates wrong load_avg accounting which can be significant when small tasks are involved in the scheduling. Let take the example of a task a that is dequeued of its task group A: root (cfs_rq) \ (se) A (cfs_rq) \ (se) a Task "a" was the only task in task group A which becomes idle when a is dequeued. We have the sequence: - dequeue_entity a->se - update_load_avg(a->se) - dequeue_entity_load_avg(A->cfs_rq, a->se) - update_cfs_shares(A->cfs_rq) A->cfs_rq->load.weight == 0 A->se->load.weight is updated with the new share (0 in this case) - dequeue_entity A->se - update_load_avg(A->se) but its weight is now null so the last time slot (up to a tick) will be accounted with a weight of 0 instead of its real weight during the time slot. The last time slot will be accounted as an idle one whereas it was a running one. If the running time of task a is short enough that no tick happens when it runs, all running time of group entity A->se will be accounted as idle time. Instead, we should update the share of a cfs_rq (in fact the weight of its group entity) only after having updated the load_avg of the group_entity. update_cfs_shares() now takes the sched_entity as a parameter instead of the cfs_rq, and the weight of the group_entity is updated only once its load_avg has been synced with current time. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: pjt@google.com Link: http://lkml.kernel.org/r/1482335426-7664-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-12-21 15:50:26 +00:00
* - For group entity, update its weight to reflect the new share
* of its group cfs_rq.
*/
update_load_avg(cfs_rq, se, UPDATE_TG);
sched/pelt: Add a new runnable average signal Now that runnable_load_avg has been removed, we can replace it by a new signal that will highlight the runnable pressure on a cfs_rq. This signal track the waiting time of tasks on rq and can help to better define the state of rqs. At now, only util_avg is used to define the state of a rq: A rq with more that around 80% of utilization and more than 1 tasks is considered as overloaded. But the util_avg signal of a rq can become temporaly low after that a task migrated onto another rq which can bias the classification of the rq. When tasks compete for the same rq, their runnable average signal will be higher than util_avg as it will include the waiting time and we can use this signal to better classify cfs_rqs. The new runnable_avg will track the runnable time of a task which simply adds the waiting time to the running time. The runnable _avg of cfs_rq will be the /Sum of se's runnable_avg and the runnable_avg of group entity will follow the one of the rq similarly to util_avg. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: "Dietmar Eggemann <dietmar.eggemann@arm.com>" Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Valentin Schneider <valentin.schneider@arm.com> Cc: Phil Auld <pauld@redhat.com> Cc: Hillf Danton <hdanton@sina.com> Link: https://lore.kernel.org/r/20200224095223.13361-9-mgorman@techsingularity.net
2020-02-24 09:52:18 +00:00
se_update_runnable(se);
update_stats_dequeue(cfs_rq, se, flags);
clear_buddies(cfs_rq, se);
if (se != cfs_rq->curr)
__dequeue_entity(cfs_rq, se);
se->on_rq = 0;
account_entity_dequeue(cfs_rq, se);
sched: Remove the cfs_rq dependency from set_task_cpu() In order to remove the cfs_rq dependency from set_task_cpu() we need to ensure the task is cfs_rq invariant for all callsites. The simple approach is to substract cfs_rq->min_vruntime from se->vruntime on dequeue, and add cfs_rq->min_vruntime on enqueue. However, this has the downside of breaking FAIR_SLEEPERS since we loose the old vruntime as we only maintain the relative position. To solve this, we observe that we only migrate runnable tasks, we do this using deactivate_task(.sleep=0) and activate_task(.wakeup=0), therefore we can restrain the min_vruntime invariance to that state. The only other case is wakeup balancing, since we want to maintain the old vruntime we cannot make it relative on dequeue, but since we don't migrate inactive tasks, we can do so right before we activate it again. This is where we need the new pre-wakeup hook, we need to call this while still holding the old rq->lock. We could fold it into ->select_task_rq(), but since that has multiple callsites and would obfuscate the locking requirements, that seems like a fudge. This leaves the fork() case, simply make sure that ->task_fork() leaves the ->vruntime in a relative state. This covers all cases where set_task_cpu() gets called, and ensures it sees a relative vruntime. Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> LKML-Reference: <20091216170518.191697025@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-12-16 17:04:41 +00:00
/*
* Normalize after update_curr(); which will also have moved
* min_vruntime if @se is the one holding it back. But before doing
* update_min_vruntime() again, which will discount @se's position and
* can move min_vruntime forward still more.
sched: Remove the cfs_rq dependency from set_task_cpu() In order to remove the cfs_rq dependency from set_task_cpu() we need to ensure the task is cfs_rq invariant for all callsites. The simple approach is to substract cfs_rq->min_vruntime from se->vruntime on dequeue, and add cfs_rq->min_vruntime on enqueue. However, this has the downside of breaking FAIR_SLEEPERS since we loose the old vruntime as we only maintain the relative position. To solve this, we observe that we only migrate runnable tasks, we do this using deactivate_task(.sleep=0) and activate_task(.wakeup=0), therefore we can restrain the min_vruntime invariance to that state. The only other case is wakeup balancing, since we want to maintain the old vruntime we cannot make it relative on dequeue, but since we don't migrate inactive tasks, we can do so right before we activate it again. This is where we need the new pre-wakeup hook, we need to call this while still holding the old rq->lock. We could fold it into ->select_task_rq(), but since that has multiple callsites and would obfuscate the locking requirements, that seems like a fudge. This leaves the fork() case, simply make sure that ->task_fork() leaves the ->vruntime in a relative state. This covers all cases where set_task_cpu() gets called, and ensures it sees a relative vruntime. Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> LKML-Reference: <20091216170518.191697025@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-12-16 17:04:41 +00:00
*/
if (!(flags & DEQUEUE_SLEEP))
sched: Remove the cfs_rq dependency from set_task_cpu() In order to remove the cfs_rq dependency from set_task_cpu() we need to ensure the task is cfs_rq invariant for all callsites. The simple approach is to substract cfs_rq->min_vruntime from se->vruntime on dequeue, and add cfs_rq->min_vruntime on enqueue. However, this has the downside of breaking FAIR_SLEEPERS since we loose the old vruntime as we only maintain the relative position. To solve this, we observe that we only migrate runnable tasks, we do this using deactivate_task(.sleep=0) and activate_task(.wakeup=0), therefore we can restrain the min_vruntime invariance to that state. The only other case is wakeup balancing, since we want to maintain the old vruntime we cannot make it relative on dequeue, but since we don't migrate inactive tasks, we can do so right before we activate it again. This is where we need the new pre-wakeup hook, we need to call this while still holding the old rq->lock. We could fold it into ->select_task_rq(), but since that has multiple callsites and would obfuscate the locking requirements, that seems like a fudge. This leaves the fork() case, simply make sure that ->task_fork() leaves the ->vruntime in a relative state. This covers all cases where set_task_cpu() gets called, and ensures it sees a relative vruntime. Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> LKML-Reference: <20091216170518.191697025@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-12-16 17:04:41 +00:00
se->vruntime -= cfs_rq->min_vruntime;
sched: Return unused runtime on group dequeue When a local cfs_rq blocks we return the majority of its remaining quota to the global bandwidth pool for use by other runqueues. We do this only when the quota is current and there is more than min_cfs_rq_quota [1ms by default] of runtime remaining on the rq. In the case where there are throttled runqueues and we have sufficient bandwidth to meter out a slice, a second timer is kicked off to handle this delivery, unthrottling where appropriate. Using a 'worst case' antagonist which executes on each cpu for 1ms before moving onto the next on a fairly large machine: no quota generations: 197.47 ms /cgroup/a/cpuacct.usage 199.46 ms /cgroup/a/cpuacct.usage 205.46 ms /cgroup/a/cpuacct.usage 198.46 ms /cgroup/a/cpuacct.usage 208.39 ms /cgroup/a/cpuacct.usage Since we are allowed to use "stale" quota our usage is effectively bounded by the rate of input into the global pool and performance is relatively stable. with quota generations [1s increments]: 119.58 ms /cgroup/a/cpuacct.usage 119.65 ms /cgroup/a/cpuacct.usage 119.64 ms /cgroup/a/cpuacct.usage 119.63 ms /cgroup/a/cpuacct.usage 119.60 ms /cgroup/a/cpuacct.usage The large deficit here is due to quota generations (/intentionally/) preventing us from now using previously stranded slack quota. The cost is that this quota becomes unavailable. with quota generations and quota return: 200.09 ms /cgroup/a/cpuacct.usage 200.09 ms /cgroup/a/cpuacct.usage 198.09 ms /cgroup/a/cpuacct.usage 200.09 ms /cgroup/a/cpuacct.usage 200.06 ms /cgroup/a/cpuacct.usage By returning unused quota we're able to both stably consume our desired quota and prevent unintentional overages due to the abuse of slack quota from previous quota periods (especially on a large machine). Signed-off-by: Paul Turner <pjt@google.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/20110721184758.306848658@google.com Signed-off-by: Ingo Molnar <mingo@elte.hu>
2011-07-21 16:43:41 +00:00
/* return excess runtime on last dequeue */
return_cfs_rq_runtime(cfs_rq);
update_cfs_group(se);
/*
* Now advance min_vruntime if @se was the entity holding it back,
* except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
* put back on, and if we advance min_vruntime, we'll be placed back
* further than we started -- ie. we'll be penalized.
*/
if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
update_min_vruntime(cfs_rq);
}
/*
* Preempt the current task with a newly woken task if needed:
*/
static void
check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
{
unsigned long ideal_runtime, delta_exec;
struct sched_entity *se;
s64 delta;
ideal_runtime = sched_slice(cfs_rq, curr);
delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
if (delta_exec > ideal_runtime) {
resched_curr(rq_of(cfs_rq));
/*
* The current task ran long enough, ensure it doesn't get
* re-elected due to buddy favours.
*/
clear_buddies(cfs_rq, curr);
sched: Strengthen buddies and mitigate buddy induced latencies This patch restores the effectiveness of LAST_BUDDY in preventing pgsql+oltp from collapsing due to wakeup preemption. It also switches LAST_BUDDY to exclusively do what it does best, namely mitigate the effects of aggressive wakeup preemption, which improves vmark throughput markedly, and restores mysql+oltp scalability. Since buddies are about scalability, enable them beginning at the point where we begin expanding sched_latency, namely sched_nr_latency. Previously, buddies were cleared aggressively, which seriously reduced their effectiveness. Not clearing aggressively however, produces a small drop in mysql+oltp throughput immediately after peak, indicating that LAST_BUDDY is actually doing some harm. This is right at the point where X on the desktop in competition with another load wants low latency service. Ergo, do not enable until we need to scale. To mitigate latency induced by buddies, or by a task just missing wakeup preemption, check latency at tick time. Last hunk prevents buddies from stymieing BALANCE_NEWIDLE via CACHE_HOT_BUDDY. Supporting performance tests: tip = v2.6.32-rc5-1497-ga525b32 tipx = NO_GENTLE_FAIR_SLEEPERS NEXT_BUDDY granularity knobs = 31 knobs + 31 buddies tip+x = NO_GENTLE_FAIR_SLEEPERS granularity knobs = 31 knobs (Three run averages except where noted.) vmark: ------ tip 108466 messages per second tip+ 125307 messages per second tip+x 125335 messages per second tipx 117781 messages per second 2.6.31.3 122729 messages per second mysql+oltp: ----------- clients 1 2 4 8 16 32 64 128 256 .......................................................................................... tip 9949.89 18690.20 34801.24 34460.04 32682.88 30765.97 28305.27 25059.64 19548.08 tip+ 10013.90 18526.84 34900.38 34420.14 33069.83 32083.40 30578.30 28010.71 25605.47 tipx 9698.71 18002.70 34477.56 33420.01 32634.30 31657.27 29932.67 26827.52 21487.18 2.6.31.3 8243.11 18784.20 34404.83 33148.38 31900.32 31161.90 29663.81 25995.94 18058.86 pgsql+oltp: ----------- clients 1 2 4 8 16 32 64 128 256 .......................................................................................... tip 13686.37 26609.25 51934.28 51347.81 49479.51 45312.65 36691.91 26851.57 24145.35 tip+ (1x) 13907.85 27135.87 52951.98 52514.04 51742.52 50705.43 49947.97 48374.19 46227.94 tip+x 13906.78 27065.81 52951.19 52542.59 52176.11 51815.94 50838.90 49439.46 46891.00 tipx 13742.46 26769.81 52351.99 51891.73 51320.79 50938.98 50248.65 48908.70 46553.84 2.6.31.3 13815.35 26906.46 52683.34 52061.31 51937.10 51376.80 50474.28 49394.47 47003.25 Signed-off-by: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> LKML-Reference: <new-submission> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-10-23 21:09:22 +00:00
return;
}
/*
* Ensure that a task that missed wakeup preemption by a
* narrow margin doesn't have to wait for a full slice.
* This also mitigates buddy induced latencies under load.
*/
if (delta_exec < sysctl_sched_min_granularity)
return;
se = __pick_first_entity(cfs_rq);
delta = curr->vruntime - se->vruntime;
sched: Strengthen buddies and mitigate buddy induced latencies This patch restores the effectiveness of LAST_BUDDY in preventing pgsql+oltp from collapsing due to wakeup preemption. It also switches LAST_BUDDY to exclusively do what it does best, namely mitigate the effects of aggressive wakeup preemption, which improves vmark throughput markedly, and restores mysql+oltp scalability. Since buddies are about scalability, enable them beginning at the point where we begin expanding sched_latency, namely sched_nr_latency. Previously, buddies were cleared aggressively, which seriously reduced their effectiveness. Not clearing aggressively however, produces a small drop in mysql+oltp throughput immediately after peak, indicating that LAST_BUDDY is actually doing some harm. This is right at the point where X on the desktop in competition with another load wants low latency service. Ergo, do not enable until we need to scale. To mitigate latency induced by buddies, or by a task just missing wakeup preemption, check latency at tick time. Last hunk prevents buddies from stymieing BALANCE_NEWIDLE via CACHE_HOT_BUDDY. Supporting performance tests: tip = v2.6.32-rc5-1497-ga525b32 tipx = NO_GENTLE_FAIR_SLEEPERS NEXT_BUDDY granularity knobs = 31 knobs + 31 buddies tip+x = NO_GENTLE_FAIR_SLEEPERS granularity knobs = 31 knobs (Three run averages except where noted.) vmark: ------ tip 108466 messages per second tip+ 125307 messages per second tip+x 125335 messages per second tipx 117781 messages per second 2.6.31.3 122729 messages per second mysql+oltp: ----------- clients 1 2 4 8 16 32 64 128 256 .......................................................................................... tip 9949.89 18690.20 34801.24 34460.04 32682.88 30765.97 28305.27 25059.64 19548.08 tip+ 10013.90 18526.84 34900.38 34420.14 33069.83 32083.40 30578.30 28010.71 25605.47 tipx 9698.71 18002.70 34477.56 33420.01 32634.30 31657.27 29932.67 26827.52 21487.18 2.6.31.3 8243.11 18784.20 34404.83 33148.38 31900.32 31161.90 29663.81 25995.94 18058.86 pgsql+oltp: ----------- clients 1 2 4 8 16 32 64 128 256 .......................................................................................... tip 13686.37 26609.25 51934.28 51347.81 49479.51 45312.65 36691.91 26851.57 24145.35 tip+ (1x) 13907.85 27135.87 52951.98 52514.04 51742.52 50705.43 49947.97 48374.19 46227.94 tip+x 13906.78 27065.81 52951.19 52542.59 52176.11 51815.94 50838.90 49439.46 46891.00 tipx 13742.46 26769.81 52351.99 51891.73 51320.79 50938.98 50248.65 48908.70 46553.84 2.6.31.3 13815.35 26906.46 52683.34 52061.31 51937.10 51376.80 50474.28 49394.47 47003.25 Signed-off-by: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> LKML-Reference: <new-submission> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-10-23 21:09:22 +00:00
if (delta < 0)
return;
if (delta > ideal_runtime)
resched_curr(rq_of(cfs_rq));
}
static void
set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
{
/* 'current' is not kept within the tree. */
if (se->on_rq) {
/*
* Any task has to be enqueued before it get to execute on
* a CPU. So account for the time it spent waiting on the
* runqueue.
*/
update_stats_wait_end(cfs_rq, se);
__dequeue_entity(cfs_rq, se);
update_load_avg(cfs_rq, se, UPDATE_TG);
}
update_stats_curr_start(cfs_rq, se);
cfs_rq->curr = se;
/*
* Track our maximum slice length, if the CPU's load is at
* least twice that of our own weight (i.e. dont track it
* when there are only lesser-weight tasks around):
*/
if (schedstat_enabled() &&
rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) {
schedstat_set(se->statistics.slice_max,
max((u64)schedstat_val(se->statistics.slice_max),
se->sum_exec_runtime - se->prev_sum_exec_runtime));
}
se->prev_sum_exec_runtime = se->sum_exec_runtime;
}
static int
wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
/*
* Pick the next process, keeping these things in mind, in this order:
* 1) keep things fair between processes/task groups
* 2) pick the "next" process, since someone really wants that to run
* 3) pick the "last" process, for cache locality
* 4) do not run the "skip" process, if something else is available
*/
static struct sched_entity *
pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
{
struct sched_entity *left = __pick_first_entity(cfs_rq);
struct sched_entity *se;
/*
* If curr is set we have to see if its left of the leftmost entity
* still in the tree, provided there was anything in the tree at all.
*/
if (!left || (curr && entity_before(curr, left)))
left = curr;
se = left; /* ideally we run the leftmost entity */
/*
* Avoid running the skip buddy, if running something else can
* be done without getting too unfair.
*/
if (cfs_rq->skip == se) {
struct sched_entity *second;
if (se == curr) {
second = __pick_first_entity(cfs_rq);
} else {
second = __pick_next_entity(se);
if (!second || (curr && entity_before(curr, second)))
second = curr;
}
if (second && wakeup_preempt_entity(second, left) < 1)
se = second;
}
sched: Strengthen buddies and mitigate buddy induced latencies This patch restores the effectiveness of LAST_BUDDY in preventing pgsql+oltp from collapsing due to wakeup preemption. It also switches LAST_BUDDY to exclusively do what it does best, namely mitigate the effects of aggressive wakeup preemption, which improves vmark throughput markedly, and restores mysql+oltp scalability. Since buddies are about scalability, enable them beginning at the point where we begin expanding sched_latency, namely sched_nr_latency. Previously, buddies were cleared aggressively, which seriously reduced their effectiveness. Not clearing aggressively however, produces a small drop in mysql+oltp throughput immediately after peak, indicating that LAST_BUDDY is actually doing some harm. This is right at the point where X on the desktop in competition with another load wants low latency service. Ergo, do not enable until we need to scale. To mitigate latency induced by buddies, or by a task just missing wakeup preemption, check latency at tick time. Last hunk prevents buddies from stymieing BALANCE_NEWIDLE via CACHE_HOT_BUDDY. Supporting performance tests: tip = v2.6.32-rc5-1497-ga525b32 tipx = NO_GENTLE_FAIR_SLEEPERS NEXT_BUDDY granularity knobs = 31 knobs + 31 buddies tip+x = NO_GENTLE_FAIR_SLEEPERS granularity knobs = 31 knobs (Three run averages except where noted.) vmark: ------ tip 108466 messages per second tip+ 125307 messages per second tip+x 125335 messages per second tipx 117781 messages per second 2.6.31.3 122729 messages per second mysql+oltp: ----------- clients 1 2 4 8 16 32 64 128 256 .......................................................................................... tip 9949.89 18690.20 34801.24 34460.04 32682.88 30765.97 28305.27 25059.64 19548.08 tip+ 10013.90 18526.84 34900.38 34420.14 33069.83 32083.40 30578.30 28010.71 25605.47 tipx 9698.71 18002.70 34477.56 33420.01 32634.30 31657.27 29932.67 26827.52 21487.18 2.6.31.3 8243.11 18784.20 34404.83 33148.38 31900.32 31161.90 29663.81 25995.94 18058.86 pgsql+oltp: ----------- clients 1 2 4 8 16 32 64 128 256 .......................................................................................... tip 13686.37 26609.25 51934.28 51347.81 49479.51 45312.65 36691.91 26851.57 24145.35 tip+ (1x) 13907.85 27135.87 52951.98 52514.04 51742.52 50705.43 49947.97 48374.19 46227.94 tip+x 13906.78 27065.81 52951.19 52542.59 52176.11 51815.94 50838.90 49439.46 46891.00 tipx 13742.46 26769.81 52351.99 51891.73 51320.79 50938.98 50248.65 48908.70 46553.84 2.6.31.3 13815.35 26906.46 52683.34 52061.31 51937.10 51376.80 50474.28 49394.47 47003.25 Signed-off-by: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> LKML-Reference: <new-submission> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-10-23 21:09:22 +00:00
/*
* Prefer last buddy, try to return the CPU to a preempted task.
*/
if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
se = cfs_rq->last;
/*
* Someone really wants this to run. If it's not unfair, run it.
*/
if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
se = cfs_rq->next;
sched: Strengthen buddies and mitigate buddy induced latencies This patch restores the effectiveness of LAST_BUDDY in preventing pgsql+oltp from collapsing due to wakeup preemption. It also switches LAST_BUDDY to exclusively do what it does best, namely mitigate the effects of aggressive wakeup preemption, which improves vmark throughput markedly, and restores mysql+oltp scalability. Since buddies are about scalability, enable them beginning at the point where we begin expanding sched_latency, namely sched_nr_latency. Previously, buddies were cleared aggressively, which seriously reduced their effectiveness. Not clearing aggressively however, produces a small drop in mysql+oltp throughput immediately after peak, indicating that LAST_BUDDY is actually doing some harm. This is right at the point where X on the desktop in competition with another load wants low latency service. Ergo, do not enable until we need to scale. To mitigate latency induced by buddies, or by a task just missing wakeup preemption, check latency at tick time. Last hunk prevents buddies from stymieing BALANCE_NEWIDLE via CACHE_HOT_BUDDY. Supporting performance tests: tip = v2.6.32-rc5-1497-ga525b32 tipx = NO_GENTLE_FAIR_SLEEPERS NEXT_BUDDY granularity knobs = 31 knobs + 31 buddies tip+x = NO_GENTLE_FAIR_SLEEPERS granularity knobs = 31 knobs (Three run averages except where noted.) vmark: ------ tip 108466 messages per second tip+ 125307 messages per second tip+x 125335 messages per second tipx 117781 messages per second 2.6.31.3 122729 messages per second mysql+oltp: ----------- clients 1 2 4 8 16 32 64 128 256 .......................................................................................... tip 9949.89 18690.20 34801.24 34460.04 32682.88 30765.97 28305.27 25059.64 19548.08 tip+ 10013.90 18526.84 34900.38 34420.14 33069.83 32083.40 30578.30 28010.71 25605.47 tipx 9698.71 18002.70 34477.56 33420.01 32634.30 31657.27 29932.67 26827.52 21487.18 2.6.31.3 8243.11 18784.20 34404.83 33148.38 31900.32 31161.90 29663.81 25995.94 18058.86 pgsql+oltp: ----------- clients 1 2 4 8 16 32 64 128 256 .......................................................................................... tip 13686.37 26609.25 51934.28 51347.81 49479.51 45312.65 36691.91 26851.57 24145.35 tip+ (1x) 13907.85 27135.87 52951.98 52514.04 51742.52 50705.43 49947.97 48374.19 46227.94 tip+x 13906.78 27065.81 52951.19 52542.59 52176.11 51815.94 50838.90 49439.46 46891.00 tipx 13742.46 26769.81 52351.99 51891.73 51320.79 50938.98 50248.65 48908.70 46553.84 2.6.31.3 13815.35 26906.46 52683.34 52061.31 51937.10 51376.80 50474.28 49394.47 47003.25 Signed-off-by: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> LKML-Reference: <new-submission> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-10-23 21:09:22 +00:00
clear_buddies(cfs_rq, se);
return se;
}
static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
{
/*
* If still on the runqueue then deactivate_task()
* was not called and update_curr() has to be done:
*/
if (prev->on_rq)
update_curr(cfs_rq);
/* throttle cfs_rqs exceeding runtime */
check_cfs_rq_runtime(cfs_rq);
check_spread(cfs_rq, prev);
sched/debug: Make schedstats a runtime tunable that is disabled by default schedstats is very useful during debugging and performance tuning but it incurs overhead to calculate the stats. As such, even though it can be disabled at build time, it is often enabled as the information is useful. This patch adds a kernel command-line and sysctl tunable to enable or disable schedstats on demand (when it's built in). It is disabled by default as someone who knows they need it can also learn to enable it when necessary. The benefits are dependent on how scheduler-intensive the workload is. If it is then the patch reduces the number of cycles spent calculating the stats with a small benefit from reducing the cache footprint of the scheduler. These measurements were taken from a 48-core 2-socket machine with Xeon(R) E5-2670 v3 cpus although they were also tested on a single socket machine 8-core machine with Intel i7-3770 processors. netperf-tcp 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v3r1 Hmean 64 560.45 ( 0.00%) 575.98 ( 2.77%) Hmean 128 766.66 ( 0.00%) 795.79 ( 3.80%) Hmean 256 950.51 ( 0.00%) 981.50 ( 3.26%) Hmean 1024 1433.25 ( 0.00%) 1466.51 ( 2.32%) Hmean 2048 2810.54 ( 0.00%) 2879.75 ( 2.46%) Hmean 3312 4618.18 ( 0.00%) 4682.09 ( 1.38%) Hmean 4096 5306.42 ( 0.00%) 5346.39 ( 0.75%) Hmean 8192 10581.44 ( 0.00%) 10698.15 ( 1.10%) Hmean 16384 18857.70 ( 0.00%) 18937.61 ( 0.42%) Small gains here, UDP_STREAM showed nothing intresting and neither did the TCP_RR tests. The gains on the 8-core machine were very similar. tbench4 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v3r1 Hmean mb/sec-1 500.85 ( 0.00%) 522.43 ( 4.31%) Hmean mb/sec-2 984.66 ( 0.00%) 1018.19 ( 3.41%) Hmean mb/sec-4 1827.91 ( 0.00%) 1847.78 ( 1.09%) Hmean mb/sec-8 3561.36 ( 0.00%) 3611.28 ( 1.40%) Hmean mb/sec-16 5824.52 ( 0.00%) 5929.03 ( 1.79%) Hmean mb/sec-32 10943.10 ( 0.00%) 10802.83 ( -1.28%) Hmean mb/sec-64 15950.81 ( 0.00%) 16211.31 ( 1.63%) Hmean mb/sec-128 15302.17 ( 0.00%) 15445.11 ( 0.93%) Hmean mb/sec-256 14866.18 ( 0.00%) 15088.73 ( 1.50%) Hmean mb/sec-512 15223.31 ( 0.00%) 15373.69 ( 0.99%) Hmean mb/sec-1024 14574.25 ( 0.00%) 14598.02 ( 0.16%) Hmean mb/sec-2048 13569.02 ( 0.00%) 13733.86 ( 1.21%) Hmean mb/sec-3072 12865.98 ( 0.00%) 13209.23 ( 2.67%) Small gains of 2-4% at low thread counts and otherwise flat. The gains on the 8-core machine were slightly different tbench4 on 8-core i7-3770 single socket machine Hmean mb/sec-1 442.59 ( 0.00%) 448.73 ( 1.39%) Hmean mb/sec-2 796.68 ( 0.00%) 794.39 ( -0.29%) Hmean mb/sec-4 1322.52 ( 0.00%) 1343.66 ( 1.60%) Hmean mb/sec-8 2611.65 ( 0.00%) 2694.86 ( 3.19%) Hmean mb/sec-16 2537.07 ( 0.00%) 2609.34 ( 2.85%) Hmean mb/sec-32 2506.02 ( 0.00%) 2578.18 ( 2.88%) Hmean mb/sec-64 2511.06 ( 0.00%) 2569.16 ( 2.31%) Hmean mb/sec-128 2313.38 ( 0.00%) 2395.50 ( 3.55%) Hmean mb/sec-256 2110.04 ( 0.00%) 2177.45 ( 3.19%) Hmean mb/sec-512 2072.51 ( 0.00%) 2053.97 ( -0.89%) In constract, this shows a relatively steady 2-3% gain at higher thread counts. Due to the nature of the patch and the type of workload, it's not a surprise that the result will depend on the CPU used. hackbench-pipes 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v3r1 Amean 1 0.0637 ( 0.00%) 0.0660 ( -3.59%) Amean 4 0.1229 ( 0.00%) 0.1181 ( 3.84%) Amean 7 0.1921 ( 0.00%) 0.1911 ( 0.52%) Amean 12 0.3117 ( 0.00%) 0.2923 ( 6.23%) Amean 21 0.4050 ( 0.00%) 0.3899 ( 3.74%) Amean 30 0.4586 ( 0.00%) 0.4433 ( 3.33%) Amean 48 0.5910 ( 0.00%) 0.5694 ( 3.65%) Amean 79 0.8663 ( 0.00%) 0.8626 ( 0.43%) Amean 110 1.1543 ( 0.00%) 1.1517 ( 0.22%) Amean 141 1.4457 ( 0.00%) 1.4290 ( 1.16%) Amean 172 1.7090 ( 0.00%) 1.6924 ( 0.97%) Amean 192 1.9126 ( 0.00%) 1.9089 ( 0.19%) Some small gains and losses and while the variance data is not included, it's close to the noise. The UMA machine did not show anything particularly different pipetest 4.5.0-rc1 4.5.0-rc1 vanilla nostats-v2r2 Min Time 4.13 ( 0.00%) 3.99 ( 3.39%) 1st-qrtle Time 4.38 ( 0.00%) 4.27 ( 2.51%) 2nd-qrtle Time 4.46 ( 0.00%) 4.39 ( 1.57%) 3rd-qrtle Time 4.56 ( 0.00%) 4.51 ( 1.10%) Max-90% Time 4.67 ( 0.00%) 4.60 ( 1.50%) Max-93% Time 4.71 ( 0.00%) 4.65 ( 1.27%) Max-95% Time 4.74 ( 0.00%) 4.71 ( 0.63%) Max-99% Time 4.88 ( 0.00%) 4.79 ( 1.84%) Max Time 4.93 ( 0.00%) 4.83 ( 2.03%) Mean Time 4.48 ( 0.00%) 4.39 ( 1.91%) Best99%Mean Time 4.47 ( 0.00%) 4.39 ( 1.91%) Best95%Mean Time 4.46 ( 0.00%) 4.38 ( 1.93%) Best90%Mean Time 4.45 ( 0.00%) 4.36 ( 1.98%) Best50%Mean Time 4.36 ( 0.00%) 4.25 ( 2.49%) Best10%Mean Time 4.23 ( 0.00%) 4.10 ( 3.13%) Best5%Mean Time 4.19 ( 0.00%) 4.06 ( 3.20%) Best1%Mean Time 4.13 ( 0.00%) 4.00 ( 3.39%) Small improvement and similar gains were seen on the UMA machine. The gain is small but it stands to reason that doing less work in the scheduler is a good thing. The downside is that the lack of schedstats and tracepoints may be surprising to experts doing performance analysis until they find the existence of the schedstats= parameter or schedstats sysctl. It will be automatically activated for latencytop and sleep profiling to alleviate the problem. For tracepoints, there is a simple warning as it's not safe to activate schedstats in the context when it's known the tracepoint may be wanted but is unavailable. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Reviewed-by: Matt Fleming <matt@codeblueprint.co.uk> Reviewed-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <mgalbraith@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1454663316-22048-1-git-send-email-mgorman@techsingularity.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-02-05 09:08:36 +00:00
if (prev->on_rq) {
update_stats_wait_start(cfs_rq, prev);
/* Put 'current' back into the tree. */
__enqueue_entity(cfs_rq, prev);
/* in !on_rq case, update occurred at dequeue */
update_load_avg(cfs_rq, prev, 0);
}
cfs_rq->curr = NULL;
}
static void
entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
{
/*
* Update run-time statistics of the 'current'.
*/
update_curr(cfs_rq);
/*
* Ensure that runnable average is periodically updated.
*/
update_load_avg(cfs_rq, curr, UPDATE_TG);
update_cfs_group(curr);
#ifdef CONFIG_SCHED_HRTICK
/*
* queued ticks are scheduled to match the slice, so don't bother
* validating it and just reschedule.
*/
if (queued) {
resched_curr(rq_of(cfs_rq));
return;
}
/*
* don't let the period tick interfere with the hrtick preemption
*/
if (!sched_feat(DOUBLE_TICK) &&
hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
return;
#endif
if (cfs_rq->nr_running > 1)
check_preempt_tick(cfs_rq, curr);
}
/**************************************************
* CFS bandwidth control machinery
*/
#ifdef CONFIG_CFS_BANDWIDTH
#ifdef CONFIG_JUMP_LABEL
static keys: Introduce 'struct static_key', static_key_true()/false() and static_key_slow_[inc|dec]() So here's a boot tested patch on top of Jason's series that does all the cleanups I talked about and turns jump labels into a more intuitive to use facility. It should also address the various misconceptions and confusions that surround jump labels. Typical usage scenarios: #include <linux/static_key.h> struct static_key key = STATIC_KEY_INIT_TRUE; if (static_key_false(&key)) do unlikely code else do likely code Or: if (static_key_true(&key)) do likely code else do unlikely code The static key is modified via: static_key_slow_inc(&key); ... static_key_slow_dec(&key); The 'slow' prefix makes it abundantly clear that this is an expensive operation. I've updated all in-kernel code to use this everywhere. Note that I (intentionally) have not pushed through the rename blindly through to the lowest levels: the actual jump-label patching arch facility should be named like that, so we want to decouple jump labels from the static-key facility a bit. On non-jump-label enabled architectures static keys default to likely()/unlikely() branches. Signed-off-by: Ingo Molnar <mingo@elte.hu> Acked-by: Jason Baron <jbaron@redhat.com> Acked-by: Steven Rostedt <rostedt@goodmis.org> Cc: a.p.zijlstra@chello.nl Cc: mathieu.desnoyers@efficios.com Cc: davem@davemloft.net Cc: ddaney.cavm@gmail.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/20120222085809.GA26397@elte.hu Signed-off-by: Ingo Molnar <mingo@elte.hu>
2012-02-24 07:31:31 +00:00
static struct static_key __cfs_bandwidth_used;
static inline bool cfs_bandwidth_used(void)
{
static keys: Introduce 'struct static_key', static_key_true()/false() and static_key_slow_[inc|dec]() So here's a boot tested patch on top of Jason's series that does all the cleanups I talked about and turns jump labels into a more intuitive to use facility. It should also address the various misconceptions and confusions that surround jump labels. Typical usage scenarios: #include <linux/static_key.h> struct static_key key = STATIC_KEY_INIT_TRUE; if (static_key_false(&key)) do unlikely code else do likely code Or: if (static_key_true(&key)) do likely code else do unlikely code The static key is modified via: static_key_slow_inc(&key); ... static_key_slow_dec(&key); The 'slow' prefix makes it abundantly clear that this is an expensive operation. I've updated all in-kernel code to use this everywhere. Note that I (intentionally) have not pushed through the rename blindly through to the lowest levels: the actual jump-label patching arch facility should be named like that, so we want to decouple jump labels from the static-key facility a bit. On non-jump-label enabled architectures static keys default to likely()/unlikely() branches. Signed-off-by: Ingo Molnar <mingo@elte.hu> Acked-by: Jason Baron <jbaron@redhat.com> Acked-by: Steven Rostedt <rostedt@goodmis.org> Cc: a.p.zijlstra@chello.nl Cc: mathieu.desnoyers@efficios.com Cc: davem@davemloft.net Cc: ddaney.cavm@gmail.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/20120222085809.GA26397@elte.hu Signed-off-by: Ingo Molnar <mingo@elte.hu>
2012-02-24 07:31:31 +00:00
return static_key_false(&__cfs_bandwidth_used);
}
void cfs_bandwidth_usage_inc(void)
{
static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
}
void cfs_bandwidth_usage_dec(void)
{
static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
}
#else /* CONFIG_JUMP_LABEL */
static bool cfs_bandwidth_used(void)
{
return true;
}
void cfs_bandwidth_usage_inc(void) {}
void cfs_bandwidth_usage_dec(void) {}
#endif /* CONFIG_JUMP_LABEL */
/*
* default period for cfs group bandwidth.
* default: 0.1s, units: nanoseconds
*/
static inline u64 default_cfs_period(void)
{
return 100000000ULL;
}
static inline u64 sched_cfs_bandwidth_slice(void)
{
return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
}
/*
* Replenish runtime according to assigned quota. We use sched_clock_cpu
* directly instead of rq->clock to avoid adding additional synchronization
* around rq->lock.
*
* requires cfs_b->lock
*/
void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
{
if (cfs_b->quota != RUNTIME_INF)
cfs_b->runtime = cfs_b->quota;
}
static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
{
return &tg->cfs_bandwidth;
}
/* returns 0 on failure to allocate runtime */
static int __assign_cfs_rq_runtime(struct cfs_bandwidth *cfs_b,
struct cfs_rq *cfs_rq, u64 target_runtime)
{
u64 min_amount, amount = 0;
lockdep_assert_held(&cfs_b->lock);
/* note: this is a positive sum as runtime_remaining <= 0 */
min_amount = target_runtime - cfs_rq->runtime_remaining;
if (cfs_b->quota == RUNTIME_INF)
amount = min_amount;
else {
sched: Cleanup bandwidth timers Roman reported a 3 cpu lockup scenario involving __start_cfs_bandwidth(). The more I look at that code the more I'm convinced its crack, that entire __start_cfs_bandwidth() thing is brain melting, we don't need to cancel a timer before starting it, *hrtimer_start*() will happily remove the timer for you if its still enqueued. Removing that, removes a big part of the problem, no more ugly cancel loop to get stuck in. So now, if I understand things right, the entire reason you have this cfs_b->lock guarded ->timer_active nonsense is to make sure we don't accidentally lose the timer. It appears to me that it should be possible to guarantee that same by unconditionally (re)starting the timer when !queued. Because regardless what hrtimer::function will return, if we beat it to (re)enqueue the timer, it doesn't matter. Now, because hrtimers don't come with any serialization guarantees we must ensure both handler and (re)start loop serialize their access to the hrtimer to avoid both trying to forward the timer at the same time. Update the rt bandwidth timer to match. This effectively reverts: 09dc4ab03936 ("sched/fair: Fix tg_set_cfs_bandwidth() deadlock on rq->lock"). Reported-by: Roman Gushchin <klamm@yandex-team.ru> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Ben Segall <bsegall@google.com> Cc: Paul Turner <pjt@google.com> Link: http://lkml.kernel.org/r/20150415095011.804589208@infradead.org Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
2015-04-15 09:41:57 +00:00
start_cfs_bandwidth(cfs_b);
if (cfs_b->runtime > 0) {
amount = min(cfs_b->runtime, min_amount);
cfs_b->runtime -= amount;
cfs_b->idle = 0;
}
}
cfs_rq->runtime_remaining += amount;
return cfs_rq->runtime_remaining > 0;
}
/* returns 0 on failure to allocate runtime */
static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
{
struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
int ret;
raw_spin_lock(&cfs_b->lock);
ret = __assign_cfs_rq_runtime(cfs_b, cfs_rq, sched_cfs_bandwidth_slice());
raw_spin_unlock(&cfs_b->lock);
return ret;
}
sched/fair: Rework sched_fair time accounting Christian suffers from a bad BIOS that wrecks his i5's TSC sync. This results in him occasionally seeing time going backwards - which crashes the scheduler ... Most of our time accounting can actually handle that except the most common one; the tick time update of sched_fair. There is a further problem with that code; previously we assumed that because we get a tick every TICK_NSEC our time delta could never exceed 32bits and math was simpler. However, ever since Frederic managed to get NO_HZ_FULL merged; this is no longer the case since now a task can run for a long time indeed without getting a tick. It only takes about ~4.2 seconds to overflow our u32 in nanoseconds. This means we not only need to better deal with time going backwards; but also means we need to be able to deal with large deltas. This patch reworks the entire code and uses mul_u64_u32_shr() as proposed by Andy a long while ago. We express our virtual time scale factor in a u32 multiplier and shift right and the 32bit mul_u64_u32_shr() implementation reduces to a single 32x32->64 multiply if the time delta is still short (common case). For 64bit a 64x64->128 multiply can be used if ARCH_SUPPORTS_INT128. Reported-and-Tested-by: Christian Engelmayer <cengelma@gmx.at> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Cc: fweisbec@gmail.com Cc: Paul Turner <pjt@google.com> Cc: Stanislaw Gruszka <sgruszka@redhat.com> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Link: http://lkml.kernel.org/r/20131118172706.GI3866@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-11-18 17:27:06 +00:00
static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
{
/* dock delta_exec before expiring quota (as it could span periods) */
cfs_rq->runtime_remaining -= delta_exec;
if (likely(cfs_rq->runtime_remaining > 0))
return;
sched/fair: Don't assign runtime for throttled cfs_rq do_sched_cfs_period_timer() will refill cfs_b runtime and call distribute_cfs_runtime to unthrottle cfs_rq, sometimes cfs_b->runtime will allocate all quota to one cfs_rq incorrectly, then other cfs_rqs attached to this cfs_b can't get runtime and will be throttled. We find that one throttled cfs_rq has non-negative cfs_rq->runtime_remaining and cause an unexpetced cast from s64 to u64 in snippet: distribute_cfs_runtime() { runtime = -cfs_rq->runtime_remaining + 1; } The runtime here will change to a large number and consume all cfs_b->runtime in this cfs_b period. According to Ben Segall, the throttled cfs_rq can have account_cfs_rq_runtime called on it because it is throttled before idle_balance, and the idle_balance calls update_rq_clock to add time that is accounted to the task. This commit prevents cfs_rq to be assgined new runtime if it has been throttled until that distribute_cfs_runtime is called. Signed-off-by: Liangyan <liangyan.peng@linux.alibaba.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Valentin Schneider <valentin.schneider@arm.com> Reviewed-by: Ben Segall <bsegall@google.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: shanpeic@linux.alibaba.com Cc: stable@vger.kernel.org Cc: xlpang@linux.alibaba.com Fixes: d3d9dc330236 ("sched: Throttle entities exceeding their allowed bandwidth") Link: https://lkml.kernel.org/r/20190826121633.6538-1-liangyan.peng@linux.alibaba.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-08-26 12:16:33 +00:00
if (cfs_rq->throttled)
return;
/*
* if we're unable to extend our runtime we resched so that the active
* hierarchy can be throttled
*/
if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
resched_curr(rq_of(cfs_rq));
}
static __always_inline
sched/fair: Rework sched_fair time accounting Christian suffers from a bad BIOS that wrecks his i5's TSC sync. This results in him occasionally seeing time going backwards - which crashes the scheduler ... Most of our time accounting can actually handle that except the most common one; the tick time update of sched_fair. There is a further problem with that code; previously we assumed that because we get a tick every TICK_NSEC our time delta could never exceed 32bits and math was simpler. However, ever since Frederic managed to get NO_HZ_FULL merged; this is no longer the case since now a task can run for a long time indeed without getting a tick. It only takes about ~4.2 seconds to overflow our u32 in nanoseconds. This means we not only need to better deal with time going backwards; but also means we need to be able to deal with large deltas. This patch reworks the entire code and uses mul_u64_u32_shr() as proposed by Andy a long while ago. We express our virtual time scale factor in a u32 multiplier and shift right and the 32bit mul_u64_u32_shr() implementation reduces to a single 32x32->64 multiply if the time delta is still short (common case). For 64bit a 64x64->128 multiply can be used if ARCH_SUPPORTS_INT128. Reported-and-Tested-by: Christian Engelmayer <cengelma@gmx.at> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Cc: fweisbec@gmail.com Cc: Paul Turner <pjt@google.com> Cc: Stanislaw Gruszka <sgruszka@redhat.com> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Link: http://lkml.kernel.org/r/20131118172706.GI3866@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-11-18 17:27:06 +00:00
void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
{
if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
return;
__account_cfs_rq_runtime(cfs_rq, delta_exec);
}
static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
{
return cfs_bandwidth_used() && cfs_rq->throttled;
}
/* check whether cfs_rq, or any parent, is throttled */
static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
{
return cfs_bandwidth_used() && cfs_rq->throttle_count;
}
/*
* Ensure that neither of the group entities corresponding to src_cpu or
* dest_cpu are members of a throttled hierarchy when performing group
* load-balance operations.
*/
static inline int throttled_lb_pair(struct task_group *tg,
int src_cpu, int dest_cpu)
{
struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
src_cfs_rq = tg->cfs_rq[src_cpu];
dest_cfs_rq = tg->cfs_rq[dest_cpu];
return throttled_hierarchy(src_cfs_rq) ||
throttled_hierarchy(dest_cfs_rq);
}
static int tg_unthrottle_up(struct task_group *tg, void *data)
{
struct rq *rq = data;
struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
cfs_rq->throttle_count--;
if (!cfs_rq->throttle_count) {
cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
cfs_rq->throttled_clock_task;
sched/fair: Optimize update_blocked_averages() Removing a cfs_rq from rq->leaf_cfs_rq_list can break the parent/child ordering of the list when it will be added back. In order to remove an empty and fully decayed cfs_rq, we must remove its children too, so they will be added back in the right order next time. With a normal decay of PELT, a parent will be empty and fully decayed if all children are empty and fully decayed too. In such a case, we just have to ensure that the whole branch will be added when a new task is enqueued. This is default behavior since : commit f6783319737f ("sched/fair: Fix insertion in rq->leaf_cfs_rq_list") In case of throttling, the PELT of throttled cfs_rq will not be updated whereas the parent will. This breaks the assumption made above unless we remove the children of a cfs_rq that is throttled. Then, they will be added back when unthrottled and a sched_entity will be enqueued. As throttled cfs_rq are now removed from the list, we can remove the associated test in update_blocked_averages(). Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: sargun@sargun.me Cc: tj@kernel.org Cc: xiexiuqi@huawei.com Cc: xiezhipeng1@huawei.com Link: https://lkml.kernel.org/r/1549469662-13614-2-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-02-06 16:14:21 +00:00
/* Add cfs_rq with already running entity in the list */
if (cfs_rq->nr_running >= 1)
list_add_leaf_cfs_rq(cfs_rq);
}
return 0;
}
static int tg_throttle_down(struct task_group *tg, void *data)
{
struct rq *rq = data;
struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
/* group is entering throttled state, stop time */
sched/fair: Optimize update_blocked_averages() Removing a cfs_rq from rq->leaf_cfs_rq_list can break the parent/child ordering of the list when it will be added back. In order to remove an empty and fully decayed cfs_rq, we must remove its children too, so they will be added back in the right order next time. With a normal decay of PELT, a parent will be empty and fully decayed if all children are empty and fully decayed too. In such a case, we just have to ensure that the whole branch will be added when a new task is enqueued. This is default behavior since : commit f6783319737f ("sched/fair: Fix insertion in rq->leaf_cfs_rq_list") In case of throttling, the PELT of throttled cfs_rq will not be updated whereas the parent will. This breaks the assumption made above unless we remove the children of a cfs_rq that is throttled. Then, they will be added back when unthrottled and a sched_entity will be enqueued. As throttled cfs_rq are now removed from the list, we can remove the associated test in update_blocked_averages(). Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: sargun@sargun.me Cc: tj@kernel.org Cc: xiexiuqi@huawei.com Cc: xiezhipeng1@huawei.com Link: https://lkml.kernel.org/r/1549469662-13614-2-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-02-06 16:14:21 +00:00
if (!cfs_rq->throttle_count) {
cfs_rq->throttled_clock_task = rq_clock_task(rq);
sched/fair: Optimize update_blocked_averages() Removing a cfs_rq from rq->leaf_cfs_rq_list can break the parent/child ordering of the list when it will be added back. In order to remove an empty and fully decayed cfs_rq, we must remove its children too, so they will be added back in the right order next time. With a normal decay of PELT, a parent will be empty and fully decayed if all children are empty and fully decayed too. In such a case, we just have to ensure that the whole branch will be added when a new task is enqueued. This is default behavior since : commit f6783319737f ("sched/fair: Fix insertion in rq->leaf_cfs_rq_list") In case of throttling, the PELT of throttled cfs_rq will not be updated whereas the parent will. This breaks the assumption made above unless we remove the children of a cfs_rq that is throttled. Then, they will be added back when unthrottled and a sched_entity will be enqueued. As throttled cfs_rq are now removed from the list, we can remove the associated test in update_blocked_averages(). Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: sargun@sargun.me Cc: tj@kernel.org Cc: xiexiuqi@huawei.com Cc: xiezhipeng1@huawei.com Link: https://lkml.kernel.org/r/1549469662-13614-2-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-02-06 16:14:21 +00:00
list_del_leaf_cfs_rq(cfs_rq);
}
cfs_rq->throttle_count++;
return 0;
}
static bool throttle_cfs_rq(struct cfs_rq *cfs_rq)
{
struct rq *rq = rq_of(cfs_rq);
struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
struct sched_entity *se;
long task_delta, idle_task_delta, dequeue = 1;
raw_spin_lock(&cfs_b->lock);
/* This will start the period timer if necessary */
if (__assign_cfs_rq_runtime(cfs_b, cfs_rq, 1)) {
/*
* We have raced with bandwidth becoming available, and if we
* actually throttled the timer might not unthrottle us for an
* entire period. We additionally needed to make sure that any
* subsequent check_cfs_rq_runtime calls agree not to throttle
* us, as we may commit to do cfs put_prev+pick_next, so we ask
* for 1ns of runtime rather than just check cfs_b.
*/
dequeue = 0;
} else {
list_add_tail_rcu(&cfs_rq->throttled_list,
&cfs_b->throttled_cfs_rq);
}
raw_spin_unlock(&cfs_b->lock);
if (!dequeue)
return false; /* Throttle no longer required. */
se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
/* freeze hierarchy runnable averages while throttled */
rcu_read_lock();
walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
rcu_read_unlock();
task_delta = cfs_rq->h_nr_running;
idle_task_delta = cfs_rq->idle_h_nr_running;
for_each_sched_entity(se) {
struct cfs_rq *qcfs_rq = cfs_rq_of(se);
/* throttled entity or throttle-on-deactivate */
if (!se->on_rq)
break;
if (dequeue) {
dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
} else {
update_load_avg(qcfs_rq, se, 0);
se_update_runnable(se);
}
qcfs_rq->h_nr_running -= task_delta;
qcfs_rq->idle_h_nr_running -= idle_task_delta;
if (qcfs_rq->load.weight)
dequeue = 0;
}
if (!se)
sub_nr_running(rq, task_delta);
/*
* Note: distribution will already see us throttled via the
* throttled-list. rq->lock protects completion.
*/
cfs_rq->throttled = 1;
cfs_rq->throttled_clock = rq_clock(rq);
return true;
}
void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
{
struct rq *rq = rq_of(cfs_rq);
struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
struct sched_entity *se;
int enqueue = 1;
long task_delta, idle_task_delta;
se = cfs_rq->tg->se[cpu_of(rq)];
cfs_rq->throttled = 0;
update_rq_clock(rq);
raw_spin_lock(&cfs_b->lock);
cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
list_del_rcu(&cfs_rq->throttled_list);
raw_spin_unlock(&cfs_b->lock);
/* update hierarchical throttle state */
walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
if (!cfs_rq->load.weight)
return;
task_delta = cfs_rq->h_nr_running;
idle_task_delta = cfs_rq->idle_h_nr_running;
for_each_sched_entity(se) {
if (se->on_rq)
enqueue = 0;
cfs_rq = cfs_rq_of(se);
if (enqueue) {
enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
} else {
update_load_avg(cfs_rq, se, 0);
se_update_runnable(se);
}
cfs_rq->h_nr_running += task_delta;
cfs_rq->idle_h_nr_running += idle_task_delta;
if (cfs_rq_throttled(cfs_rq))
break;
}
if (!se)
add_nr_running(rq, task_delta);
/*
* The cfs_rq_throttled() breaks in the above iteration can result in
* incomplete leaf list maintenance, resulting in triggering the
* assertion below.
*/
for_each_sched_entity(se) {
cfs_rq = cfs_rq_of(se);
list_add_leaf_cfs_rq(cfs_rq);
}
assert_list_leaf_cfs_rq(rq);
/* Determine whether we need to wake up potentially idle CPU: */
if (rq->curr == rq->idle && rq->cfs.nr_running)
resched_curr(rq);
}
sched/fair: Fix race between runtime distribution and assignment Currently, there is a potential race between distribute_cfs_runtime() and assign_cfs_rq_runtime(). Race happens when cfs_b->runtime is read, distributes without holding lock and finds out there is not enough runtime to charge against after distribution. Because assign_cfs_rq_runtime() might be called during distribution, and use cfs_b->runtime at the same time. Fibtest is the tool to test this race. Assume all gcfs_rq is throttled and cfs period timer runs, slow threads might run and sleep, returning unused cfs_rq runtime and keeping min_cfs_rq_runtime in their local pool. If all this happens sufficiently quickly, cfs_b->runtime will drop a lot. If runtime distributed is large too, over-use of runtime happens. A runtime over-using by about 70 percent of quota is seen when we test fibtest on a 96-core machine. We run fibtest with 1 fast thread and 95 slow threads in test group, configure 10ms quota for this group and see the CPU usage of fibtest is 17.0%, which is far more than the expected 10%. On a smaller machine with 32 cores, we also run fibtest with 96 threads. CPU usage is more than 12%, which is also more than expected 10%. This shows that on similar workloads, this race do affect CPU bandwidth control. Solve this by holding lock inside distribute_cfs_runtime(). Fixes: c06f04c70489 ("sched: Fix potential near-infinite distribute_cfs_runtime() loop") Reviewed-by: Ben Segall <bsegall@google.com> Signed-off-by: Huaixin Chang <changhuaixin@linux.alibaba.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Link: https://lore.kernel.org/lkml/20200325092602.22471-1-changhuaixin@linux.alibaba.com/
2020-03-27 03:26:25 +00:00
static void distribute_cfs_runtime(struct cfs_bandwidth *cfs_b)
{
struct cfs_rq *cfs_rq;
sched/fair: Fix race between runtime distribution and assignment Currently, there is a potential race between distribute_cfs_runtime() and assign_cfs_rq_runtime(). Race happens when cfs_b->runtime is read, distributes without holding lock and finds out there is not enough runtime to charge against after distribution. Because assign_cfs_rq_runtime() might be called during distribution, and use cfs_b->runtime at the same time. Fibtest is the tool to test this race. Assume all gcfs_rq is throttled and cfs period timer runs, slow threads might run and sleep, returning unused cfs_rq runtime and keeping min_cfs_rq_runtime in their local pool. If all this happens sufficiently quickly, cfs_b->runtime will drop a lot. If runtime distributed is large too, over-use of runtime happens. A runtime over-using by about 70 percent of quota is seen when we test fibtest on a 96-core machine. We run fibtest with 1 fast thread and 95 slow threads in test group, configure 10ms quota for this group and see the CPU usage of fibtest is 17.0%, which is far more than the expected 10%. On a smaller machine with 32 cores, we also run fibtest with 96 threads. CPU usage is more than 12%, which is also more than expected 10%. This shows that on similar workloads, this race do affect CPU bandwidth control. Solve this by holding lock inside distribute_cfs_runtime(). Fixes: c06f04c70489 ("sched: Fix potential near-infinite distribute_cfs_runtime() loop") Reviewed-by: Ben Segall <bsegall@google.com> Signed-off-by: Huaixin Chang <changhuaixin@linux.alibaba.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Link: https://lore.kernel.org/lkml/20200325092602.22471-1-changhuaixin@linux.alibaba.com/
2020-03-27 03:26:25 +00:00
u64 runtime, remaining = 1;
rcu_read_lock();
list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
throttled_list) {
struct rq *rq = rq_of(cfs_rq);
struct rq_flags rf;
rq_lock_irqsave(rq, &rf);
if (!cfs_rq_throttled(cfs_rq))
goto next;
sched/fair: Don't assign runtime for throttled cfs_rq do_sched_cfs_period_timer() will refill cfs_b runtime and call distribute_cfs_runtime to unthrottle cfs_rq, sometimes cfs_b->runtime will allocate all quota to one cfs_rq incorrectly, then other cfs_rqs attached to this cfs_b can't get runtime and will be throttled. We find that one throttled cfs_rq has non-negative cfs_rq->runtime_remaining and cause an unexpetced cast from s64 to u64 in snippet: distribute_cfs_runtime() { runtime = -cfs_rq->runtime_remaining + 1; } The runtime here will change to a large number and consume all cfs_b->runtime in this cfs_b period. According to Ben Segall, the throttled cfs_rq can have account_cfs_rq_runtime called on it because it is throttled before idle_balance, and the idle_balance calls update_rq_clock to add time that is accounted to the task. This commit prevents cfs_rq to be assgined new runtime if it has been throttled until that distribute_cfs_runtime is called. Signed-off-by: Liangyan <liangyan.peng@linux.alibaba.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Valentin Schneider <valentin.schneider@arm.com> Reviewed-by: Ben Segall <bsegall@google.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: shanpeic@linux.alibaba.com Cc: stable@vger.kernel.org Cc: xlpang@linux.alibaba.com Fixes: d3d9dc330236 ("sched: Throttle entities exceeding their allowed bandwidth") Link: https://lkml.kernel.org/r/20190826121633.6538-1-liangyan.peng@linux.alibaba.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-08-26 12:16:33 +00:00
/* By the above check, this should never be true */
SCHED_WARN_ON(cfs_rq->runtime_remaining > 0);
sched/fair: Fix race between runtime distribution and assignment Currently, there is a potential race between distribute_cfs_runtime() and assign_cfs_rq_runtime(). Race happens when cfs_b->runtime is read, distributes without holding lock and finds out there is not enough runtime to charge against after distribution. Because assign_cfs_rq_runtime() might be called during distribution, and use cfs_b->runtime at the same time. Fibtest is the tool to test this race. Assume all gcfs_rq is throttled and cfs period timer runs, slow threads might run and sleep, returning unused cfs_rq runtime and keeping min_cfs_rq_runtime in their local pool. If all this happens sufficiently quickly, cfs_b->runtime will drop a lot. If runtime distributed is large too, over-use of runtime happens. A runtime over-using by about 70 percent of quota is seen when we test fibtest on a 96-core machine. We run fibtest with 1 fast thread and 95 slow threads in test group, configure 10ms quota for this group and see the CPU usage of fibtest is 17.0%, which is far more than the expected 10%. On a smaller machine with 32 cores, we also run fibtest with 96 threads. CPU usage is more than 12%, which is also more than expected 10%. This shows that on similar workloads, this race do affect CPU bandwidth control. Solve this by holding lock inside distribute_cfs_runtime(). Fixes: c06f04c70489 ("sched: Fix potential near-infinite distribute_cfs_runtime() loop") Reviewed-by: Ben Segall <bsegall@google.com> Signed-off-by: Huaixin Chang <changhuaixin@linux.alibaba.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Link: https://lore.kernel.org/lkml/20200325092602.22471-1-changhuaixin@linux.alibaba.com/
2020-03-27 03:26:25 +00:00
raw_spin_lock(&cfs_b->lock);
runtime = -cfs_rq->runtime_remaining + 1;
sched/fair: Fix race between runtime distribution and assignment Currently, there is a potential race between distribute_cfs_runtime() and assign_cfs_rq_runtime(). Race happens when cfs_b->runtime is read, distributes without holding lock and finds out there is not enough runtime to charge against after distribution. Because assign_cfs_rq_runtime() might be called during distribution, and use cfs_b->runtime at the same time. Fibtest is the tool to test this race. Assume all gcfs_rq is throttled and cfs period timer runs, slow threads might run and sleep, returning unused cfs_rq runtime and keeping min_cfs_rq_runtime in their local pool. If all this happens sufficiently quickly, cfs_b->runtime will drop a lot. If runtime distributed is large too, over-use of runtime happens. A runtime over-using by about 70 percent of quota is seen when we test fibtest on a 96-core machine. We run fibtest with 1 fast thread and 95 slow threads in test group, configure 10ms quota for this group and see the CPU usage of fibtest is 17.0%, which is far more than the expected 10%. On a smaller machine with 32 cores, we also run fibtest with 96 threads. CPU usage is more than 12%, which is also more than expected 10%. This shows that on similar workloads, this race do affect CPU bandwidth control. Solve this by holding lock inside distribute_cfs_runtime(). Fixes: c06f04c70489 ("sched: Fix potential near-infinite distribute_cfs_runtime() loop") Reviewed-by: Ben Segall <bsegall@google.com> Signed-off-by: Huaixin Chang <changhuaixin@linux.alibaba.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Link: https://lore.kernel.org/lkml/20200325092602.22471-1-changhuaixin@linux.alibaba.com/
2020-03-27 03:26:25 +00:00
if (runtime > cfs_b->runtime)
runtime = cfs_b->runtime;
cfs_b->runtime -= runtime;
remaining = cfs_b->runtime;
raw_spin_unlock(&cfs_b->lock);
cfs_rq->runtime_remaining += runtime;
/* we check whether we're throttled above */
if (cfs_rq->runtime_remaining > 0)
unthrottle_cfs_rq(cfs_rq);
next:
rq_unlock_irqrestore(rq, &rf);
if (!remaining)
break;
}
rcu_read_unlock();
}
/*
* Responsible for refilling a task_group's bandwidth and unthrottling its
* cfs_rqs as appropriate. If there has been no activity within the last
* period the timer is deactivated until scheduling resumes; cfs_b->idle is
* used to track this state.
*/
static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags)
{
int throttled;
/* no need to continue the timer with no bandwidth constraint */
if (cfs_b->quota == RUNTIME_INF)
goto out_deactivate;
throttled = !list_empty(&cfs_b->throttled_cfs_rq);
cfs_b->nr_periods += overrun;
/*
* idle depends on !throttled (for the case of a large deficit), and if
* we're going inactive then everything else can be deferred
*/
if (cfs_b->idle && !throttled)
goto out_deactivate;
__refill_cfs_bandwidth_runtime(cfs_b);
if (!throttled) {
/* mark as potentially idle for the upcoming period */
cfs_b->idle = 1;
return 0;
}
/* account preceding periods in which throttling occurred */
cfs_b->nr_throttled += overrun;
/*
sched/fair: Fix race between runtime distribution and assignment Currently, there is a potential race between distribute_cfs_runtime() and assign_cfs_rq_runtime(). Race happens when cfs_b->runtime is read, distributes without holding lock and finds out there is not enough runtime to charge against after distribution. Because assign_cfs_rq_runtime() might be called during distribution, and use cfs_b->runtime at the same time. Fibtest is the tool to test this race. Assume all gcfs_rq is throttled and cfs period timer runs, slow threads might run and sleep, returning unused cfs_rq runtime and keeping min_cfs_rq_runtime in their local pool. If all this happens sufficiently quickly, cfs_b->runtime will drop a lot. If runtime distributed is large too, over-use of runtime happens. A runtime over-using by about 70 percent of quota is seen when we test fibtest on a 96-core machine. We run fibtest with 1 fast thread and 95 slow threads in test group, configure 10ms quota for this group and see the CPU usage of fibtest is 17.0%, which is far more than the expected 10%. On a smaller machine with 32 cores, we also run fibtest with 96 threads. CPU usage is more than 12%, which is also more than expected 10%. This shows that on similar workloads, this race do affect CPU bandwidth control. Solve this by holding lock inside distribute_cfs_runtime(). Fixes: c06f04c70489 ("sched: Fix potential near-infinite distribute_cfs_runtime() loop") Reviewed-by: Ben Segall <bsegall@google.com> Signed-off-by: Huaixin Chang <changhuaixin@linux.alibaba.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Link: https://lore.kernel.org/lkml/20200325092602.22471-1-changhuaixin@linux.alibaba.com/
2020-03-27 03:26:25 +00:00
* This check is repeated as we release cfs_b->lock while we unthrottle.
*/
sched/fair: Fix throttle_list starvation with low CFS quota With a very low cpu.cfs_quota_us setting, such as the minimum of 1000, distribute_cfs_runtime may not empty the throttled_list before it runs out of runtime to distribute. In that case, due to the change from c06f04c7048 to put throttled entries at the head of the list, later entries on the list will starve. Essentially, the same X processes will get pulled off the list, given CPU time and then, when expired, get put back on the head of the list where distribute_cfs_runtime will give runtime to the same set of processes leaving the rest. Fix the issue by setting a bit in struct cfs_bandwidth when distribute_cfs_runtime is running, so that the code in throttle_cfs_rq can decide to put the throttled entry on the tail or the head of the list. The bit is set/cleared by the callers of distribute_cfs_runtime while they hold cfs_bandwidth->lock. This is easy to reproduce with a handful of CPU consumers. I use 'crash' on the live system. In some cases you can simply look at the throttled list and see the later entries are not changing: crash> list cfs_rq.throttled_list -H 0xffff90b54f6ade40 -s cfs_rq.runtime_remaining | paste - - | awk '{print $1" "$4}' | pr -t -n3 1 ffff90b56cb2d200 -976050 2 ffff90b56cb2cc00 -484925 3 ffff90b56cb2bc00 -658814 4 ffff90b56cb2ba00 -275365 5 ffff90b166a45600 -135138 6 ffff90b56cb2da00 -282505 7 ffff90b56cb2e000 -148065 8 ffff90b56cb2fa00 -872591 9 ffff90b56cb2c000 -84687 10 ffff90b56cb2f000 -87237 11 ffff90b166a40a00 -164582 crash> list cfs_rq.throttled_list -H 0xffff90b54f6ade40 -s cfs_rq.runtime_remaining | paste - - | awk '{print $1" "$4}' | pr -t -n3 1 ffff90b56cb2d200 -994147 2 ffff90b56cb2cc00 -306051 3 ffff90b56cb2bc00 -961321 4 ffff90b56cb2ba00 -24490 5 ffff90b166a45600 -135138 6 ffff90b56cb2da00 -282505 7 ffff90b56cb2e000 -148065 8 ffff90b56cb2fa00 -872591 9 ffff90b56cb2c000 -84687 10 ffff90b56cb2f000 -87237 11 ffff90b166a40a00 -164582 Sometimes it is easier to see by finding a process getting starved and looking at the sched_info: crash> task ffff8eb765994500 sched_info PID: 7800 TASK: ffff8eb765994500 CPU: 16 COMMAND: "cputest" sched_info = { pcount = 8, run_delay = 697094208, last_arrival = 240260125039, last_queued = 240260327513 }, crash> task ffff8eb765994500 sched_info PID: 7800 TASK: ffff8eb765994500 CPU: 16 COMMAND: "cputest" sched_info = { pcount = 8, run_delay = 697094208, last_arrival = 240260125039, last_queued = 240260327513 }, Signed-off-by: Phil Auld <pauld@redhat.com> Reviewed-by: Ben Segall <bsegall@google.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: stable@vger.kernel.org Fixes: c06f04c70489 ("sched: Fix potential near-infinite distribute_cfs_runtime() loop") Link: http://lkml.kernel.org/r/20181008143639.GA4019@pauld.bos.csb Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-10-08 14:36:40 +00:00
while (throttled && cfs_b->runtime > 0 && !cfs_b->distribute_running) {
cfs_b->distribute_running = 1;
raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
/* we can't nest cfs_b->lock while distributing bandwidth */
sched/fair: Fix race between runtime distribution and assignment Currently, there is a potential race between distribute_cfs_runtime() and assign_cfs_rq_runtime(). Race happens when cfs_b->runtime is read, distributes without holding lock and finds out there is not enough runtime to charge against after distribution. Because assign_cfs_rq_runtime() might be called during distribution, and use cfs_b->runtime at the same time. Fibtest is the tool to test this race. Assume all gcfs_rq is throttled and cfs period timer runs, slow threads might run and sleep, returning unused cfs_rq runtime and keeping min_cfs_rq_runtime in their local pool. If all this happens sufficiently quickly, cfs_b->runtime will drop a lot. If runtime distributed is large too, over-use of runtime happens. A runtime over-using by about 70 percent of quota is seen when we test fibtest on a 96-core machine. We run fibtest with 1 fast thread and 95 slow threads in test group, configure 10ms quota for this group and see the CPU usage of fibtest is 17.0%, which is far more than the expected 10%. On a smaller machine with 32 cores, we also run fibtest with 96 threads. CPU usage is more than 12%, which is also more than expected 10%. This shows that on similar workloads, this race do affect CPU bandwidth control. Solve this by holding lock inside distribute_cfs_runtime(). Fixes: c06f04c70489 ("sched: Fix potential near-infinite distribute_cfs_runtime() loop") Reviewed-by: Ben Segall <bsegall@google.com> Signed-off-by: Huaixin Chang <changhuaixin@linux.alibaba.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Link: https://lore.kernel.org/lkml/20200325092602.22471-1-changhuaixin@linux.alibaba.com/
2020-03-27 03:26:25 +00:00
distribute_cfs_runtime(cfs_b);
raw_spin_lock_irqsave(&cfs_b->lock, flags);
sched/fair: Fix throttle_list starvation with low CFS quota With a very low cpu.cfs_quota_us setting, such as the minimum of 1000, distribute_cfs_runtime may not empty the throttled_list before it runs out of runtime to distribute. In that case, due to the change from c06f04c7048 to put throttled entries at the head of the list, later entries on the list will starve. Essentially, the same X processes will get pulled off the list, given CPU time and then, when expired, get put back on the head of the list where distribute_cfs_runtime will give runtime to the same set of processes leaving the rest. Fix the issue by setting a bit in struct cfs_bandwidth when distribute_cfs_runtime is running, so that the code in throttle_cfs_rq can decide to put the throttled entry on the tail or the head of the list. The bit is set/cleared by the callers of distribute_cfs_runtime while they hold cfs_bandwidth->lock. This is easy to reproduce with a handful of CPU consumers. I use 'crash' on the live system. In some cases you can simply look at the throttled list and see the later entries are not changing: crash> list cfs_rq.throttled_list -H 0xffff90b54f6ade40 -s cfs_rq.runtime_remaining | paste - - | awk '{print $1" "$4}' | pr -t -n3 1 ffff90b56cb2d200 -976050 2 ffff90b56cb2cc00 -484925 3 ffff90b56cb2bc00 -658814 4 ffff90b56cb2ba00 -275365 5 ffff90b166a45600 -135138 6 ffff90b56cb2da00 -282505 7 ffff90b56cb2e000 -148065 8 ffff90b56cb2fa00 -872591 9 ffff90b56cb2c000 -84687 10 ffff90b56cb2f000 -87237 11 ffff90b166a40a00 -164582 crash> list cfs_rq.throttled_list -H 0xffff90b54f6ade40 -s cfs_rq.runtime_remaining | paste - - | awk '{print $1" "$4}' | pr -t -n3 1 ffff90b56cb2d200 -994147 2 ffff90b56cb2cc00 -306051 3 ffff90b56cb2bc00 -961321 4 ffff90b56cb2ba00 -24490 5 ffff90b166a45600 -135138 6 ffff90b56cb2da00 -282505 7 ffff90b56cb2e000 -148065 8 ffff90b56cb2fa00 -872591 9 ffff90b56cb2c000 -84687 10 ffff90b56cb2f000 -87237 11 ffff90b166a40a00 -164582 Sometimes it is easier to see by finding a process getting starved and looking at the sched_info: crash> task ffff8eb765994500 sched_info PID: 7800 TASK: ffff8eb765994500 CPU: 16 COMMAND: "cputest" sched_info = { pcount = 8, run_delay = 697094208, last_arrival = 240260125039, last_queued = 240260327513 }, crash> task ffff8eb765994500 sched_info PID: 7800 TASK: ffff8eb765994500 CPU: 16 COMMAND: "cputest" sched_info = { pcount = 8, run_delay = 697094208, last_arrival = 240260125039, last_queued = 240260327513 }, Signed-off-by: Phil Auld <pauld@redhat.com> Reviewed-by: Ben Segall <bsegall@google.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: stable@vger.kernel.org Fixes: c06f04c70489 ("sched: Fix potential near-infinite distribute_cfs_runtime() loop") Link: http://lkml.kernel.org/r/20181008143639.GA4019@pauld.bos.csb Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-10-08 14:36:40 +00:00
cfs_b->distribute_running = 0;
throttled = !list_empty(&cfs_b->throttled_cfs_rq);
}
/*
* While we are ensured activity in the period following an
* unthrottle, this also covers the case in which the new bandwidth is
* insufficient to cover the existing bandwidth deficit. (Forcing the
* timer to remain active while there are any throttled entities.)
*/
cfs_b->idle = 0;
return 0;
out_deactivate:
return 1;
}
sched: Return unused runtime on group dequeue When a local cfs_rq blocks we return the majority of its remaining quota to the global bandwidth pool for use by other runqueues. We do this only when the quota is current and there is more than min_cfs_rq_quota [1ms by default] of runtime remaining on the rq. In the case where there are throttled runqueues and we have sufficient bandwidth to meter out a slice, a second timer is kicked off to handle this delivery, unthrottling where appropriate. Using a 'worst case' antagonist which executes on each cpu for 1ms before moving onto the next on a fairly large machine: no quota generations: 197.47 ms /cgroup/a/cpuacct.usage 199.46 ms /cgroup/a/cpuacct.usage 205.46 ms /cgroup/a/cpuacct.usage 198.46 ms /cgroup/a/cpuacct.usage 208.39 ms /cgroup/a/cpuacct.usage Since we are allowed to use "stale" quota our usage is effectively bounded by the rate of input into the global pool and performance is relatively stable. with quota generations [1s increments]: 119.58 ms /cgroup/a/cpuacct.usage 119.65 ms /cgroup/a/cpuacct.usage 119.64 ms /cgroup/a/cpuacct.usage 119.63 ms /cgroup/a/cpuacct.usage 119.60 ms /cgroup/a/cpuacct.usage The large deficit here is due to quota generations (/intentionally/) preventing us from now using previously stranded slack quota. The cost is that this quota becomes unavailable. with quota generations and quota return: 200.09 ms /cgroup/a/cpuacct.usage 200.09 ms /cgroup/a/cpuacct.usage 198.09 ms /cgroup/a/cpuacct.usage 200.09 ms /cgroup/a/cpuacct.usage 200.06 ms /cgroup/a/cpuacct.usage By returning unused quota we're able to both stably consume our desired quota and prevent unintentional overages due to the abuse of slack quota from previous quota periods (especially on a large machine). Signed-off-by: Paul Turner <pjt@google.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/20110721184758.306848658@google.com Signed-off-by: Ingo Molnar <mingo@elte.hu>
2011-07-21 16:43:41 +00:00
/* a cfs_rq won't donate quota below this amount */
static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
/* minimum remaining period time to redistribute slack quota */
static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
/* how long we wait to gather additional slack before distributing */
static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
/*
* Are we near the end of the current quota period?
*
* Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
* hrtimer base being cleared by hrtimer_start. In the case of
* migrate_hrtimers, base is never cleared, so we are fine.
*/
sched: Return unused runtime on group dequeue When a local cfs_rq blocks we return the majority of its remaining quota to the global bandwidth pool for use by other runqueues. We do this only when the quota is current and there is more than min_cfs_rq_quota [1ms by default] of runtime remaining on the rq. In the case where there are throttled runqueues and we have sufficient bandwidth to meter out a slice, a second timer is kicked off to handle this delivery, unthrottling where appropriate. Using a 'worst case' antagonist which executes on each cpu for 1ms before moving onto the next on a fairly large machine: no quota generations: 197.47 ms /cgroup/a/cpuacct.usage 199.46 ms /cgroup/a/cpuacct.usage 205.46 ms /cgroup/a/cpuacct.usage 198.46 ms /cgroup/a/cpuacct.usage 208.39 ms /cgroup/a/cpuacct.usage Since we are allowed to use "stale" quota our usage is effectively bounded by the rate of input into the global pool and performance is relatively stable. with quota generations [1s increments]: 119.58 ms /cgroup/a/cpuacct.usage 119.65 ms /cgroup/a/cpuacct.usage 119.64 ms /cgroup/a/cpuacct.usage 119.63 ms /cgroup/a/cpuacct.usage 119.60 ms /cgroup/a/cpuacct.usage The large deficit here is due to quota generations (/intentionally/) preventing us from now using previously stranded slack quota. The cost is that this quota becomes unavailable. with quota generations and quota return: 200.09 ms /cgroup/a/cpuacct.usage 200.09 ms /cgroup/a/cpuacct.usage 198.09 ms /cgroup/a/cpuacct.usage 200.09 ms /cgroup/a/cpuacct.usage 200.06 ms /cgroup/a/cpuacct.usage By returning unused quota we're able to both stably consume our desired quota and prevent unintentional overages due to the abuse of slack quota from previous quota periods (especially on a large machine). Signed-off-by: Paul Turner <pjt@google.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/20110721184758.306848658@google.com Signed-off-by: Ingo Molnar <mingo@elte.hu>
2011-07-21 16:43:41 +00:00
static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
{
struct hrtimer *refresh_timer = &cfs_b->period_timer;
u64 remaining;
/* if the call-back is running a quota refresh is already occurring */
if (hrtimer_callback_running(refresh_timer))
return 1;
/* is a quota refresh about to occur? */
remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
if (remaining < min_expire)
return 1;
return 0;
}
static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
{
u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
/* if there's a quota refresh soon don't bother with slack */
if (runtime_refresh_within(cfs_b, min_left))
return;
/* don't push forwards an existing deferred unthrottle */
if (cfs_b->slack_started)
return;
cfs_b->slack_started = true;
sched,perf: Fix periodic timers In the below two commits (see Fixes) we have periodic timers that can stop themselves when they're no longer required, but need to be (re)-started when their idle condition changes. Further complications is that we want the timer handler to always do the forward such that it will always correctly deal with the overruns, and we do not want to race such that the handler has already decided to stop, but the (external) restart sees the timer still active and we end up with a 'lost' timer. The problem with the current code is that the re-start can come before the callback does the forward, at which point the forward from the callback will WARN about forwarding an enqueued timer. Now, conceptually its easy to detect if you're before or after the fwd by comparing the expiration time against the current time. Of course, that's expensive (and racy) because we don't have the current time. Alternatively one could cache this state inside the timer, but then everybody pays the overhead of maintaining this extra state, and that is undesired. The only other option that I could see is the external timer_active variable, which I tried to kill before. I would love a nicer interface for this seemingly simple 'problem' but alas. Fixes: 272325c4821f ("perf: Fix mux_interval hrtimer wreckage") Fixes: 77a4d1a1b9a1 ("sched: Cleanup bandwidth timers") Cc: pjt@google.com Cc: tglx@linutronix.de Cc: klamm@yandex-team.ru Cc: mingo@kernel.org Cc: bsegall@google.com Cc: hpa@zytor.com Cc: Sasha Levin <sasha.levin@oracle.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: http://lkml.kernel.org/r/20150514102311.GX21418@twins.programming.kicks-ass.net
2015-05-14 10:23:11 +00:00
hrtimer_start(&cfs_b->slack_timer,
ns_to_ktime(cfs_bandwidth_slack_period),
HRTIMER_MODE_REL);
sched: Return unused runtime on group dequeue When a local cfs_rq blocks we return the majority of its remaining quota to the global bandwidth pool for use by other runqueues. We do this only when the quota is current and there is more than min_cfs_rq_quota [1ms by default] of runtime remaining on the rq. In the case where there are throttled runqueues and we have sufficient bandwidth to meter out a slice, a second timer is kicked off to handle this delivery, unthrottling where appropriate. Using a 'worst case' antagonist which executes on each cpu for 1ms before moving onto the next on a fairly large machine: no quota generations: 197.47 ms /cgroup/a/cpuacct.usage 199.46 ms /cgroup/a/cpuacct.usage 205.46 ms /cgroup/a/cpuacct.usage 198.46 ms /cgroup/a/cpuacct.usage 208.39 ms /cgroup/a/cpuacct.usage Since we are allowed to use "stale" quota our usage is effectively bounded by the rate of input into the global pool and performance is relatively stable. with quota generations [1s increments]: 119.58 ms /cgroup/a/cpuacct.usage 119.65 ms /cgroup/a/cpuacct.usage 119.64 ms /cgroup/a/cpuacct.usage 119.63 ms /cgroup/a/cpuacct.usage 119.60 ms /cgroup/a/cpuacct.usage The large deficit here is due to quota generations (/intentionally/) preventing us from now using previously stranded slack quota. The cost is that this quota becomes unavailable. with quota generations and quota return: 200.09 ms /cgroup/a/cpuacct.usage 200.09 ms /cgroup/a/cpuacct.usage 198.09 ms /cgroup/a/cpuacct.usage 200.09 ms /cgroup/a/cpuacct.usage 200.06 ms /cgroup/a/cpuacct.usage By returning unused quota we're able to both stably consume our desired quota and prevent unintentional overages due to the abuse of slack quota from previous quota periods (especially on a large machine). Signed-off-by: Paul Turner <pjt@google.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/20110721184758.306848658@google.com Signed-off-by: Ingo Molnar <mingo@elte.hu>
2011-07-21 16:43:41 +00:00
}
/* we know any runtime found here is valid as update_curr() precedes return */
static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
{
struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
if (slack_runtime <= 0)
return;
raw_spin_lock(&cfs_b->lock);
sched/fair: Fix low cpu usage with high throttling by removing expiration of cpu-local slices It has been observed, that highly-threaded, non-cpu-bound applications running under cpu.cfs_quota_us constraints can hit a high percentage of periods throttled while simultaneously not consuming the allocated amount of quota. This use case is typical of user-interactive non-cpu bound applications, such as those running in kubernetes or mesos when run on multiple cpu cores. This has been root caused to cpu-local run queue being allocated per cpu bandwidth slices, and then not fully using that slice within the period. At which point the slice and quota expires. This expiration of unused slice results in applications not being able to utilize the quota for which they are allocated. The non-expiration of per-cpu slices was recently fixed by 'commit 512ac999d275 ("sched/fair: Fix bandwidth timer clock drift condition")'. Prior to that it appears that this had been broken since at least 'commit 51f2176d74ac ("sched/fair: Fix unlocked reads of some cfs_b->quota/period")' which was introduced in v3.16-rc1 in 2014. That added the following conditional which resulted in slices never being expired. if (cfs_rq->runtime_expires != cfs_b->runtime_expires) { /* extend local deadline, drift is bounded above by 2 ticks */ cfs_rq->runtime_expires += TICK_NSEC; Because this was broken for nearly 5 years, and has recently been fixed and is now being noticed by many users running kubernetes (https://github.com/kubernetes/kubernetes/issues/67577) it is my opinion that the mechanisms around expiring runtime should be removed altogether. This allows quota already allocated to per-cpu run-queues to live longer than the period boundary. This allows threads on runqueues that do not use much CPU to continue to use their remaining slice over a longer period of time than cpu.cfs_period_us. However, this helps prevent the above condition of hitting throttling while also not fully utilizing your cpu quota. This theoretically allows a machine to use slightly more than its allotted quota in some periods. This overflow would be bounded by the remaining quota left on each per-cpu runqueueu. This is typically no more than min_cfs_rq_runtime=1ms per cpu. For CPU bound tasks this will change nothing, as they should theoretically fully utilize all of their quota in each period. For user-interactive tasks as described above this provides a much better user/application experience as their cpu utilization will more closely match the amount they requested when they hit throttling. This means that cpu limits no longer strictly apply per period for non-cpu bound applications, but that they are still accurate over longer timeframes. This greatly improves performance of high-thread-count, non-cpu bound applications with low cfs_quota_us allocation on high-core-count machines. In the case of an artificial testcase (10ms/100ms of quota on 80 CPU machine), this commit resulted in almost 30x performance improvement, while still maintaining correct cpu quota restrictions. That testcase is available at https://github.com/indeedeng/fibtest. Fixes: 512ac999d275 ("sched/fair: Fix bandwidth timer clock drift condition") Signed-off-by: Dave Chiluk <chiluk+linux@indeed.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Phil Auld <pauld@redhat.com> Reviewed-by: Ben Segall <bsegall@google.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: John Hammond <jhammond@indeed.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Kyle Anderson <kwa@yelp.com> Cc: Gabriel Munos <gmunoz@netflix.com> Cc: Peter Oskolkov <posk@posk.io> Cc: Cong Wang <xiyou.wangcong@gmail.com> Cc: Brendan Gregg <bgregg@netflix.com> Link: https://lkml.kernel.org/r/1563900266-19734-2-git-send-email-chiluk+linux@indeed.com
2019-07-23 16:44:26 +00:00
if (cfs_b->quota != RUNTIME_INF) {
sched: Return unused runtime on group dequeue When a local cfs_rq blocks we return the majority of its remaining quota to the global bandwidth pool for use by other runqueues. We do this only when the quota is current and there is more than min_cfs_rq_quota [1ms by default] of runtime remaining on the rq. In the case where there are throttled runqueues and we have sufficient bandwidth to meter out a slice, a second timer is kicked off to handle this delivery, unthrottling where appropriate. Using a 'worst case' antagonist which executes on each cpu for 1ms before moving onto the next on a fairly large machine: no quota generations: 197.47 ms /cgroup/a/cpuacct.usage 199.46 ms /cgroup/a/cpuacct.usage 205.46 ms /cgroup/a/cpuacct.usage 198.46 ms /cgroup/a/cpuacct.usage 208.39 ms /cgroup/a/cpuacct.usage Since we are allowed to use "stale" quota our usage is effectively bounded by the rate of input into the global pool and performance is relatively stable. with quota generations [1s increments]: 119.58 ms /cgroup/a/cpuacct.usage 119.65 ms /cgroup/a/cpuacct.usage 119.64 ms /cgroup/a/cpuacct.usage 119.63 ms /cgroup/a/cpuacct.usage 119.60 ms /cgroup/a/cpuacct.usage The large deficit here is due to quota generations (/intentionally/) preventing us from now using previously stranded slack quota. The cost is that this quota becomes unavailable. with quota generations and quota return: 200.09 ms /cgroup/a/cpuacct.usage 200.09 ms /cgroup/a/cpuacct.usage 198.09 ms /cgroup/a/cpuacct.usage 200.09 ms /cgroup/a/cpuacct.usage 200.06 ms /cgroup/a/cpuacct.usage By returning unused quota we're able to both stably consume our desired quota and prevent unintentional overages due to the abuse of slack quota from previous quota periods (especially on a large machine). Signed-off-by: Paul Turner <pjt@google.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/20110721184758.306848658@google.com Signed-off-by: Ingo Molnar <mingo@elte.hu>
2011-07-21 16:43:41 +00:00
cfs_b->runtime += slack_runtime;
/* we are under rq->lock, defer unthrottling using a timer */
if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
!list_empty(&cfs_b->throttled_cfs_rq))
start_cfs_slack_bandwidth(cfs_b);
}
raw_spin_unlock(&cfs_b->lock);
/* even if it's not valid for return we don't want to try again */
cfs_rq->runtime_remaining -= slack_runtime;
}
static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
{
if (!cfs_bandwidth_used())
return;
if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
sched: Return unused runtime on group dequeue When a local cfs_rq blocks we return the majority of its remaining quota to the global bandwidth pool for use by other runqueues. We do this only when the quota is current and there is more than min_cfs_rq_quota [1ms by default] of runtime remaining on the rq. In the case where there are throttled runqueues and we have sufficient bandwidth to meter out a slice, a second timer is kicked off to handle this delivery, unthrottling where appropriate. Using a 'worst case' antagonist which executes on each cpu for 1ms before moving onto the next on a fairly large machine: no quota generations: 197.47 ms /cgroup/a/cpuacct.usage 199.46 ms /cgroup/a/cpuacct.usage 205.46 ms /cgroup/a/cpuacct.usage 198.46 ms /cgroup/a/cpuacct.usage 208.39 ms /cgroup/a/cpuacct.usage Since we are allowed to use "stale" quota our usage is effectively bounded by the rate of input into the global pool and performance is relatively stable. with quota generations [1s increments]: 119.58 ms /cgroup/a/cpuacct.usage 119.65 ms /cgroup/a/cpuacct.usage 119.64 ms /cgroup/a/cpuacct.usage 119.63 ms /cgroup/a/cpuacct.usage 119.60 ms /cgroup/a/cpuacct.usage The large deficit here is due to quota generations (/intentionally/) preventing us from now using previously stranded slack quota. The cost is that this quota becomes unavailable. with quota generations and quota return: 200.09 ms /cgroup/a/cpuacct.usage 200.09 ms /cgroup/a/cpuacct.usage 198.09 ms /cgroup/a/cpuacct.usage 200.09 ms /cgroup/a/cpuacct.usage 200.06 ms /cgroup/a/cpuacct.usage By returning unused quota we're able to both stably consume our desired quota and prevent unintentional overages due to the abuse of slack quota from previous quota periods (especially on a large machine). Signed-off-by: Paul Turner <pjt@google.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/20110721184758.306848658@google.com Signed-off-by: Ingo Molnar <mingo@elte.hu>
2011-07-21 16:43:41 +00:00
return;
__return_cfs_rq_runtime(cfs_rq);
}
/*
* This is done with a timer (instead of inline with bandwidth return) since
* it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
*/
static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
{
u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
unsigned long flags;
sched: Return unused runtime on group dequeue When a local cfs_rq blocks we return the majority of its remaining quota to the global bandwidth pool for use by other runqueues. We do this only when the quota is current and there is more than min_cfs_rq_quota [1ms by default] of runtime remaining on the rq. In the case where there are throttled runqueues and we have sufficient bandwidth to meter out a slice, a second timer is kicked off to handle this delivery, unthrottling where appropriate. Using a 'worst case' antagonist which executes on each cpu for 1ms before moving onto the next on a fairly large machine: no quota generations: 197.47 ms /cgroup/a/cpuacct.usage 199.46 ms /cgroup/a/cpuacct.usage 205.46 ms /cgroup/a/cpuacct.usage 198.46 ms /cgroup/a/cpuacct.usage 208.39 ms /cgroup/a/cpuacct.usage Since we are allowed to use "stale" quota our usage is effectively bounded by the rate of input into the global pool and performance is relatively stable. with quota generations [1s increments]: 119.58 ms /cgroup/a/cpuacct.usage 119.65 ms /cgroup/a/cpuacct.usage 119.64 ms /cgroup/a/cpuacct.usage 119.63 ms /cgroup/a/cpuacct.usage 119.60 ms /cgroup/a/cpuacct.usage The large deficit here is due to quota generations (/intentionally/) preventing us from now using previously stranded slack quota. The cost is that this quota becomes unavailable. with quota generations and quota return: 200.09 ms /cgroup/a/cpuacct.usage 200.09 ms /cgroup/a/cpuacct.usage 198.09 ms /cgroup/a/cpuacct.usage 200.09 ms /cgroup/a/cpuacct.usage 200.06 ms /cgroup/a/cpuacct.usage By returning unused quota we're able to both stably consume our desired quota and prevent unintentional overages due to the abuse of slack quota from previous quota periods (especially on a large machine). Signed-off-by: Paul Turner <pjt@google.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/20110721184758.306848658@google.com Signed-off-by: Ingo Molnar <mingo@elte.hu>
2011-07-21 16:43:41 +00:00
/* confirm we're still not at a refresh boundary */
raw_spin_lock_irqsave(&cfs_b->lock, flags);
cfs_b->slack_started = false;
sched/fair: Fix throttle_list starvation with low CFS quota With a very low cpu.cfs_quota_us setting, such as the minimum of 1000, distribute_cfs_runtime may not empty the throttled_list before it runs out of runtime to distribute. In that case, due to the change from c06f04c7048 to put throttled entries at the head of the list, later entries on the list will starve. Essentially, the same X processes will get pulled off the list, given CPU time and then, when expired, get put back on the head of the list where distribute_cfs_runtime will give runtime to the same set of processes leaving the rest. Fix the issue by setting a bit in struct cfs_bandwidth when distribute_cfs_runtime is running, so that the code in throttle_cfs_rq can decide to put the throttled entry on the tail or the head of the list. The bit is set/cleared by the callers of distribute_cfs_runtime while they hold cfs_bandwidth->lock. This is easy to reproduce with a handful of CPU consumers. I use 'crash' on the live system. In some cases you can simply look at the throttled list and see the later entries are not changing: crash> list cfs_rq.throttled_list -H 0xffff90b54f6ade40 -s cfs_rq.runtime_remaining | paste - - | awk '{print $1" "$4}' | pr -t -n3 1 ffff90b56cb2d200 -976050 2 ffff90b56cb2cc00 -484925 3 ffff90b56cb2bc00 -658814 4 ffff90b56cb2ba00 -275365 5 ffff90b166a45600 -135138 6 ffff90b56cb2da00 -282505 7 ffff90b56cb2e000 -148065 8 ffff90b56cb2fa00 -872591 9 ffff90b56cb2c000 -84687 10 ffff90b56cb2f000 -87237 11 ffff90b166a40a00 -164582 crash> list cfs_rq.throttled_list -H 0xffff90b54f6ade40 -s cfs_rq.runtime_remaining | paste - - | awk '{print $1" "$4}' | pr -t -n3 1 ffff90b56cb2d200 -994147 2 ffff90b56cb2cc00 -306051 3 ffff90b56cb2bc00 -961321 4 ffff90b56cb2ba00 -24490 5 ffff90b166a45600 -135138 6 ffff90b56cb2da00 -282505 7 ffff90b56cb2e000 -148065 8 ffff90b56cb2fa00 -872591 9 ffff90b56cb2c000 -84687 10 ffff90b56cb2f000 -87237 11 ffff90b166a40a00 -164582 Sometimes it is easier to see by finding a process getting starved and looking at the sched_info: crash> task ffff8eb765994500 sched_info PID: 7800 TASK: ffff8eb765994500 CPU: 16 COMMAND: "cputest" sched_info = { pcount = 8, run_delay = 697094208, last_arrival = 240260125039, last_queued = 240260327513 }, crash> task ffff8eb765994500 sched_info PID: 7800 TASK: ffff8eb765994500 CPU: 16 COMMAND: "cputest" sched_info = { pcount = 8, run_delay = 697094208, last_arrival = 240260125039, last_queued = 240260327513 }, Signed-off-by: Phil Auld <pauld@redhat.com> Reviewed-by: Ben Segall <bsegall@google.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: stable@vger.kernel.org Fixes: c06f04c70489 ("sched: Fix potential near-infinite distribute_cfs_runtime() loop") Link: http://lkml.kernel.org/r/20181008143639.GA4019@pauld.bos.csb Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-10-08 14:36:40 +00:00
if (cfs_b->distribute_running) {
raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
sched/fair: Fix throttle_list starvation with low CFS quota With a very low cpu.cfs_quota_us setting, such as the minimum of 1000, distribute_cfs_runtime may not empty the throttled_list before it runs out of runtime to distribute. In that case, due to the change from c06f04c7048 to put throttled entries at the head of the list, later entries on the list will starve. Essentially, the same X processes will get pulled off the list, given CPU time and then, when expired, get put back on the head of the list where distribute_cfs_runtime will give runtime to the same set of processes leaving the rest. Fix the issue by setting a bit in struct cfs_bandwidth when distribute_cfs_runtime is running, so that the code in throttle_cfs_rq can decide to put the throttled entry on the tail or the head of the list. The bit is set/cleared by the callers of distribute_cfs_runtime while they hold cfs_bandwidth->lock. This is easy to reproduce with a handful of CPU consumers. I use 'crash' on the live system. In some cases you can simply look at the throttled list and see the later entries are not changing: crash> list cfs_rq.throttled_list -H 0xffff90b54f6ade40 -s cfs_rq.runtime_remaining | paste - - | awk '{print $1" "$4}' | pr -t -n3 1 ffff90b56cb2d200 -976050 2 ffff90b56cb2cc00 -484925 3 ffff90b56cb2bc00 -658814 4 ffff90b56cb2ba00 -275365 5 ffff90b166a45600 -135138 6 ffff90b56cb2da00 -282505 7 ffff90b56cb2e000 -148065 8 ffff90b56cb2fa00 -872591 9 ffff90b56cb2c000 -84687 10 ffff90b56cb2f000 -87237 11 ffff90b166a40a00 -164582 crash> list cfs_rq.throttled_list -H 0xffff90b54f6ade40 -s cfs_rq.runtime_remaining | paste - - | awk '{print $1" "$4}' | pr -t -n3 1 ffff90b56cb2d200 -994147 2 ffff90b56cb2cc00 -306051 3 ffff90b56cb2bc00 -961321 4 ffff90b56cb2ba00 -24490 5 ffff90b166a45600 -135138 6 ffff90b56cb2da00 -282505 7 ffff90b56cb2e000 -148065 8 ffff90b56cb2fa00 -872591 9 ffff90b56cb2c000 -84687 10 ffff90b56cb2f000 -87237 11 ffff90b166a40a00 -164582 Sometimes it is easier to see by finding a process getting starved and looking at the sched_info: crash> task ffff8eb765994500 sched_info PID: 7800 TASK: ffff8eb765994500 CPU: 16 COMMAND: "cputest" sched_info = { pcount = 8, run_delay = 697094208, last_arrival = 240260125039, last_queued = 240260327513 }, crash> task ffff8eb765994500 sched_info PID: 7800 TASK: ffff8eb765994500 CPU: 16 COMMAND: "cputest" sched_info = { pcount = 8, run_delay = 697094208, last_arrival = 240260125039, last_queued = 240260327513 }, Signed-off-by: Phil Auld <pauld@redhat.com> Reviewed-by: Ben Segall <bsegall@google.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: stable@vger.kernel.org Fixes: c06f04c70489 ("sched: Fix potential near-infinite distribute_cfs_runtime() loop") Link: http://lkml.kernel.org/r/20181008143639.GA4019@pauld.bos.csb Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-10-08 14:36:40 +00:00
return;
}
if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
sched: Return unused runtime on group dequeue When a local cfs_rq blocks we return the majority of its remaining quota to the global bandwidth pool for use by other runqueues. We do this only when the quota is current and there is more than min_cfs_rq_quota [1ms by default] of runtime remaining on the rq. In the case where there are throttled runqueues and we have sufficient bandwidth to meter out a slice, a second timer is kicked off to handle this delivery, unthrottling where appropriate. Using a 'worst case' antagonist which executes on each cpu for 1ms before moving onto the next on a fairly large machine: no quota generations: 197.47 ms /cgroup/a/cpuacct.usage 199.46 ms /cgroup/a/cpuacct.usage 205.46 ms /cgroup/a/cpuacct.usage 198.46 ms /cgroup/a/cpuacct.usage 208.39 ms /cgroup/a/cpuacct.usage Since we are allowed to use "stale" quota our usage is effectively bounded by the rate of input into the global pool and performance is relatively stable. with quota generations [1s increments]: 119.58 ms /cgroup/a/cpuacct.usage 119.65 ms /cgroup/a/cpuacct.usage 119.64 ms /cgroup/a/cpuacct.usage 119.63 ms /cgroup/a/cpuacct.usage 119.60 ms /cgroup/a/cpuacct.usage The large deficit here is due to quota generations (/intentionally/) preventing us from now using previously stranded slack quota. The cost is that this quota becomes unavailable. with quota generations and quota return: 200.09 ms /cgroup/a/cpuacct.usage 200.09 ms /cgroup/a/cpuacct.usage 198.09 ms /cgroup/a/cpuacct.usage 200.09 ms /cgroup/a/cpuacct.usage 200.06 ms /cgroup/a/cpuacct.usage By returning unused quota we're able to both stably consume our desired quota and prevent unintentional overages due to the abuse of slack quota from previous quota periods (especially on a large machine). Signed-off-by: Paul Turner <pjt@google.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/20110721184758.306848658@google.com Signed-off-by: Ingo Molnar <mingo@elte.hu>
2011-07-21 16:43:41 +00:00
return;
}
sched: Return unused runtime on group dequeue When a local cfs_rq blocks we return the majority of its remaining quota to the global bandwidth pool for use by other runqueues. We do this only when the quota is current and there is more than min_cfs_rq_quota [1ms by default] of runtime remaining on the rq. In the case where there are throttled runqueues and we have sufficient bandwidth to meter out a slice, a second timer is kicked off to handle this delivery, unthrottling where appropriate. Using a 'worst case' antagonist which executes on each cpu for 1ms before moving onto the next on a fairly large machine: no quota generations: 197.47 ms /cgroup/a/cpuacct.usage 199.46 ms /cgroup/a/cpuacct.usage 205.46 ms /cgroup/a/cpuacct.usage 198.46 ms /cgroup/a/cpuacct.usage 208.39 ms /cgroup/a/cpuacct.usage Since we are allowed to use "stale" quota our usage is effectively bounded by the rate of input into the global pool and performance is relatively stable. with quota generations [1s increments]: 119.58 ms /cgroup/a/cpuacct.usage 119.65 ms /cgroup/a/cpuacct.usage 119.64 ms /cgroup/a/cpuacct.usage 119.63 ms /cgroup/a/cpuacct.usage 119.60 ms /cgroup/a/cpuacct.usage The large deficit here is due to quota generations (/intentionally/) preventing us from now using previously stranded slack quota. The cost is that this quota becomes unavailable. with quota generations and quota return: 200.09 ms /cgroup/a/cpuacct.usage 200.09 ms /cgroup/a/cpuacct.usage 198.09 ms /cgroup/a/cpuacct.usage 200.09 ms /cgroup/a/cpuacct.usage 200.06 ms /cgroup/a/cpuacct.usage By returning unused quota we're able to both stably consume our desired quota and prevent unintentional overages due to the abuse of slack quota from previous quota periods (especially on a large machine). Signed-off-by: Paul Turner <pjt@google.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/20110721184758.306848658@google.com Signed-off-by: Ingo Molnar <mingo@elte.hu>
2011-07-21 16:43:41 +00:00
if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
sched: Return unused runtime on group dequeue When a local cfs_rq blocks we return the majority of its remaining quota to the global bandwidth pool for use by other runqueues. We do this only when the quota is current and there is more than min_cfs_rq_quota [1ms by default] of runtime remaining on the rq. In the case where there are throttled runqueues and we have sufficient bandwidth to meter out a slice, a second timer is kicked off to handle this delivery, unthrottling where appropriate. Using a 'worst case' antagonist which executes on each cpu for 1ms before moving onto the next on a fairly large machine: no quota generations: 197.47 ms /cgroup/a/cpuacct.usage 199.46 ms /cgroup/a/cpuacct.usage 205.46 ms /cgroup/a/cpuacct.usage 198.46 ms /cgroup/a/cpuacct.usage 208.39 ms /cgroup/a/cpuacct.usage Since we are allowed to use "stale" quota our usage is effectively bounded by the rate of input into the global pool and performance is relatively stable. with quota generations [1s increments]: 119.58 ms /cgroup/a/cpuacct.usage 119.65 ms /cgroup/a/cpuacct.usage 119.64 ms /cgroup/a/cpuacct.usage 119.63 ms /cgroup/a/cpuacct.usage 119.60 ms /cgroup/a/cpuacct.usage The large deficit here is due to quota generations (/intentionally/) preventing us from now using previously stranded slack quota. The cost is that this quota becomes unavailable. with quota generations and quota return: 200.09 ms /cgroup/a/cpuacct.usage 200.09 ms /cgroup/a/cpuacct.usage 198.09 ms /cgroup/a/cpuacct.usage 200.09 ms /cgroup/a/cpuacct.usage 200.06 ms /cgroup/a/cpuacct.usage By returning unused quota we're able to both stably consume our desired quota and prevent unintentional overages due to the abuse of slack quota from previous quota periods (especially on a large machine). Signed-off-by: Paul Turner <pjt@google.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/20110721184758.306848658@google.com Signed-off-by: Ingo Molnar <mingo@elte.hu>
2011-07-21 16:43:41 +00:00
runtime = cfs_b->runtime;
sched/fair: Fix throttle_list starvation with low CFS quota With a very low cpu.cfs_quota_us setting, such as the minimum of 1000, distribute_cfs_runtime may not empty the throttled_list before it runs out of runtime to distribute. In that case, due to the change from c06f04c7048 to put throttled entries at the head of the list, later entries on the list will starve. Essentially, the same X processes will get pulled off the list, given CPU time and then, when expired, get put back on the head of the list where distribute_cfs_runtime will give runtime to the same set of processes leaving the rest. Fix the issue by setting a bit in struct cfs_bandwidth when distribute_cfs_runtime is running, so that the code in throttle_cfs_rq can decide to put the throttled entry on the tail or the head of the list. The bit is set/cleared by the callers of distribute_cfs_runtime while they hold cfs_bandwidth->lock. This is easy to reproduce with a handful of CPU consumers. I use 'crash' on the live system. In some cases you can simply look at the throttled list and see the later entries are not changing: crash> list cfs_rq.throttled_list -H 0xffff90b54f6ade40 -s cfs_rq.runtime_remaining | paste - - | awk '{print $1" "$4}' | pr -t -n3 1 ffff90b56cb2d200 -976050 2 ffff90b56cb2cc00 -484925 3 ffff90b56cb2bc00 -658814 4 ffff90b56cb2ba00 -275365 5 ffff90b166a45600 -135138 6 ffff90b56cb2da00 -282505 7 ffff90b56cb2e000 -148065 8 ffff90b56cb2fa00 -872591 9 ffff90b56cb2c000 -84687 10 ffff90b56cb2f000 -87237 11 ffff90b166a40a00 -164582 crash> list cfs_rq.throttled_list -H 0xffff90b54f6ade40 -s cfs_rq.runtime_remaining | paste - - | awk '{print $1" "$4}' | pr -t -n3 1 ffff90b56cb2d200 -994147 2 ffff90b56cb2cc00 -306051 3 ffff90b56cb2bc00 -961321 4 ffff90b56cb2ba00 -24490 5 ffff90b166a45600 -135138 6 ffff90b56cb2da00 -282505 7 ffff90b56cb2e000 -148065 8 ffff90b56cb2fa00 -872591 9 ffff90b56cb2c000 -84687 10 ffff90b56cb2f000 -87237 11 ffff90b166a40a00 -164582 Sometimes it is easier to see by finding a process getting starved and looking at the sched_info: crash> task ffff8eb765994500 sched_info PID: 7800 TASK: ffff8eb765994500 CPU: 16 COMMAND: "cputest" sched_info = { pcount = 8, run_delay = 697094208, last_arrival = 240260125039, last_queued = 240260327513 }, crash> task ffff8eb765994500 sched_info PID: 7800 TASK: ffff8eb765994500 CPU: 16 COMMAND: "cputest" sched_info = { pcount = 8, run_delay = 697094208, last_arrival = 240260125039, last_queued = 240260327513 }, Signed-off-by: Phil Auld <pauld@redhat.com> Reviewed-by: Ben Segall <bsegall@google.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: stable@vger.kernel.org Fixes: c06f04c70489 ("sched: Fix potential near-infinite distribute_cfs_runtime() loop") Link: http://lkml.kernel.org/r/20181008143639.GA4019@pauld.bos.csb Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-10-08 14:36:40 +00:00
if (runtime)
cfs_b->distribute_running = 1;
raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
sched: Return unused runtime on group dequeue When a local cfs_rq blocks we return the majority of its remaining quota to the global bandwidth pool for use by other runqueues. We do this only when the quota is current and there is more than min_cfs_rq_quota [1ms by default] of runtime remaining on the rq. In the case where there are throttled runqueues and we have sufficient bandwidth to meter out a slice, a second timer is kicked off to handle this delivery, unthrottling where appropriate. Using a 'worst case' antagonist which executes on each cpu for 1ms before moving onto the next on a fairly large machine: no quota generations: 197.47 ms /cgroup/a/cpuacct.usage 199.46 ms /cgroup/a/cpuacct.usage 205.46 ms /cgroup/a/cpuacct.usage 198.46 ms /cgroup/a/cpuacct.usage 208.39 ms /cgroup/a/cpuacct.usage Since we are allowed to use "stale" quota our usage is effectively bounded by the rate of input into the global pool and performance is relatively stable. with quota generations [1s increments]: 119.58 ms /cgroup/a/cpuacct.usage 119.65 ms /cgroup/a/cpuacct.usage 119.64 ms /cgroup/a/cpuacct.usage 119.63 ms /cgroup/a/cpuacct.usage 119.60 ms /cgroup/a/cpuacct.usage The large deficit here is due to quota generations (/intentionally/) preventing us from now using previously stranded slack quota. The cost is that this quota becomes unavailable. with quota generations and quota return: 200.09 ms /cgroup/a/cpuacct.usage 200.09 ms /cgroup/a/cpuacct.usage 198.09 ms /cgroup/a/cpuacct.usage 200.09 ms /cgroup/a/cpuacct.usage 200.06 ms /cgroup/a/cpuacct.usage By returning unused quota we're able to both stably consume our desired quota and prevent unintentional overages due to the abuse of slack quota from previous quota periods (especially on a large machine). Signed-off-by: Paul Turner <pjt@google.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/20110721184758.306848658@google.com Signed-off-by: Ingo Molnar <mingo@elte.hu>
2011-07-21 16:43:41 +00:00
if (!runtime)
return;
sched/fair: Fix race between runtime distribution and assignment Currently, there is a potential race between distribute_cfs_runtime() and assign_cfs_rq_runtime(). Race happens when cfs_b->runtime is read, distributes without holding lock and finds out there is not enough runtime to charge against after distribution. Because assign_cfs_rq_runtime() might be called during distribution, and use cfs_b->runtime at the same time. Fibtest is the tool to test this race. Assume all gcfs_rq is throttled and cfs period timer runs, slow threads might run and sleep, returning unused cfs_rq runtime and keeping min_cfs_rq_runtime in their local pool. If all this happens sufficiently quickly, cfs_b->runtime will drop a lot. If runtime distributed is large too, over-use of runtime happens. A runtime over-using by about 70 percent of quota is seen when we test fibtest on a 96-core machine. We run fibtest with 1 fast thread and 95 slow threads in test group, configure 10ms quota for this group and see the CPU usage of fibtest is 17.0%, which is far more than the expected 10%. On a smaller machine with 32 cores, we also run fibtest with 96 threads. CPU usage is more than 12%, which is also more than expected 10%. This shows that on similar workloads, this race do affect CPU bandwidth control. Solve this by holding lock inside distribute_cfs_runtime(). Fixes: c06f04c70489 ("sched: Fix potential near-infinite distribute_cfs_runtime() loop") Reviewed-by: Ben Segall <bsegall@google.com> Signed-off-by: Huaixin Chang <changhuaixin@linux.alibaba.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Link: https://lore.kernel.org/lkml/20200325092602.22471-1-changhuaixin@linux.alibaba.com/
2020-03-27 03:26:25 +00:00
distribute_cfs_runtime(cfs_b);
sched: Return unused runtime on group dequeue When a local cfs_rq blocks we return the majority of its remaining quota to the global bandwidth pool for use by other runqueues. We do this only when the quota is current and there is more than min_cfs_rq_quota [1ms by default] of runtime remaining on the rq. In the case where there are throttled runqueues and we have sufficient bandwidth to meter out a slice, a second timer is kicked off to handle this delivery, unthrottling where appropriate. Using a 'worst case' antagonist which executes on each cpu for 1ms before moving onto the next on a fairly large machine: no quota generations: 197.47 ms /cgroup/a/cpuacct.usage 199.46 ms /cgroup/a/cpuacct.usage 205.46 ms /cgroup/a/cpuacct.usage 198.46 ms /cgroup/a/cpuacct.usage 208.39 ms /cgroup/a/cpuacct.usage Since we are allowed to use "stale" quota our usage is effectively bounded by the rate of input into the global pool and performance is relatively stable. with quota generations [1s increments]: 119.58 ms /cgroup/a/cpuacct.usage 119.65 ms /cgroup/a/cpuacct.usage 119.64 ms /cgroup/a/cpuacct.usage 119.63 ms /cgroup/a/cpuacct.usage 119.60 ms /cgroup/a/cpuacct.usage The large deficit here is due to quota generations (/intentionally/) preventing us from now using previously stranded slack quota. The cost is that this quota becomes unavailable. with quota generations and quota return: 200.09 ms /cgroup/a/cpuacct.usage 200.09 ms /cgroup/a/cpuacct.usage 198.09 ms /cgroup/a/cpuacct.usage 200.09 ms /cgroup/a/cpuacct.usage 200.06 ms /cgroup/a/cpuacct.usage By returning unused quota we're able to both stably consume our desired quota and prevent unintentional overages due to the abuse of slack quota from previous quota periods (especially on a large machine). Signed-off-by: Paul Turner <pjt@google.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/20110721184758.306848658@google.com Signed-off-by: Ingo Molnar <mingo@elte.hu>
2011-07-21 16:43:41 +00:00
raw_spin_lock_irqsave(&cfs_b->lock, flags);
sched/fair: Fix throttle_list starvation with low CFS quota With a very low cpu.cfs_quota_us setting, such as the minimum of 1000, distribute_cfs_runtime may not empty the throttled_list before it runs out of runtime to distribute. In that case, due to the change from c06f04c7048 to put throttled entries at the head of the list, later entries on the list will starve. Essentially, the same X processes will get pulled off the list, given CPU time and then, when expired, get put back on the head of the list where distribute_cfs_runtime will give runtime to the same set of processes leaving the rest. Fix the issue by setting a bit in struct cfs_bandwidth when distribute_cfs_runtime is running, so that the code in throttle_cfs_rq can decide to put the throttled entry on the tail or the head of the list. The bit is set/cleared by the callers of distribute_cfs_runtime while they hold cfs_bandwidth->lock. This is easy to reproduce with a handful of CPU consumers. I use 'crash' on the live system. In some cases you can simply look at the throttled list and see the later entries are not changing: crash> list cfs_rq.throttled_list -H 0xffff90b54f6ade40 -s cfs_rq.runtime_remaining | paste - - | awk '{print $1" "$4}' | pr -t -n3 1 ffff90b56cb2d200 -976050 2 ffff90b56cb2cc00 -484925 3 ffff90b56cb2bc00 -658814 4 ffff90b56cb2ba00 -275365 5 ffff90b166a45600 -135138 6 ffff90b56cb2da00 -282505 7 ffff90b56cb2e000 -148065 8 ffff90b56cb2fa00 -872591 9 ffff90b56cb2c000 -84687 10 ffff90b56cb2f000 -87237 11 ffff90b166a40a00 -164582 crash> list cfs_rq.throttled_list -H 0xffff90b54f6ade40 -s cfs_rq.runtime_remaining | paste - - | awk '{print $1" "$4}' | pr -t -n3 1 ffff90b56cb2d200 -994147 2 ffff90b56cb2cc00 -306051 3 ffff90b56cb2bc00 -961321 4 ffff90b56cb2ba00 -24490 5 ffff90b166a45600 -135138 6 ffff90b56cb2da00 -282505 7 ffff90b56cb2e000 -148065 8 ffff90b56cb2fa00 -872591 9 ffff90b56cb2c000 -84687 10 ffff90b56cb2f000 -87237 11 ffff90b166a40a00 -164582 Sometimes it is easier to see by finding a process getting starved and looking at the sched_info: crash> task ffff8eb765994500 sched_info PID: 7800 TASK: ffff8eb765994500 CPU: 16 COMMAND: "cputest" sched_info = { pcount = 8, run_delay = 697094208, last_arrival = 240260125039, last_queued = 240260327513 }, crash> task ffff8eb765994500 sched_info PID: 7800 TASK: ffff8eb765994500 CPU: 16 COMMAND: "cputest" sched_info = { pcount = 8, run_delay = 697094208, last_arrival = 240260125039, last_queued = 240260327513 }, Signed-off-by: Phil Auld <pauld@redhat.com> Reviewed-by: Ben Segall <bsegall@google.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: stable@vger.kernel.org Fixes: c06f04c70489 ("sched: Fix potential near-infinite distribute_cfs_runtime() loop") Link: http://lkml.kernel.org/r/20181008143639.GA4019@pauld.bos.csb Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-10-08 14:36:40 +00:00
cfs_b->distribute_running = 0;
raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
sched: Return unused runtime on group dequeue When a local cfs_rq blocks we return the majority of its remaining quota to the global bandwidth pool for use by other runqueues. We do this only when the quota is current and there is more than min_cfs_rq_quota [1ms by default] of runtime remaining on the rq. In the case where there are throttled runqueues and we have sufficient bandwidth to meter out a slice, a second timer is kicked off to handle this delivery, unthrottling where appropriate. Using a 'worst case' antagonist which executes on each cpu for 1ms before moving onto the next on a fairly large machine: no quota generations: 197.47 ms /cgroup/a/cpuacct.usage 199.46 ms /cgroup/a/cpuacct.usage 205.46 ms /cgroup/a/cpuacct.usage 198.46 ms /cgroup/a/cpuacct.usage 208.39 ms /cgroup/a/cpuacct.usage Since we are allowed to use "stale" quota our usage is effectively bounded by the rate of input into the global pool and performance is relatively stable. with quota generations [1s increments]: 119.58 ms /cgroup/a/cpuacct.usage 119.65 ms /cgroup/a/cpuacct.usage 119.64 ms /cgroup/a/cpuacct.usage 119.63 ms /cgroup/a/cpuacct.usage 119.60 ms /cgroup/a/cpuacct.usage The large deficit here is due to quota generations (/intentionally/) preventing us from now using previously stranded slack quota. The cost is that this quota becomes unavailable. with quota generations and quota return: 200.09 ms /cgroup/a/cpuacct.usage 200.09 ms /cgroup/a/cpuacct.usage 198.09 ms /cgroup/a/cpuacct.usage 200.09 ms /cgroup/a/cpuacct.usage 200.06 ms /cgroup/a/cpuacct.usage By returning unused quota we're able to both stably consume our desired quota and prevent unintentional overages due to the abuse of slack quota from previous quota periods (especially on a large machine). Signed-off-by: Paul Turner <pjt@google.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/20110721184758.306848658@google.com Signed-off-by: Ingo Molnar <mingo@elte.hu>
2011-07-21 16:43:41 +00:00
}
/*
* When a group wakes up we want to make sure that its quota is not already
* expired/exceeded, otherwise it may be allowed to steal additional ticks of
* runtime as update_curr() throttling can not not trigger until it's on-rq.
*/
static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
{
if (!cfs_bandwidth_used())
return;
/* an active group must be handled by the update_curr()->put() path */
if (!cfs_rq->runtime_enabled || cfs_rq->curr)
return;
/* ensure the group is not already throttled */
if (cfs_rq_throttled(cfs_rq))
return;
/* update runtime allocation */
account_cfs_rq_runtime(cfs_rq, 0);
if (cfs_rq->runtime_remaining <= 0)
throttle_cfs_rq(cfs_rq);
}
static void sync_throttle(struct task_group *tg, int cpu)
{
struct cfs_rq *pcfs_rq, *cfs_rq;
if (!cfs_bandwidth_used())
return;
if (!tg->parent)
return;
cfs_rq = tg->cfs_rq[cpu];
pcfs_rq = tg->parent->cfs_rq[cpu];
cfs_rq->throttle_count = pcfs_rq->throttle_count;
cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
}
/* conditionally throttle active cfs_rq's from put_prev_entity() */
static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
{
if (!cfs_bandwidth_used())
return false;
if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
return false;
/*
* it's possible for a throttled entity to be forced into a running
* state (e.g. set_curr_task), in this case we're finished.
*/
if (cfs_rq_throttled(cfs_rq))
return true;
return throttle_cfs_rq(cfs_rq);
}
static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
{
struct cfs_bandwidth *cfs_b =
container_of(timer, struct cfs_bandwidth, slack_timer);
sched: Cleanup bandwidth timers Roman reported a 3 cpu lockup scenario involving __start_cfs_bandwidth(). The more I look at that code the more I'm convinced its crack, that entire __start_cfs_bandwidth() thing is brain melting, we don't need to cancel a timer before starting it, *hrtimer_start*() will happily remove the timer for you if its still enqueued. Removing that, removes a big part of the problem, no more ugly cancel loop to get stuck in. So now, if I understand things right, the entire reason you have this cfs_b->lock guarded ->timer_active nonsense is to make sure we don't accidentally lose the timer. It appears to me that it should be possible to guarantee that same by unconditionally (re)starting the timer when !queued. Because regardless what hrtimer::function will return, if we beat it to (re)enqueue the timer, it doesn't matter. Now, because hrtimers don't come with any serialization guarantees we must ensure both handler and (re)start loop serialize their access to the hrtimer to avoid both trying to forward the timer at the same time. Update the rt bandwidth timer to match. This effectively reverts: 09dc4ab03936 ("sched/fair: Fix tg_set_cfs_bandwidth() deadlock on rq->lock"). Reported-by: Roman Gushchin <klamm@yandex-team.ru> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Ben Segall <bsegall@google.com> Cc: Paul Turner <pjt@google.com> Link: http://lkml.kernel.org/r/20150415095011.804589208@infradead.org Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
2015-04-15 09:41:57 +00:00
do_sched_cfs_slack_timer(cfs_b);
return HRTIMER_NORESTART;
}
sched/fair: Limit sched_cfs_period_timer() loop to avoid hard lockup With extremely short cfs_period_us setting on a parent task group with a large number of children the for loop in sched_cfs_period_timer() can run until the watchdog fires. There is no guarantee that the call to hrtimer_forward_now() will ever return 0. The large number of children can make do_sched_cfs_period_timer() take longer than the period. NMI watchdog: Watchdog detected hard LOCKUP on cpu 24 RIP: 0010:tg_nop+0x0/0x10 <IRQ> walk_tg_tree_from+0x29/0xb0 unthrottle_cfs_rq+0xe0/0x1a0 distribute_cfs_runtime+0xd3/0xf0 sched_cfs_period_timer+0xcb/0x160 ? sched_cfs_slack_timer+0xd0/0xd0 __hrtimer_run_queues+0xfb/0x270 hrtimer_interrupt+0x122/0x270 smp_apic_timer_interrupt+0x6a/0x140 apic_timer_interrupt+0xf/0x20 </IRQ> To prevent this we add protection to the loop that detects when the loop has run too many times and scales the period and quota up, proportionally, so that the timer can complete before then next period expires. This preserves the relative runtime quota while preventing the hard lockup. A warning is issued reporting this state and the new values. Signed-off-by: Phil Auld <pauld@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: <stable@vger.kernel.org> Cc: Anton Blanchard <anton@ozlabs.org> Cc: Ben Segall <bsegall@google.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: https://lkml.kernel.org/r/20190319130005.25492-1-pauld@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-03-19 13:00:05 +00:00
extern const u64 max_cfs_quota_period;
static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
{
struct cfs_bandwidth *cfs_b =
container_of(timer, struct cfs_bandwidth, period_timer);
unsigned long flags;
int overrun;
int idle = 0;
sched/fair: Limit sched_cfs_period_timer() loop to avoid hard lockup With extremely short cfs_period_us setting on a parent task group with a large number of children the for loop in sched_cfs_period_timer() can run until the watchdog fires. There is no guarantee that the call to hrtimer_forward_now() will ever return 0. The large number of children can make do_sched_cfs_period_timer() take longer than the period. NMI watchdog: Watchdog detected hard LOCKUP on cpu 24 RIP: 0010:tg_nop+0x0/0x10 <IRQ> walk_tg_tree_from+0x29/0xb0 unthrottle_cfs_rq+0xe0/0x1a0 distribute_cfs_runtime+0xd3/0xf0 sched_cfs_period_timer+0xcb/0x160 ? sched_cfs_slack_timer+0xd0/0xd0 __hrtimer_run_queues+0xfb/0x270 hrtimer_interrupt+0x122/0x270 smp_apic_timer_interrupt+0x6a/0x140 apic_timer_interrupt+0xf/0x20 </IRQ> To prevent this we add protection to the loop that detects when the loop has run too many times and scales the period and quota up, proportionally, so that the timer can complete before then next period expires. This preserves the relative runtime quota while preventing the hard lockup. A warning is issued reporting this state and the new values. Signed-off-by: Phil Auld <pauld@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: <stable@vger.kernel.org> Cc: Anton Blanchard <anton@ozlabs.org> Cc: Ben Segall <bsegall@google.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: https://lkml.kernel.org/r/20190319130005.25492-1-pauld@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-03-19 13:00:05 +00:00
int count = 0;
raw_spin_lock_irqsave(&cfs_b->lock, flags);
for (;;) {
sched: Cleanup bandwidth timers Roman reported a 3 cpu lockup scenario involving __start_cfs_bandwidth(). The more I look at that code the more I'm convinced its crack, that entire __start_cfs_bandwidth() thing is brain melting, we don't need to cancel a timer before starting it, *hrtimer_start*() will happily remove the timer for you if its still enqueued. Removing that, removes a big part of the problem, no more ugly cancel loop to get stuck in. So now, if I understand things right, the entire reason you have this cfs_b->lock guarded ->timer_active nonsense is to make sure we don't accidentally lose the timer. It appears to me that it should be possible to guarantee that same by unconditionally (re)starting the timer when !queued. Because regardless what hrtimer::function will return, if we beat it to (re)enqueue the timer, it doesn't matter. Now, because hrtimers don't come with any serialization guarantees we must ensure both handler and (re)start loop serialize their access to the hrtimer to avoid both trying to forward the timer at the same time. Update the rt bandwidth timer to match. This effectively reverts: 09dc4ab03936 ("sched/fair: Fix tg_set_cfs_bandwidth() deadlock on rq->lock"). Reported-by: Roman Gushchin <klamm@yandex-team.ru> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Ben Segall <bsegall@google.com> Cc: Paul Turner <pjt@google.com> Link: http://lkml.kernel.org/r/20150415095011.804589208@infradead.org Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
2015-04-15 09:41:57 +00:00
overrun = hrtimer_forward_now(timer, cfs_b->period);
if (!overrun)
break;
sched/fair: Limit sched_cfs_period_timer() loop to avoid hard lockup With extremely short cfs_period_us setting on a parent task group with a large number of children the for loop in sched_cfs_period_timer() can run until the watchdog fires. There is no guarantee that the call to hrtimer_forward_now() will ever return 0. The large number of children can make do_sched_cfs_period_timer() take longer than the period. NMI watchdog: Watchdog detected hard LOCKUP on cpu 24 RIP: 0010:tg_nop+0x0/0x10 <IRQ> walk_tg_tree_from+0x29/0xb0 unthrottle_cfs_rq+0xe0/0x1a0 distribute_cfs_runtime+0xd3/0xf0 sched_cfs_period_timer+0xcb/0x160 ? sched_cfs_slack_timer+0xd0/0xd0 __hrtimer_run_queues+0xfb/0x270 hrtimer_interrupt+0x122/0x270 smp_apic_timer_interrupt+0x6a/0x140 apic_timer_interrupt+0xf/0x20 </IRQ> To prevent this we add protection to the loop that detects when the loop has run too many times and scales the period and quota up, proportionally, so that the timer can complete before then next period expires. This preserves the relative runtime quota while preventing the hard lockup. A warning is issued reporting this state and the new values. Signed-off-by: Phil Auld <pauld@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: <stable@vger.kernel.org> Cc: Anton Blanchard <anton@ozlabs.org> Cc: Ben Segall <bsegall@google.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: https://lkml.kernel.org/r/20190319130005.25492-1-pauld@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-03-19 13:00:05 +00:00
if (++count > 3) {
u64 new, old = ktime_to_ns(cfs_b->period);
sched/fair: Scale bandwidth quota and period without losing quota/period ratio precision The quota/period ratio is used to ensure a child task group won't get more bandwidth than the parent task group, and is calculated as: normalized_cfs_quota() = [(quota_us << 20) / period_us] If the quota/period ratio was changed during this scaling due to precision loss, it will cause inconsistency between parent and child task groups. See below example: A userspace container manager (kubelet) does three operations: 1) Create a parent cgroup, set quota to 1,000us and period to 10,000us. 2) Create a few children cgroups. 3) Set quota to 1,000us and period to 10,000us on a child cgroup. These operations are expected to succeed. However, if the scaling of 147/128 happens before step 3, quota and period of the parent cgroup will be changed: new_quota: 1148437ns, 1148us new_period: 11484375ns, 11484us And when step 3 comes in, the ratio of the child cgroup will be 104857, which will be larger than the parent cgroup ratio (104821), and will fail. Scaling them by a factor of 2 will fix the problem. Tested-by: Phil Auld <pauld@redhat.com> Signed-off-by: Xuewei Zhang <xueweiz@google.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Phil Auld <pauld@redhat.com> Cc: Anton Blanchard <anton@ozlabs.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vincent Guittot <vincent.guittot@linaro.org> Fixes: 2e8e19226398 ("sched/fair: Limit sched_cfs_period_timer() loop to avoid hard lockup") Link: https://lkml.kernel.org/r/20191004001243.140897-1-xueweiz@google.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-04 00:12:43 +00:00
/*
* Grow period by a factor of 2 to avoid losing precision.
* Precision loss in the quota/period ratio can cause __cfs_schedulable
* to fail.
*/
new = old * 2;
if (new < max_cfs_quota_period) {
cfs_b->period = ns_to_ktime(new);
cfs_b->quota *= 2;
pr_warn_ratelimited(
"cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n",
smp_processor_id(),
div_u64(new, NSEC_PER_USEC),
div_u64(cfs_b->quota, NSEC_PER_USEC));
} else {
pr_warn_ratelimited(
"cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n",
smp_processor_id(),
div_u64(old, NSEC_PER_USEC),
div_u64(cfs_b->quota, NSEC_PER_USEC));
}
sched/fair: Limit sched_cfs_period_timer() loop to avoid hard lockup With extremely short cfs_period_us setting on a parent task group with a large number of children the for loop in sched_cfs_period_timer() can run until the watchdog fires. There is no guarantee that the call to hrtimer_forward_now() will ever return 0. The large number of children can make do_sched_cfs_period_timer() take longer than the period. NMI watchdog: Watchdog detected hard LOCKUP on cpu 24 RIP: 0010:tg_nop+0x0/0x10 <IRQ> walk_tg_tree_from+0x29/0xb0 unthrottle_cfs_rq+0xe0/0x1a0 distribute_cfs_runtime+0xd3/0xf0 sched_cfs_period_timer+0xcb/0x160 ? sched_cfs_slack_timer+0xd0/0xd0 __hrtimer_run_queues+0xfb/0x270 hrtimer_interrupt+0x122/0x270 smp_apic_timer_interrupt+0x6a/0x140 apic_timer_interrupt+0xf/0x20 </IRQ> To prevent this we add protection to the loop that detects when the loop has run too many times and scales the period and quota up, proportionally, so that the timer can complete before then next period expires. This preserves the relative runtime quota while preventing the hard lockup. A warning is issued reporting this state and the new values. Signed-off-by: Phil Auld <pauld@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: <stable@vger.kernel.org> Cc: Anton Blanchard <anton@ozlabs.org> Cc: Ben Segall <bsegall@google.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: https://lkml.kernel.org/r/20190319130005.25492-1-pauld@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-03-19 13:00:05 +00:00
/* reset count so we don't come right back in here */
count = 0;
}
idle = do_sched_cfs_period_timer(cfs_b, overrun, flags);
}
sched,perf: Fix periodic timers In the below two commits (see Fixes) we have periodic timers that can stop themselves when they're no longer required, but need to be (re)-started when their idle condition changes. Further complications is that we want the timer handler to always do the forward such that it will always correctly deal with the overruns, and we do not want to race such that the handler has already decided to stop, but the (external) restart sees the timer still active and we end up with a 'lost' timer. The problem with the current code is that the re-start can come before the callback does the forward, at which point the forward from the callback will WARN about forwarding an enqueued timer. Now, conceptually its easy to detect if you're before or after the fwd by comparing the expiration time against the current time. Of course, that's expensive (and racy) because we don't have the current time. Alternatively one could cache this state inside the timer, but then everybody pays the overhead of maintaining this extra state, and that is undesired. The only other option that I could see is the external timer_active variable, which I tried to kill before. I would love a nicer interface for this seemingly simple 'problem' but alas. Fixes: 272325c4821f ("perf: Fix mux_interval hrtimer wreckage") Fixes: 77a4d1a1b9a1 ("sched: Cleanup bandwidth timers") Cc: pjt@google.com Cc: tglx@linutronix.de Cc: klamm@yandex-team.ru Cc: mingo@kernel.org Cc: bsegall@google.com Cc: hpa@zytor.com Cc: Sasha Levin <sasha.levin@oracle.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: http://lkml.kernel.org/r/20150514102311.GX21418@twins.programming.kicks-ass.net
2015-05-14 10:23:11 +00:00
if (idle)
cfs_b->period_active = 0;
raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
}
void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
{
raw_spin_lock_init(&cfs_b->lock);
cfs_b->runtime = 0;
cfs_b->quota = RUNTIME_INF;
cfs_b->period = ns_to_ktime(default_cfs_period());
INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
sched,perf: Fix periodic timers In the below two commits (see Fixes) we have periodic timers that can stop themselves when they're no longer required, but need to be (re)-started when their idle condition changes. Further complications is that we want the timer handler to always do the forward such that it will always correctly deal with the overruns, and we do not want to race such that the handler has already decided to stop, but the (external) restart sees the timer still active and we end up with a 'lost' timer. The problem with the current code is that the re-start can come before the callback does the forward, at which point the forward from the callback will WARN about forwarding an enqueued timer. Now, conceptually its easy to detect if you're before or after the fwd by comparing the expiration time against the current time. Of course, that's expensive (and racy) because we don't have the current time. Alternatively one could cache this state inside the timer, but then everybody pays the overhead of maintaining this extra state, and that is undesired. The only other option that I could see is the external timer_active variable, which I tried to kill before. I would love a nicer interface for this seemingly simple 'problem' but alas. Fixes: 272325c4821f ("perf: Fix mux_interval hrtimer wreckage") Fixes: 77a4d1a1b9a1 ("sched: Cleanup bandwidth timers") Cc: pjt@google.com Cc: tglx@linutronix.de Cc: klamm@yandex-team.ru Cc: mingo@kernel.org Cc: bsegall@google.com Cc: hpa@zytor.com Cc: Sasha Levin <sasha.levin@oracle.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: http://lkml.kernel.org/r/20150514102311.GX21418@twins.programming.kicks-ass.net
2015-05-14 10:23:11 +00:00
hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
cfs_b->period_timer.function = sched_cfs_period_timer;
hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
cfs_b->slack_timer.function = sched_cfs_slack_timer;
sched/fair: Fix throttle_list starvation with low CFS quota With a very low cpu.cfs_quota_us setting, such as the minimum of 1000, distribute_cfs_runtime may not empty the throttled_list before it runs out of runtime to distribute. In that case, due to the change from c06f04c7048 to put throttled entries at the head of the list, later entries on the list will starve. Essentially, the same X processes will get pulled off the list, given CPU time and then, when expired, get put back on the head of the list where distribute_cfs_runtime will give runtime to the same set of processes leaving the rest. Fix the issue by setting a bit in struct cfs_bandwidth when distribute_cfs_runtime is running, so that the code in throttle_cfs_rq can decide to put the throttled entry on the tail or the head of the list. The bit is set/cleared by the callers of distribute_cfs_runtime while they hold cfs_bandwidth->lock. This is easy to reproduce with a handful of CPU consumers. I use 'crash' on the live system. In some cases you can simply look at the throttled list and see the later entries are not changing: crash> list cfs_rq.throttled_list -H 0xffff90b54f6ade40 -s cfs_rq.runtime_remaining | paste - - | awk '{print $1" "$4}' | pr -t -n3 1 ffff90b56cb2d200 -976050 2 ffff90b56cb2cc00 -484925 3 ffff90b56cb2bc00 -658814 4 ffff90b56cb2ba00 -275365 5 ffff90b166a45600 -135138 6 ffff90b56cb2da00 -282505 7 ffff90b56cb2e000 -148065 8 ffff90b56cb2fa00 -872591 9 ffff90b56cb2c000 -84687 10 ffff90b56cb2f000 -87237 11 ffff90b166a40a00 -164582 crash> list cfs_rq.throttled_list -H 0xffff90b54f6ade40 -s cfs_rq.runtime_remaining | paste - - | awk '{print $1" "$4}' | pr -t -n3 1 ffff90b56cb2d200 -994147 2 ffff90b56cb2cc00 -306051 3 ffff90b56cb2bc00 -961321 4 ffff90b56cb2ba00 -24490 5 ffff90b166a45600 -135138 6 ffff90b56cb2da00 -282505 7 ffff90b56cb2e000 -148065 8 ffff90b56cb2fa00 -872591 9 ffff90b56cb2c000 -84687 10 ffff90b56cb2f000 -87237 11 ffff90b166a40a00 -164582 Sometimes it is easier to see by finding a process getting starved and looking at the sched_info: crash> task ffff8eb765994500 sched_info PID: 7800 TASK: ffff8eb765994500 CPU: 16 COMMAND: "cputest" sched_info = { pcount = 8, run_delay = 697094208, last_arrival = 240260125039, last_queued = 240260327513 }, crash> task ffff8eb765994500 sched_info PID: 7800 TASK: ffff8eb765994500 CPU: 16 COMMAND: "cputest" sched_info = { pcount = 8, run_delay = 697094208, last_arrival = 240260125039, last_queued = 240260327513 }, Signed-off-by: Phil Auld <pauld@redhat.com> Reviewed-by: Ben Segall <bsegall@google.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: stable@vger.kernel.org Fixes: c06f04c70489 ("sched: Fix potential near-infinite distribute_cfs_runtime() loop") Link: http://lkml.kernel.org/r/20181008143639.GA4019@pauld.bos.csb Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-10-08 14:36:40 +00:00
cfs_b->distribute_running = 0;
cfs_b->slack_started = false;
}
static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
{
cfs_rq->runtime_enabled = 0;
INIT_LIST_HEAD(&cfs_rq->throttled_list);
}
sched: Cleanup bandwidth timers Roman reported a 3 cpu lockup scenario involving __start_cfs_bandwidth(). The more I look at that code the more I'm convinced its crack, that entire __start_cfs_bandwidth() thing is brain melting, we don't need to cancel a timer before starting it, *hrtimer_start*() will happily remove the timer for you if its still enqueued. Removing that, removes a big part of the problem, no more ugly cancel loop to get stuck in. So now, if I understand things right, the entire reason you have this cfs_b->lock guarded ->timer_active nonsense is to make sure we don't accidentally lose the timer. It appears to me that it should be possible to guarantee that same by unconditionally (re)starting the timer when !queued. Because regardless what hrtimer::function will return, if we beat it to (re)enqueue the timer, it doesn't matter. Now, because hrtimers don't come with any serialization guarantees we must ensure both handler and (re)start loop serialize their access to the hrtimer to avoid both trying to forward the timer at the same time. Update the rt bandwidth timer to match. This effectively reverts: 09dc4ab03936 ("sched/fair: Fix tg_set_cfs_bandwidth() deadlock on rq->lock"). Reported-by: Roman Gushchin <klamm@yandex-team.ru> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Ben Segall <bsegall@google.com> Cc: Paul Turner <pjt@google.com> Link: http://lkml.kernel.org/r/20150415095011.804589208@infradead.org Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
2015-04-15 09:41:57 +00:00
void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
{
sched,perf: Fix periodic timers In the below two commits (see Fixes) we have periodic timers that can stop themselves when they're no longer required, but need to be (re)-started when their idle condition changes. Further complications is that we want the timer handler to always do the forward such that it will always correctly deal with the overruns, and we do not want to race such that the handler has already decided to stop, but the (external) restart sees the timer still active and we end up with a 'lost' timer. The problem with the current code is that the re-start can come before the callback does the forward, at which point the forward from the callback will WARN about forwarding an enqueued timer. Now, conceptually its easy to detect if you're before or after the fwd by comparing the expiration time against the current time. Of course, that's expensive (and racy) because we don't have the current time. Alternatively one could cache this state inside the timer, but then everybody pays the overhead of maintaining this extra state, and that is undesired. The only other option that I could see is the external timer_active variable, which I tried to kill before. I would love a nicer interface for this seemingly simple 'problem' but alas. Fixes: 272325c4821f ("perf: Fix mux_interval hrtimer wreckage") Fixes: 77a4d1a1b9a1 ("sched: Cleanup bandwidth timers") Cc: pjt@google.com Cc: tglx@linutronix.de Cc: klamm@yandex-team.ru Cc: mingo@kernel.org Cc: bsegall@google.com Cc: hpa@zytor.com Cc: Sasha Levin <sasha.levin@oracle.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: http://lkml.kernel.org/r/20150514102311.GX21418@twins.programming.kicks-ass.net
2015-05-14 10:23:11 +00:00
lockdep_assert_held(&cfs_b->lock);
if (cfs_b->period_active)
return;
cfs_b->period_active = 1;
hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
}
static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
{
sched/fair: Fix RCU stall upon -ENOMEM in sched_create_group() When alloc_fair_sched_group() in sched_create_group() fails, free_sched_group() is called, and free_fair_sched_group() is called by free_sched_group(). Since destroy_cfs_bandwidth() is called by free_fair_sched_group() without calling init_cfs_bandwidth(), RCU stall occurs at hrtimer_cancel(): INFO: rcu_sched self-detected stall on CPU { 1} (t=60000 jiffies g=13074 c=13073 q=0) Task dump for CPU 1: (fprintd) R running task 0 6249 1 0x00000088 ... Call Trace: <IRQ> [<ffffffff81094988>] sched_show_task+0xa8/0x110 [<ffffffff81097acd>] dump_cpu_task+0x3d/0x50 [<ffffffff810c3a80>] rcu_dump_cpu_stacks+0x90/0xd0 [<ffffffff810c7751>] rcu_check_callbacks+0x491/0x700 [<ffffffff810cbf2b>] update_process_times+0x4b/0x80 [<ffffffff810db046>] tick_sched_handle.isra.20+0x36/0x50 [<ffffffff810db0a2>] tick_sched_timer+0x42/0x70 [<ffffffff810ccb19>] __run_hrtimer+0x69/0x1a0 [<ffffffff810db060>] ? tick_sched_handle.isra.20+0x50/0x50 [<ffffffff810ccedf>] hrtimer_interrupt+0xef/0x230 [<ffffffff810452cb>] local_apic_timer_interrupt+0x3b/0x70 [<ffffffff8164a465>] smp_apic_timer_interrupt+0x45/0x60 [<ffffffff816485bd>] apic_timer_interrupt+0x6d/0x80 <EOI> [<ffffffff810cc588>] ? lock_hrtimer_base.isra.23+0x18/0x50 [<ffffffff81193cf1>] ? __kmalloc+0x211/0x230 [<ffffffff810cc9d2>] hrtimer_try_to_cancel+0x22/0xd0 [<ffffffff81193cf1>] ? __kmalloc+0x211/0x230 [<ffffffff810ccaa2>] hrtimer_cancel+0x22/0x30 [<ffffffff810a3cb5>] free_fair_sched_group+0x25/0xd0 [<ffffffff8108df46>] free_sched_group+0x16/0x40 [<ffffffff810971bb>] sched_create_group+0x4b/0x80 [<ffffffff810aa383>] sched_autogroup_create_attach+0x43/0x1c0 [<ffffffff8107dc9c>] sys_setsid+0x7c/0x110 [<ffffffff81647729>] system_call_fastpath+0x12/0x17 Check whether init_cfs_bandwidth() was called before calling destroy_cfs_bandwidth(). Signed-off-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp> [ Move the check into destroy_cfs_bandwidth() to aid compilability. ] Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Paul Turner <pjt@google.com> Cc: Ben Segall <bsegall@google.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/201412252210.GCC30204.SOMVFFOtQJFLOH@I-love.SAKURA.ne.jp Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-12-25 06:51:21 +00:00
/* init_cfs_bandwidth() was not called */
if (!cfs_b->throttled_cfs_rq.next)
return;
hrtimer_cancel(&cfs_b->period_timer);
hrtimer_cancel(&cfs_b->slack_timer);
}
/*
* Both these CPU hotplug callbacks race against unregister_fair_sched_group()
*
* The race is harmless, since modifying bandwidth settings of unhooked group
* bits doesn't do much.
*/
/* cpu online calback */
sched/fair: Disable runtime_enabled on dying rq We kill rq->rd on the CPU_DOWN_PREPARE stage: cpuset_cpu_inactive -> cpuset_update_active_cpus -> partition_sched_domains -> -> cpu_attach_domain -> rq_attach_root -> set_rq_offline This unthrottles all throttled cfs_rqs. But the cpu is still able to call schedule() till take_cpu_down->__cpu_disable() is called from stop_machine. This case the tasks from just unthrottled cfs_rqs are pickable in a standard scheduler way, and they are picked by dying cpu. The cfs_rqs becomes throttled again, and migrate_tasks() in migration_call skips their tasks (one more unthrottle in migrate_tasks()->CPU_DYING does not happen, because rq->rd is already NULL). Patch sets runtime_enabled to zero. This guarantees, the runtime is not accounted, and the cfs_rqs won't exceed given cfs_rq->runtime_remaining = 1, and tasks will be pickable in migrate_tasks(). runtime_enabled is recalculated again when rq becomes online again. Ben Segall also noticed, we always enable runtime in tg_set_cfs_bandwidth(). Actually, we should do that for online cpus only. To prevent races with unthrottle_offline_cfs_rqs() we take get_online_cpus() lock. Reviewed-by: Ben Segall <bsegall@google.com> Reviewed-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Kirill Tkhai <ktkhai@parallels.com> CC: Konstantin Khorenko <khorenko@parallels.com> CC: Paul Turner <pjt@google.com> CC: Mike Galbraith <umgwanakikbuti@gmail.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1403684382.3462.42.camel@tkhai Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-06-25 08:19:42 +00:00
static void __maybe_unused update_runtime_enabled(struct rq *rq)
{
struct task_group *tg;
sched/fair: Disable runtime_enabled on dying rq We kill rq->rd on the CPU_DOWN_PREPARE stage: cpuset_cpu_inactive -> cpuset_update_active_cpus -> partition_sched_domains -> -> cpu_attach_domain -> rq_attach_root -> set_rq_offline This unthrottles all throttled cfs_rqs. But the cpu is still able to call schedule() till take_cpu_down->__cpu_disable() is called from stop_machine. This case the tasks from just unthrottled cfs_rqs are pickable in a standard scheduler way, and they are picked by dying cpu. The cfs_rqs becomes throttled again, and migrate_tasks() in migration_call skips their tasks (one more unthrottle in migrate_tasks()->CPU_DYING does not happen, because rq->rd is already NULL). Patch sets runtime_enabled to zero. This guarantees, the runtime is not accounted, and the cfs_rqs won't exceed given cfs_rq->runtime_remaining = 1, and tasks will be pickable in migrate_tasks(). runtime_enabled is recalculated again when rq becomes online again. Ben Segall also noticed, we always enable runtime in tg_set_cfs_bandwidth(). Actually, we should do that for online cpus only. To prevent races with unthrottle_offline_cfs_rqs() we take get_online_cpus() lock. Reviewed-by: Ben Segall <bsegall@google.com> Reviewed-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Kirill Tkhai <ktkhai@parallels.com> CC: Konstantin Khorenko <khorenko@parallels.com> CC: Paul Turner <pjt@google.com> CC: Mike Galbraith <umgwanakikbuti@gmail.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1403684382.3462.42.camel@tkhai Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-06-25 08:19:42 +00:00
lockdep_assert_held(&rq->lock);
rcu_read_lock();
list_for_each_entry_rcu(tg, &task_groups, list) {
struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
sched/fair: Disable runtime_enabled on dying rq We kill rq->rd on the CPU_DOWN_PREPARE stage: cpuset_cpu_inactive -> cpuset_update_active_cpus -> partition_sched_domains -> -> cpu_attach_domain -> rq_attach_root -> set_rq_offline This unthrottles all throttled cfs_rqs. But the cpu is still able to call schedule() till take_cpu_down->__cpu_disable() is called from stop_machine. This case the tasks from just unthrottled cfs_rqs are pickable in a standard scheduler way, and they are picked by dying cpu. The cfs_rqs becomes throttled again, and migrate_tasks() in migration_call skips their tasks (one more unthrottle in migrate_tasks()->CPU_DYING does not happen, because rq->rd is already NULL). Patch sets runtime_enabled to zero. This guarantees, the runtime is not accounted, and the cfs_rqs won't exceed given cfs_rq->runtime_remaining = 1, and tasks will be pickable in migrate_tasks(). runtime_enabled is recalculated again when rq becomes online again. Ben Segall also noticed, we always enable runtime in tg_set_cfs_bandwidth(). Actually, we should do that for online cpus only. To prevent races with unthrottle_offline_cfs_rqs() we take get_online_cpus() lock. Reviewed-by: Ben Segall <bsegall@google.com> Reviewed-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Kirill Tkhai <ktkhai@parallels.com> CC: Konstantin Khorenko <khorenko@parallels.com> CC: Paul Turner <pjt@google.com> CC: Mike Galbraith <umgwanakikbuti@gmail.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1403684382.3462.42.camel@tkhai Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-06-25 08:19:42 +00:00
raw_spin_lock(&cfs_b->lock);
cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
raw_spin_unlock(&cfs_b->lock);
}
rcu_read_unlock();
sched/fair: Disable runtime_enabled on dying rq We kill rq->rd on the CPU_DOWN_PREPARE stage: cpuset_cpu_inactive -> cpuset_update_active_cpus -> partition_sched_domains -> -> cpu_attach_domain -> rq_attach_root -> set_rq_offline This unthrottles all throttled cfs_rqs. But the cpu is still able to call schedule() till take_cpu_down->__cpu_disable() is called from stop_machine. This case the tasks from just unthrottled cfs_rqs are pickable in a standard scheduler way, and they are picked by dying cpu. The cfs_rqs becomes throttled again, and migrate_tasks() in migration_call skips their tasks (one more unthrottle in migrate_tasks()->CPU_DYING does not happen, because rq->rd is already NULL). Patch sets runtime_enabled to zero. This guarantees, the runtime is not accounted, and the cfs_rqs won't exceed given cfs_rq->runtime_remaining = 1, and tasks will be pickable in migrate_tasks(). runtime_enabled is recalculated again when rq becomes online again. Ben Segall also noticed, we always enable runtime in tg_set_cfs_bandwidth(). Actually, we should do that for online cpus only. To prevent races with unthrottle_offline_cfs_rqs() we take get_online_cpus() lock. Reviewed-by: Ben Segall <bsegall@google.com> Reviewed-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Kirill Tkhai <ktkhai@parallels.com> CC: Konstantin Khorenko <khorenko@parallels.com> CC: Paul Turner <pjt@google.com> CC: Mike Galbraith <umgwanakikbuti@gmail.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1403684382.3462.42.camel@tkhai Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-06-25 08:19:42 +00:00
}
/* cpu offline callback */
static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
{
struct task_group *tg;
lockdep_assert_held(&rq->lock);
rcu_read_lock();
list_for_each_entry_rcu(tg, &task_groups, list) {
struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
if (!cfs_rq->runtime_enabled)
continue;
/*
* clock_task is not advancing so we just need to make sure
* there's some valid quota amount
*/
cfs_rq->runtime_remaining = 1;
sched/fair: Disable runtime_enabled on dying rq We kill rq->rd on the CPU_DOWN_PREPARE stage: cpuset_cpu_inactive -> cpuset_update_active_cpus -> partition_sched_domains -> -> cpu_attach_domain -> rq_attach_root -> set_rq_offline This unthrottles all throttled cfs_rqs. But the cpu is still able to call schedule() till take_cpu_down->__cpu_disable() is called from stop_machine. This case the tasks from just unthrottled cfs_rqs are pickable in a standard scheduler way, and they are picked by dying cpu. The cfs_rqs becomes throttled again, and migrate_tasks() in migration_call skips their tasks (one more unthrottle in migrate_tasks()->CPU_DYING does not happen, because rq->rd is already NULL). Patch sets runtime_enabled to zero. This guarantees, the runtime is not accounted, and the cfs_rqs won't exceed given cfs_rq->runtime_remaining = 1, and tasks will be pickable in migrate_tasks(). runtime_enabled is recalculated again when rq becomes online again. Ben Segall also noticed, we always enable runtime in tg_set_cfs_bandwidth(). Actually, we should do that for online cpus only. To prevent races with unthrottle_offline_cfs_rqs() we take get_online_cpus() lock. Reviewed-by: Ben Segall <bsegall@google.com> Reviewed-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Kirill Tkhai <ktkhai@parallels.com> CC: Konstantin Khorenko <khorenko@parallels.com> CC: Paul Turner <pjt@google.com> CC: Mike Galbraith <umgwanakikbuti@gmail.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1403684382.3462.42.camel@tkhai Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-06-25 08:19:42 +00:00
/*
* Offline rq is schedulable till CPU is completely disabled
sched/fair: Disable runtime_enabled on dying rq We kill rq->rd on the CPU_DOWN_PREPARE stage: cpuset_cpu_inactive -> cpuset_update_active_cpus -> partition_sched_domains -> -> cpu_attach_domain -> rq_attach_root -> set_rq_offline This unthrottles all throttled cfs_rqs. But the cpu is still able to call schedule() till take_cpu_down->__cpu_disable() is called from stop_machine. This case the tasks from just unthrottled cfs_rqs are pickable in a standard scheduler way, and they are picked by dying cpu. The cfs_rqs becomes throttled again, and migrate_tasks() in migration_call skips their tasks (one more unthrottle in migrate_tasks()->CPU_DYING does not happen, because rq->rd is already NULL). Patch sets runtime_enabled to zero. This guarantees, the runtime is not accounted, and the cfs_rqs won't exceed given cfs_rq->runtime_remaining = 1, and tasks will be pickable in migrate_tasks(). runtime_enabled is recalculated again when rq becomes online again. Ben Segall also noticed, we always enable runtime in tg_set_cfs_bandwidth(). Actually, we should do that for online cpus only. To prevent races with unthrottle_offline_cfs_rqs() we take get_online_cpus() lock. Reviewed-by: Ben Segall <bsegall@google.com> Reviewed-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Kirill Tkhai <ktkhai@parallels.com> CC: Konstantin Khorenko <khorenko@parallels.com> CC: Paul Turner <pjt@google.com> CC: Mike Galbraith <umgwanakikbuti@gmail.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1403684382.3462.42.camel@tkhai Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-06-25 08:19:42 +00:00
* in take_cpu_down(), so we prevent new cfs throttling here.
*/
cfs_rq->runtime_enabled = 0;
if (cfs_rq_throttled(cfs_rq))
unthrottle_cfs_rq(cfs_rq);
}
rcu_read_unlock();
}
#else /* CONFIG_CFS_BANDWIDTH */
sched/fair: Fix insertion in rq->leaf_cfs_rq_list Sargun reported a crash: "I picked up c40f7d74c741a907cfaeb73a7697081881c497d0 sched/fair: Fix infinite loop in update_blocked_averages() by reverting a9e7f6544b9c and put it on top of 4.19.13. In addition to this, I uninlined list_add_leaf_cfs_rq for debugging. This revealed a new bug that we didn't get to because we kept getting crashes from the previous issue. When we are running with cgroups that are rapidly changing, with CFS bandwidth control, and in addition using the cpusets cgroup, we see this crash. Specifically, it seems to occur with cgroups that are throttled and we change the allowed cpuset." The algorithm used to order cfs_rq in rq->leaf_cfs_rq_list assumes that it will walk down to root the 1st time a cfs_rq is used and we will finish to add either a cfs_rq without parent or a cfs_rq with a parent that is already on the list. But this is not always true in presence of throttling. Because a cfs_rq can be throttled even if it has never been used but other CPUs of the cgroup have already used all the bandwdith, we are not sure to go down to the root and add all cfs_rq in the list. Ensure that all cfs_rq will be added in the list even if they are throttled. [ mingo: Fix !CGROUPS build. ] Reported-by: Sargun Dhillon <sargun@sargun.me> Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: tj@kernel.org Fixes: 9c2791f936ef ("Fix hierarchical order in rq->leaf_cfs_rq_list") Link: https://lkml.kernel.org/r/1548825767-10799-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-01-30 05:22:47 +00:00
static inline bool cfs_bandwidth_used(void)
{
return false;
}
sched/fair: Rework sched_fair time accounting Christian suffers from a bad BIOS that wrecks his i5's TSC sync. This results in him occasionally seeing time going backwards - which crashes the scheduler ... Most of our time accounting can actually handle that except the most common one; the tick time update of sched_fair. There is a further problem with that code; previously we assumed that because we get a tick every TICK_NSEC our time delta could never exceed 32bits and math was simpler. However, ever since Frederic managed to get NO_HZ_FULL merged; this is no longer the case since now a task can run for a long time indeed without getting a tick. It only takes about ~4.2 seconds to overflow our u32 in nanoseconds. This means we not only need to better deal with time going backwards; but also means we need to be able to deal with large deltas. This patch reworks the entire code and uses mul_u64_u32_shr() as proposed by Andy a long while ago. We express our virtual time scale factor in a u32 multiplier and shift right and the 32bit mul_u64_u32_shr() implementation reduces to a single 32x32->64 multiply if the time delta is still short (common case). For 64bit a 64x64->128 multiply can be used if ARCH_SUPPORTS_INT128. Reported-and-Tested-by: Christian Engelmayer <cengelma@gmx.at> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Cc: fweisbec@gmail.com Cc: Paul Turner <pjt@google.com> Cc: Stanislaw Gruszka <sgruszka@redhat.com> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Link: http://lkml.kernel.org/r/20131118172706.GI3866@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-11-18 17:27:06 +00:00
static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
static inline void sync_throttle(struct task_group *tg, int cpu) {}
static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
{
return 0;
}
static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
{
return 0;
}
static inline int throttled_lb_pair(struct task_group *tg,
int src_cpu, int dest_cpu)
{
return 0;
}
void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
#ifdef CONFIG_FAIR_GROUP_SCHED
static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
#endif
static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
{
return NULL;
}
static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
sched/fair: Disable runtime_enabled on dying rq We kill rq->rd on the CPU_DOWN_PREPARE stage: cpuset_cpu_inactive -> cpuset_update_active_cpus -> partition_sched_domains -> -> cpu_attach_domain -> rq_attach_root -> set_rq_offline This unthrottles all throttled cfs_rqs. But the cpu is still able to call schedule() till take_cpu_down->__cpu_disable() is called from stop_machine. This case the tasks from just unthrottled cfs_rqs are pickable in a standard scheduler way, and they are picked by dying cpu. The cfs_rqs becomes throttled again, and migrate_tasks() in migration_call skips their tasks (one more unthrottle in migrate_tasks()->CPU_DYING does not happen, because rq->rd is already NULL). Patch sets runtime_enabled to zero. This guarantees, the runtime is not accounted, and the cfs_rqs won't exceed given cfs_rq->runtime_remaining = 1, and tasks will be pickable in migrate_tasks(). runtime_enabled is recalculated again when rq becomes online again. Ben Segall also noticed, we always enable runtime in tg_set_cfs_bandwidth(). Actually, we should do that for online cpus only. To prevent races with unthrottle_offline_cfs_rqs() we take get_online_cpus() lock. Reviewed-by: Ben Segall <bsegall@google.com> Reviewed-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Kirill Tkhai <ktkhai@parallels.com> CC: Konstantin Khorenko <khorenko@parallels.com> CC: Paul Turner <pjt@google.com> CC: Mike Galbraith <umgwanakikbuti@gmail.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1403684382.3462.42.camel@tkhai Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-06-25 08:19:42 +00:00
static inline void update_runtime_enabled(struct rq *rq) {}
static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
#endif /* CONFIG_CFS_BANDWIDTH */
/**************************************************
* CFS operations on tasks:
*/
#ifdef CONFIG_SCHED_HRTICK
static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
{
struct sched_entity *se = &p->se;
struct cfs_rq *cfs_rq = cfs_rq_of(se);
SCHED_WARN_ON(task_rq(p) != rq);
if (rq->cfs.h_nr_running > 1) {
u64 slice = sched_slice(cfs_rq, se);
u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
s64 delta = slice - ran;
if (delta < 0) {
if (rq->curr == p)
resched_curr(rq);
return;
}
sched, x86: clean up hrtick implementation random uvesafb failures were reported against Gentoo: http://bugs.gentoo.org/show_bug.cgi?id=222799 and Mihai Moldovan bisected it back to: > 8f4d37ec073c17e2d4aa8851df5837d798606d6f is first bad commit > commit 8f4d37ec073c17e2d4aa8851df5837d798606d6f > Author: Peter Zijlstra <a.p.zijlstra@chello.nl> > Date: Fri Jan 25 21:08:29 2008 +0100 > > sched: high-res preemption tick Linus suspected it to be hrtick + vm86 interaction and observed: > Btw, Peter, Ingo: I think that commit is doing bad things. They aren't > _incorrect_ per se, but they are definitely bad. > > Why? > > Using random _TIF_WORK_MASK flags is really impolite for doing > "scheduling" work. There's a reason that arch/x86/kernel/entry_32.S > special-cases the _TIF_NEED_RESCHED flag: we don't want to exit out of > vm86 mode unnecessarily. > > See the "work_notifysig_v86" label, and how it does that > "save_v86_state()" thing etc etc. Right, I never liked having to fiddle with those TIF flags. Initially I needed it because the hrtimer base lock could not nest in the rq lock. That however is fixed these days. Currently the only reason left to fiddle with the TIF flags is remote wakeups. We cannot program a remote cpu's hrtimer. I've been thinking about using the new and improved IPI function call stuff to implement hrtimer_start_on(). However that does require that smp_call_function_single(.wait=0) works from interrupt context - /me looks at the latest series from Jens - Yes that does seem to be supported, good. Here's a stab at cleaning this stuff up ... Mihai reported test success as well. Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Tested-by: Mihai Moldovan <ionic@ionic.de> Cc: Michal Januszewski <spock@gentoo.org> Cc: Antonino Daplas <adaplas@gmail.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-07-18 16:01:23 +00:00
hrtick_start(rq, delta);
}
}
/*
* called from enqueue/dequeue and updates the hrtick when the
* current task is from our class and nr_running is low enough
* to matter.
*/
static void hrtick_update(struct rq *rq)
{
struct task_struct *curr = rq->curr;
sched: Save some hrtick_start_fair cycles hrtick_start_fair() shows up in profiles even when disabled. v3.0.6 taskset -c 3 pipe-test PerfTop: 997 irqs/sec kernel:89.5% exact: 0.0% [1000Hz cycles], (all, CPU: 3) ------------------------------------------------------------------------------------------------ Virgin Patched samples pcnt function samples pcnt function _______ _____ ___________________________ _______ _____ ___________________________ 2880.00 10.2% __schedule 3136.00 11.3% __schedule 1634.00 5.8% pipe_read 1615.00 5.8% pipe_read 1458.00 5.2% system_call 1534.00 5.5% system_call 1382.00 4.9% _raw_spin_lock_irqsave 1412.00 5.1% _raw_spin_lock_irqsave 1202.00 4.3% pipe_write 1255.00 4.5% copy_user_generic_string 1164.00 4.1% copy_user_generic_string 1241.00 4.5% __switch_to 1097.00 3.9% __switch_to 929.00 3.3% mutex_lock 872.00 3.1% mutex_lock 846.00 3.0% mutex_unlock 687.00 2.4% mutex_unlock 804.00 2.9% pipe_write 682.00 2.4% native_sched_clock 713.00 2.6% native_sched_clock 643.00 2.3% system_call_after_swapgs 653.00 2.3% _raw_spin_unlock_irqrestore 617.00 2.2% sched_clock_local 633.00 2.3% fsnotify 612.00 2.2% fsnotify 605.00 2.2% sched_clock_local 596.00 2.1% _raw_spin_unlock_irqrestore 593.00 2.1% system_call_after_swapgs 542.00 1.9% sysret_check 559.00 2.0% sysret_check 467.00 1.7% fget_light 472.00 1.7% fget_light 462.00 1.6% finish_task_switch 461.00 1.7% finish_task_switch 437.00 1.5% vfs_write 442.00 1.6% vfs_write 431.00 1.5% do_sync_write 428.00 1.5% do_sync_write 413.00 1.5% select_task_rq_fair 404.00 1.5% _raw_spin_lock_irq 386.00 1.4% update_curr 402.00 1.4% update_curr 385.00 1.4% rw_verify_area 389.00 1.4% do_sync_read 377.00 1.3% _raw_spin_lock_irq 378.00 1.4% vfs_read 369.00 1.3% do_sync_read 340.00 1.2% pipe_iov_copy_from_user 360.00 1.3% vfs_read 316.00 1.1% __wake_up_sync_key * 342.00 1.2% hrtick_start_fair 313.00 1.1% __wake_up_common Signed-off-by: Mike Galbraith <efault@gmx.de> [ fixed !CONFIG_SCHED_HRTICK borkage ] Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/1321971607.6855.17.camel@marge.simson.net Signed-off-by: Ingo Molnar <mingo@elte.hu>
2011-11-22 14:20:07 +00:00
if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
return;
if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
hrtick_start_fair(rq, curr);
}
#else /* !CONFIG_SCHED_HRTICK */
static inline void
hrtick_start_fair(struct rq *rq, struct task_struct *p)
{
}
static inline void hrtick_update(struct rq *rq)
{
}
#endif
sched/fair: Add over-utilization/tipping point indicator Energy-aware scheduling is only meant to be active while the system is _not_ over-utilized. That is, there are spare cycles available to shift tasks around based on their actual utilization to get a more energy-efficient task distribution without depriving any tasks. When above the tipping point task placement is done the traditional way based on load_avg, spreading the tasks across as many cpus as possible based on priority scaled load to preserve smp_nice. Below the tipping point we want to use util_avg instead. We need to define a criteria for when we make the switch. The util_avg for each cpu converges towards 100% regardless of how many additional tasks we may put on it. If we define over-utilized as: sum_{cpus}(rq.cfs.avg.util_avg) + margin > sum_{cpus}(rq.capacity) some individual cpus may be over-utilized running multiple tasks even when the above condition is false. That should be okay as long as we try to spread the tasks out to avoid per-cpu over-utilization as much as possible and if all tasks have the _same_ priority. If the latter isn't true, we have to consider priority to preserve smp_nice. For example, we could have n_cpus nice=-10 util_avg=55% tasks and n_cpus/2 nice=0 util_avg=60% tasks. Balancing based on util_avg we are likely to end up with nice=-10 tasks sharing cpus and nice=0 tasks getting their own as we 1.5*n_cpus tasks in total and 55%+55% is less over-utilized than 55%+60% for those cpus that have to be shared. The system utilization is only 85% of the system capacity, but we are breaking smp_nice. To be sure not to break smp_nice, we have defined over-utilization conservatively as when any cpu in the system is fully utilized at its highest frequency instead: cpu_rq(any).cfs.avg.util_avg + margin > cpu_rq(any).capacity IOW, as soon as one cpu is (nearly) 100% utilized, we switch to load_avg to factor in priority to preserve smp_nice. With this definition, we can skip periodic load-balance as no cpu has an always-running task when the system is not over-utilized. All tasks will be periodic and we can balance them at wake-up. This conservative condition does however mean that some scenarios that could benefit from energy-aware decisions even if one cpu is fully utilized would not get those benefits. For systems where some cpus might have reduced capacity on some cpus (RT-pressure and/or big.LITTLE), we want periodic load-balance checks as soon a just a single cpu is fully utilized as it might one of those with reduced capacity and in that case we want to migrate it. [ peterz: Added a comment explaining why new tasks are not accounted during overutilization detection. ] Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Quentin Perret <quentin.perret@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: adharmap@codeaurora.org Cc: chris.redpath@arm.com Cc: currojerez@riseup.net Cc: dietmar.eggemann@arm.com Cc: edubezval@gmail.com Cc: gregkh@linuxfoundation.org Cc: javi.merino@kernel.org Cc: joel@joelfernandes.org Cc: juri.lelli@redhat.com Cc: patrick.bellasi@arm.com Cc: pkondeti@codeaurora.org Cc: rjw@rjwysocki.net Cc: skannan@codeaurora.org Cc: smuckle@google.com Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Cc: viresh.kumar@linaro.org Link: https://lkml.kernel.org/r/20181203095628.11858-13-quentin.perret@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-12-03 09:56:25 +00:00
#ifdef CONFIG_SMP
static inline unsigned long cpu_util(int cpu);
static inline bool cpu_overutilized(int cpu)
{
return !fits_capacity(cpu_util(cpu), capacity_of(cpu));
sched/fair: Add over-utilization/tipping point indicator Energy-aware scheduling is only meant to be active while the system is _not_ over-utilized. That is, there are spare cycles available to shift tasks around based on their actual utilization to get a more energy-efficient task distribution without depriving any tasks. When above the tipping point task placement is done the traditional way based on load_avg, spreading the tasks across as many cpus as possible based on priority scaled load to preserve smp_nice. Below the tipping point we want to use util_avg instead. We need to define a criteria for when we make the switch. The util_avg for each cpu converges towards 100% regardless of how many additional tasks we may put on it. If we define over-utilized as: sum_{cpus}(rq.cfs.avg.util_avg) + margin > sum_{cpus}(rq.capacity) some individual cpus may be over-utilized running multiple tasks even when the above condition is false. That should be okay as long as we try to spread the tasks out to avoid per-cpu over-utilization as much as possible and if all tasks have the _same_ priority. If the latter isn't true, we have to consider priority to preserve smp_nice. For example, we could have n_cpus nice=-10 util_avg=55% tasks and n_cpus/2 nice=0 util_avg=60% tasks. Balancing based on util_avg we are likely to end up with nice=-10 tasks sharing cpus and nice=0 tasks getting their own as we 1.5*n_cpus tasks in total and 55%+55% is less over-utilized than 55%+60% for those cpus that have to be shared. The system utilization is only 85% of the system capacity, but we are breaking smp_nice. To be sure not to break smp_nice, we have defined over-utilization conservatively as when any cpu in the system is fully utilized at its highest frequency instead: cpu_rq(any).cfs.avg.util_avg + margin > cpu_rq(any).capacity IOW, as soon as one cpu is (nearly) 100% utilized, we switch to load_avg to factor in priority to preserve smp_nice. With this definition, we can skip periodic load-balance as no cpu has an always-running task when the system is not over-utilized. All tasks will be periodic and we can balance them at wake-up. This conservative condition does however mean that some scenarios that could benefit from energy-aware decisions even if one cpu is fully utilized would not get those benefits. For systems where some cpus might have reduced capacity on some cpus (RT-pressure and/or big.LITTLE), we want periodic load-balance checks as soon a just a single cpu is fully utilized as it might one of those with reduced capacity and in that case we want to migrate it. [ peterz: Added a comment explaining why new tasks are not accounted during overutilization detection. ] Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Quentin Perret <quentin.perret@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: adharmap@codeaurora.org Cc: chris.redpath@arm.com Cc: currojerez@riseup.net Cc: dietmar.eggemann@arm.com Cc: edubezval@gmail.com Cc: gregkh@linuxfoundation.org Cc: javi.merino@kernel.org Cc: joel@joelfernandes.org Cc: juri.lelli@redhat.com Cc: patrick.bellasi@arm.com Cc: pkondeti@codeaurora.org Cc: rjw@rjwysocki.net Cc: skannan@codeaurora.org Cc: smuckle@google.com Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Cc: viresh.kumar@linaro.org Link: https://lkml.kernel.org/r/20181203095628.11858-13-quentin.perret@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-12-03 09:56:25 +00:00
}
static inline void update_overutilized_status(struct rq *rq)
{
if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu)) {
sched/fair: Add over-utilization/tipping point indicator Energy-aware scheduling is only meant to be active while the system is _not_ over-utilized. That is, there are spare cycles available to shift tasks around based on their actual utilization to get a more energy-efficient task distribution without depriving any tasks. When above the tipping point task placement is done the traditional way based on load_avg, spreading the tasks across as many cpus as possible based on priority scaled load to preserve smp_nice. Below the tipping point we want to use util_avg instead. We need to define a criteria for when we make the switch. The util_avg for each cpu converges towards 100% regardless of how many additional tasks we may put on it. If we define over-utilized as: sum_{cpus}(rq.cfs.avg.util_avg) + margin > sum_{cpus}(rq.capacity) some individual cpus may be over-utilized running multiple tasks even when the above condition is false. That should be okay as long as we try to spread the tasks out to avoid per-cpu over-utilization as much as possible and if all tasks have the _same_ priority. If the latter isn't true, we have to consider priority to preserve smp_nice. For example, we could have n_cpus nice=-10 util_avg=55% tasks and n_cpus/2 nice=0 util_avg=60% tasks. Balancing based on util_avg we are likely to end up with nice=-10 tasks sharing cpus and nice=0 tasks getting their own as we 1.5*n_cpus tasks in total and 55%+55% is less over-utilized than 55%+60% for those cpus that have to be shared. The system utilization is only 85% of the system capacity, but we are breaking smp_nice. To be sure not to break smp_nice, we have defined over-utilization conservatively as when any cpu in the system is fully utilized at its highest frequency instead: cpu_rq(any).cfs.avg.util_avg + margin > cpu_rq(any).capacity IOW, as soon as one cpu is (nearly) 100% utilized, we switch to load_avg to factor in priority to preserve smp_nice. With this definition, we can skip periodic load-balance as no cpu has an always-running task when the system is not over-utilized. All tasks will be periodic and we can balance them at wake-up. This conservative condition does however mean that some scenarios that could benefit from energy-aware decisions even if one cpu is fully utilized would not get those benefits. For systems where some cpus might have reduced capacity on some cpus (RT-pressure and/or big.LITTLE), we want periodic load-balance checks as soon a just a single cpu is fully utilized as it might one of those with reduced capacity and in that case we want to migrate it. [ peterz: Added a comment explaining why new tasks are not accounted during overutilization detection. ] Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Quentin Perret <quentin.perret@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: adharmap@codeaurora.org Cc: chris.redpath@arm.com Cc: currojerez@riseup.net Cc: dietmar.eggemann@arm.com Cc: edubezval@gmail.com Cc: gregkh@linuxfoundation.org Cc: javi.merino@kernel.org Cc: joel@joelfernandes.org Cc: juri.lelli@redhat.com Cc: patrick.bellasi@arm.com Cc: pkondeti@codeaurora.org Cc: rjw@rjwysocki.net Cc: skannan@codeaurora.org Cc: smuckle@google.com Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Cc: viresh.kumar@linaro.org Link: https://lkml.kernel.org/r/20181203095628.11858-13-quentin.perret@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-12-03 09:56:25 +00:00
WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED);
trace_sched_overutilized_tp(rq->rd, SG_OVERUTILIZED);
}
sched/fair: Add over-utilization/tipping point indicator Energy-aware scheduling is only meant to be active while the system is _not_ over-utilized. That is, there are spare cycles available to shift tasks around based on their actual utilization to get a more energy-efficient task distribution without depriving any tasks. When above the tipping point task placement is done the traditional way based on load_avg, spreading the tasks across as many cpus as possible based on priority scaled load to preserve smp_nice. Below the tipping point we want to use util_avg instead. We need to define a criteria for when we make the switch. The util_avg for each cpu converges towards 100% regardless of how many additional tasks we may put on it. If we define over-utilized as: sum_{cpus}(rq.cfs.avg.util_avg) + margin > sum_{cpus}(rq.capacity) some individual cpus may be over-utilized running multiple tasks even when the above condition is false. That should be okay as long as we try to spread the tasks out to avoid per-cpu over-utilization as much as possible and if all tasks have the _same_ priority. If the latter isn't true, we have to consider priority to preserve smp_nice. For example, we could have n_cpus nice=-10 util_avg=55% tasks and n_cpus/2 nice=0 util_avg=60% tasks. Balancing based on util_avg we are likely to end up with nice=-10 tasks sharing cpus and nice=0 tasks getting their own as we 1.5*n_cpus tasks in total and 55%+55% is less over-utilized than 55%+60% for those cpus that have to be shared. The system utilization is only 85% of the system capacity, but we are breaking smp_nice. To be sure not to break smp_nice, we have defined over-utilization conservatively as when any cpu in the system is fully utilized at its highest frequency instead: cpu_rq(any).cfs.avg.util_avg + margin > cpu_rq(any).capacity IOW, as soon as one cpu is (nearly) 100% utilized, we switch to load_avg to factor in priority to preserve smp_nice. With this definition, we can skip periodic load-balance as no cpu has an always-running task when the system is not over-utilized. All tasks will be periodic and we can balance them at wake-up. This conservative condition does however mean that some scenarios that could benefit from energy-aware decisions even if one cpu is fully utilized would not get those benefits. For systems where some cpus might have reduced capacity on some cpus (RT-pressure and/or big.LITTLE), we want periodic load-balance checks as soon a just a single cpu is fully utilized as it might one of those with reduced capacity and in that case we want to migrate it. [ peterz: Added a comment explaining why new tasks are not accounted during overutilization detection. ] Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Quentin Perret <quentin.perret@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: adharmap@codeaurora.org Cc: chris.redpath@arm.com Cc: currojerez@riseup.net Cc: dietmar.eggemann@arm.com Cc: edubezval@gmail.com Cc: gregkh@linuxfoundation.org Cc: javi.merino@kernel.org Cc: joel@joelfernandes.org Cc: juri.lelli@redhat.com Cc: patrick.bellasi@arm.com Cc: pkondeti@codeaurora.org Cc: rjw@rjwysocki.net Cc: skannan@codeaurora.org Cc: smuckle@google.com Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Cc: viresh.kumar@linaro.org Link: https://lkml.kernel.org/r/20181203095628.11858-13-quentin.perret@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-12-03 09:56:25 +00:00
}
#else
static inline void update_overutilized_status(struct rq *rq) { }
#endif
sched/fair: Load balance aggressively for SCHED_IDLE CPUs The fair scheduler performs periodic load balance on every CPU to check if it can pull some tasks from other busy CPUs. The duration of this periodic load balance is set to sd->balance_interval for the idle CPUs and is calculated by multiplying the sd->balance_interval with the sd->busy_factor (set to 32 by default) for the busy CPUs. The multiplication is done for busy CPUs to avoid doing load balance too often and rather spend more time executing actual task. While that is the right thing to do for the CPUs busy with SCHED_OTHER or SCHED_BATCH tasks, it may not be the optimal thing for CPUs running only SCHED_IDLE tasks. With the recent enhancements in the fair scheduler around SCHED_IDLE CPUs, we now prefer to enqueue a newly-woken task to a SCHED_IDLE CPU instead of other busy or idle CPUs. The same reasoning should be applied to the load balancer as well to make it migrate tasks more aggressively to a SCHED_IDLE CPU, as that will reduce the scheduling latency of the migrated (SCHED_OTHER) tasks. This patch makes minimal changes to the fair scheduler to do the next load balance soon after the last non SCHED_IDLE task is dequeued from a runqueue, i.e. making the CPU SCHED_IDLE. Also the sd->busy_factor is ignored while calculating the balance_interval for such CPUs. This is done to avoid delaying the periodic load balance by few hundred milliseconds for SCHED_IDLE CPUs. This is tested on ARM64 Hikey620 platform (octa-core) with the help of rt-app and it is verified, using kernel traces, that the newly SCHED_IDLE CPU does load balancing shortly after it becomes SCHED_IDLE and pulls tasks from other busy CPUs. Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org> Link: https://lkml.kernel.org/r/e485827eb8fe7db0943d6f3f6e0f5a4a70272781.1578471925.git.viresh.kumar@linaro.org
2020-01-08 08:27:04 +00:00
/* Runqueue only has SCHED_IDLE tasks enqueued */
static int sched_idle_rq(struct rq *rq)
{
return unlikely(rq->nr_running == rq->cfs.idle_h_nr_running &&
rq->nr_running);
}
#ifdef CONFIG_SMP
sched/fair: Load balance aggressively for SCHED_IDLE CPUs The fair scheduler performs periodic load balance on every CPU to check if it can pull some tasks from other busy CPUs. The duration of this periodic load balance is set to sd->balance_interval for the idle CPUs and is calculated by multiplying the sd->balance_interval with the sd->busy_factor (set to 32 by default) for the busy CPUs. The multiplication is done for busy CPUs to avoid doing load balance too often and rather spend more time executing actual task. While that is the right thing to do for the CPUs busy with SCHED_OTHER or SCHED_BATCH tasks, it may not be the optimal thing for CPUs running only SCHED_IDLE tasks. With the recent enhancements in the fair scheduler around SCHED_IDLE CPUs, we now prefer to enqueue a newly-woken task to a SCHED_IDLE CPU instead of other busy or idle CPUs. The same reasoning should be applied to the load balancer as well to make it migrate tasks more aggressively to a SCHED_IDLE CPU, as that will reduce the scheduling latency of the migrated (SCHED_OTHER) tasks. This patch makes minimal changes to the fair scheduler to do the next load balance soon after the last non SCHED_IDLE task is dequeued from a runqueue, i.e. making the CPU SCHED_IDLE. Also the sd->busy_factor is ignored while calculating the balance_interval for such CPUs. This is done to avoid delaying the periodic load balance by few hundred milliseconds for SCHED_IDLE CPUs. This is tested on ARM64 Hikey620 platform (octa-core) with the help of rt-app and it is verified, using kernel traces, that the newly SCHED_IDLE CPU does load balancing shortly after it becomes SCHED_IDLE and pulls tasks from other busy CPUs. Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org> Link: https://lkml.kernel.org/r/e485827eb8fe7db0943d6f3f6e0f5a4a70272781.1578471925.git.viresh.kumar@linaro.org
2020-01-08 08:27:04 +00:00
static int sched_idle_cpu(int cpu)
{
return sched_idle_rq(cpu_rq(cpu));
}
#endif
sched/fair: Load balance aggressively for SCHED_IDLE CPUs The fair scheduler performs periodic load balance on every CPU to check if it can pull some tasks from other busy CPUs. The duration of this periodic load balance is set to sd->balance_interval for the idle CPUs and is calculated by multiplying the sd->balance_interval with the sd->busy_factor (set to 32 by default) for the busy CPUs. The multiplication is done for busy CPUs to avoid doing load balance too often and rather spend more time executing actual task. While that is the right thing to do for the CPUs busy with SCHED_OTHER or SCHED_BATCH tasks, it may not be the optimal thing for CPUs running only SCHED_IDLE tasks. With the recent enhancements in the fair scheduler around SCHED_IDLE CPUs, we now prefer to enqueue a newly-woken task to a SCHED_IDLE CPU instead of other busy or idle CPUs. The same reasoning should be applied to the load balancer as well to make it migrate tasks more aggressively to a SCHED_IDLE CPU, as that will reduce the scheduling latency of the migrated (SCHED_OTHER) tasks. This patch makes minimal changes to the fair scheduler to do the next load balance soon after the last non SCHED_IDLE task is dequeued from a runqueue, i.e. making the CPU SCHED_IDLE. Also the sd->busy_factor is ignored while calculating the balance_interval for such CPUs. This is done to avoid delaying the periodic load balance by few hundred milliseconds for SCHED_IDLE CPUs. This is tested on ARM64 Hikey620 platform (octa-core) with the help of rt-app and it is verified, using kernel traces, that the newly SCHED_IDLE CPU does load balancing shortly after it becomes SCHED_IDLE and pulls tasks from other busy CPUs. Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org> Link: https://lkml.kernel.org/r/e485827eb8fe7db0943d6f3f6e0f5a4a70272781.1578471925.git.viresh.kumar@linaro.org
2020-01-08 08:27:04 +00:00
/*
* The enqueue_task method is called before nr_running is
* increased. Here we update the fair scheduling stats and
* then put the task into the rbtree:
*/
static void
enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
{
struct cfs_rq *cfs_rq;
struct sched_entity *se = &p->se;
int idle_h_nr_running = task_has_idle_policy(p);
/*
* The code below (indirectly) updates schedutil which looks at
* the cfs_rq utilization to select a frequency.
* Let's add the task's estimated utilization to the cfs_rq's
* estimated utilization, before we update schedutil.
*/
util_est_enqueue(&rq->cfs, p);
/*
* If in_iowait is set, the code below may not trigger any cpufreq
* utilization updates, so do it here explicitly with the IOWAIT flag
* passed.
*/
if (p->in_iowait)
sched: cpufreq: Allow remote cpufreq callbacks With Android UI and benchmarks the latency of cpufreq response to certain scheduling events can become very critical. Currently, callbacks into cpufreq governors are only made from the scheduler if the target CPU of the event is the same as the current CPU. This means there are certain situations where a target CPU may not run the cpufreq governor for some time. One testcase to show this behavior is where a task starts running on CPU0, then a new task is also spawned on CPU0 by a task on CPU1. If the system is configured such that the new tasks should receive maximum demand initially, this should result in CPU0 increasing frequency immediately. But because of the above mentioned limitation though, this does not occur. This patch updates the scheduler core to call the cpufreq callbacks for remote CPUs as well. The schedutil, ondemand and conservative governors are updated to process cpufreq utilization update hooks called for remote CPUs where the remote CPU is managed by the cpufreq policy of the local CPU. The intel_pstate driver is updated to always reject remote callbacks. This is tested with couple of usecases (Android: hackbench, recentfling, galleryfling, vellamo, Ubuntu: hackbench) on ARM hikey board (64 bit octa-core, single policy). Only galleryfling showed minor improvements, while others didn't had much deviation. The reason being that this patch only targets a corner case, where following are required to be true to improve performance and that doesn't happen too often with these tests: - Task is migrated to another CPU. - The task has high demand, and should take the target CPU to higher OPPs. - And the target CPU doesn't call into the cpufreq governor until the next tick. Based on initial work from Steve Muckle. Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org> Acked-by: Saravana Kannan <skannan@codeaurora.org> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2017-07-28 06:46:38 +00:00
cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
for_each_sched_entity(se) {
if (se->on_rq)
break;
cfs_rq = cfs_rq_of(se);
sched: Remove the cfs_rq dependency from set_task_cpu() In order to remove the cfs_rq dependency from set_task_cpu() we need to ensure the task is cfs_rq invariant for all callsites. The simple approach is to substract cfs_rq->min_vruntime from se->vruntime on dequeue, and add cfs_rq->min_vruntime on enqueue. However, this has the downside of breaking FAIR_SLEEPERS since we loose the old vruntime as we only maintain the relative position. To solve this, we observe that we only migrate runnable tasks, we do this using deactivate_task(.sleep=0) and activate_task(.wakeup=0), therefore we can restrain the min_vruntime invariance to that state. The only other case is wakeup balancing, since we want to maintain the old vruntime we cannot make it relative on dequeue, but since we don't migrate inactive tasks, we can do so right before we activate it again. This is where we need the new pre-wakeup hook, we need to call this while still holding the old rq->lock. We could fold it into ->select_task_rq(), but since that has multiple callsites and would obfuscate the locking requirements, that seems like a fudge. This leaves the fork() case, simply make sure that ->task_fork() leaves the ->vruntime in a relative state. This covers all cases where set_task_cpu() gets called, and ensures it sees a relative vruntime. Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> LKML-Reference: <20091216170518.191697025@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-12-16 17:04:41 +00:00
enqueue_entity(cfs_rq, se, flags);
cfs_rq->h_nr_running++;
cfs_rq->idle_h_nr_running += idle_h_nr_running;
/* end evaluation on encountering a throttled cfs_rq */
if (cfs_rq_throttled(cfs_rq))
goto enqueue_throttle;
sched: Remove the cfs_rq dependency from set_task_cpu() In order to remove the cfs_rq dependency from set_task_cpu() we need to ensure the task is cfs_rq invariant for all callsites. The simple approach is to substract cfs_rq->min_vruntime from se->vruntime on dequeue, and add cfs_rq->min_vruntime on enqueue. However, this has the downside of breaking FAIR_SLEEPERS since we loose the old vruntime as we only maintain the relative position. To solve this, we observe that we only migrate runnable tasks, we do this using deactivate_task(.sleep=0) and activate_task(.wakeup=0), therefore we can restrain the min_vruntime invariance to that state. The only other case is wakeup balancing, since we want to maintain the old vruntime we cannot make it relative on dequeue, but since we don't migrate inactive tasks, we can do so right before we activate it again. This is where we need the new pre-wakeup hook, we need to call this while still holding the old rq->lock. We could fold it into ->select_task_rq(), but since that has multiple callsites and would obfuscate the locking requirements, that seems like a fudge. This leaves the fork() case, simply make sure that ->task_fork() leaves the ->vruntime in a relative state. This covers all cases where set_task_cpu() gets called, and ensures it sees a relative vruntime. Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> LKML-Reference: <20091216170518.191697025@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-12-16 17:04:41 +00:00
flags = ENQUEUE_WAKEUP;
}
for_each_sched_entity(se) {
cfs_rq = cfs_rq_of(se);
update_load_avg(cfs_rq, se, UPDATE_TG);
sched/pelt: Add a new runnable average signal Now that runnable_load_avg has been removed, we can replace it by a new signal that will highlight the runnable pressure on a cfs_rq. This signal track the waiting time of tasks on rq and can help to better define the state of rqs. At now, only util_avg is used to define the state of a rq: A rq with more that around 80% of utilization and more than 1 tasks is considered as overloaded. But the util_avg signal of a rq can become temporaly low after that a task migrated onto another rq which can bias the classification of the rq. When tasks compete for the same rq, their runnable average signal will be higher than util_avg as it will include the waiting time and we can use this signal to better classify cfs_rqs. The new runnable_avg will track the runnable time of a task which simply adds the waiting time to the running time. The runnable _avg of cfs_rq will be the /Sum of se's runnable_avg and the runnable_avg of group entity will follow the one of the rq similarly to util_avg. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: "Dietmar Eggemann <dietmar.eggemann@arm.com>" Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Valentin Schneider <valentin.schneider@arm.com> Cc: Phil Auld <pauld@redhat.com> Cc: Hillf Danton <hdanton@sina.com> Link: https://lore.kernel.org/r/20200224095223.13361-9-mgorman@techsingularity.net
2020-02-24 09:52:18 +00:00
se_update_runnable(se);
update_cfs_group(se);
cfs_rq->h_nr_running++;
cfs_rq->idle_h_nr_running += idle_h_nr_running;
/* end evaluation on encountering a throttled cfs_rq */
if (cfs_rq_throttled(cfs_rq))
goto enqueue_throttle;
}
enqueue_throttle:
sched/fair: Add over-utilization/tipping point indicator Energy-aware scheduling is only meant to be active while the system is _not_ over-utilized. That is, there are spare cycles available to shift tasks around based on their actual utilization to get a more energy-efficient task distribution without depriving any tasks. When above the tipping point task placement is done the traditional way based on load_avg, spreading the tasks across as many cpus as possible based on priority scaled load to preserve smp_nice. Below the tipping point we want to use util_avg instead. We need to define a criteria for when we make the switch. The util_avg for each cpu converges towards 100% regardless of how many additional tasks we may put on it. If we define over-utilized as: sum_{cpus}(rq.cfs.avg.util_avg) + margin > sum_{cpus}(rq.capacity) some individual cpus may be over-utilized running multiple tasks even when the above condition is false. That should be okay as long as we try to spread the tasks out to avoid per-cpu over-utilization as much as possible and if all tasks have the _same_ priority. If the latter isn't true, we have to consider priority to preserve smp_nice. For example, we could have n_cpus nice=-10 util_avg=55% tasks and n_cpus/2 nice=0 util_avg=60% tasks. Balancing based on util_avg we are likely to end up with nice=-10 tasks sharing cpus and nice=0 tasks getting their own as we 1.5*n_cpus tasks in total and 55%+55% is less over-utilized than 55%+60% for those cpus that have to be shared. The system utilization is only 85% of the system capacity, but we are breaking smp_nice. To be sure not to break smp_nice, we have defined over-utilization conservatively as when any cpu in the system is fully utilized at its highest frequency instead: cpu_rq(any).cfs.avg.util_avg + margin > cpu_rq(any).capacity IOW, as soon as one cpu is (nearly) 100% utilized, we switch to load_avg to factor in priority to preserve smp_nice. With this definition, we can skip periodic load-balance as no cpu has an always-running task when the system is not over-utilized. All tasks will be periodic and we can balance them at wake-up. This conservative condition does however mean that some scenarios that could benefit from energy-aware decisions even if one cpu is fully utilized would not get those benefits. For systems where some cpus might have reduced capacity on some cpus (RT-pressure and/or big.LITTLE), we want periodic load-balance checks as soon a just a single cpu is fully utilized as it might one of those with reduced capacity and in that case we want to migrate it. [ peterz: Added a comment explaining why new tasks are not accounted during overutilization detection. ] Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Quentin Perret <quentin.perret@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: adharmap@codeaurora.org Cc: chris.redpath@arm.com Cc: currojerez@riseup.net Cc: dietmar.eggemann@arm.com Cc: edubezval@gmail.com Cc: gregkh@linuxfoundation.org Cc: javi.merino@kernel.org Cc: joel@joelfernandes.org Cc: juri.lelli@redhat.com Cc: patrick.bellasi@arm.com Cc: pkondeti@codeaurora.org Cc: rjw@rjwysocki.net Cc: skannan@codeaurora.org Cc: smuckle@google.com Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Cc: viresh.kumar@linaro.org Link: https://lkml.kernel.org/r/20181203095628.11858-13-quentin.perret@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-12-03 09:56:25 +00:00
if (!se) {
add_nr_running(rq, 1);
sched/fair: Add over-utilization/tipping point indicator Energy-aware scheduling is only meant to be active while the system is _not_ over-utilized. That is, there are spare cycles available to shift tasks around based on their actual utilization to get a more energy-efficient task distribution without depriving any tasks. When above the tipping point task placement is done the traditional way based on load_avg, spreading the tasks across as many cpus as possible based on priority scaled load to preserve smp_nice. Below the tipping point we want to use util_avg instead. We need to define a criteria for when we make the switch. The util_avg for each cpu converges towards 100% regardless of how many additional tasks we may put on it. If we define over-utilized as: sum_{cpus}(rq.cfs.avg.util_avg) + margin > sum_{cpus}(rq.capacity) some individual cpus may be over-utilized running multiple tasks even when the above condition is false. That should be okay as long as we try to spread the tasks out to avoid per-cpu over-utilization as much as possible and if all tasks have the _same_ priority. If the latter isn't true, we have to consider priority to preserve smp_nice. For example, we could have n_cpus nice=-10 util_avg=55% tasks and n_cpus/2 nice=0 util_avg=60% tasks. Balancing based on util_avg we are likely to end up with nice=-10 tasks sharing cpus and nice=0 tasks getting their own as we 1.5*n_cpus tasks in total and 55%+55% is less over-utilized than 55%+60% for those cpus that have to be shared. The system utilization is only 85% of the system capacity, but we are breaking smp_nice. To be sure not to break smp_nice, we have defined over-utilization conservatively as when any cpu in the system is fully utilized at its highest frequency instead: cpu_rq(any).cfs.avg.util_avg + margin > cpu_rq(any).capacity IOW, as soon as one cpu is (nearly) 100% utilized, we switch to load_avg to factor in priority to preserve smp_nice. With this definition, we can skip periodic load-balance as no cpu has an always-running task when the system is not over-utilized. All tasks will be periodic and we can balance them at wake-up. This conservative condition does however mean that some scenarios that could benefit from energy-aware decisions even if one cpu is fully utilized would not get those benefits. For systems where some cpus might have reduced capacity on some cpus (RT-pressure and/or big.LITTLE), we want periodic load-balance checks as soon a just a single cpu is fully utilized as it might one of those with reduced capacity and in that case we want to migrate it. [ peterz: Added a comment explaining why new tasks are not accounted during overutilization detection. ] Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Quentin Perret <quentin.perret@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: adharmap@codeaurora.org Cc: chris.redpath@arm.com Cc: currojerez@riseup.net Cc: dietmar.eggemann@arm.com Cc: edubezval@gmail.com Cc: gregkh@linuxfoundation.org Cc: javi.merino@kernel.org Cc: joel@joelfernandes.org Cc: juri.lelli@redhat.com Cc: patrick.bellasi@arm.com Cc: pkondeti@codeaurora.org Cc: rjw@rjwysocki.net Cc: skannan@codeaurora.org Cc: smuckle@google.com Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Cc: viresh.kumar@linaro.org Link: https://lkml.kernel.org/r/20181203095628.11858-13-quentin.perret@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-12-03 09:56:25 +00:00
/*
* Since new tasks are assigned an initial util_avg equal to
* half of the spare capacity of their CPU, tiny tasks have the
* ability to cross the overutilized threshold, which will
* result in the load balancer ruining all the task placement
* done by EAS. As a way to mitigate that effect, do not account
* for the first enqueue operation of new tasks during the
* overutilized flag detection.
*
* A better way of solving this problem would be to wait for
* the PELT signals of tasks to converge before taking them
* into account, but that is not straightforward to implement,
* and the following generally works well enough in practice.
*/
if (flags & ENQUEUE_WAKEUP)
update_overutilized_status(rq);
}
sched/fair: Fix insertion in rq->leaf_cfs_rq_list Sargun reported a crash: "I picked up c40f7d74c741a907cfaeb73a7697081881c497d0 sched/fair: Fix infinite loop in update_blocked_averages() by reverting a9e7f6544b9c and put it on top of 4.19.13. In addition to this, I uninlined list_add_leaf_cfs_rq for debugging. This revealed a new bug that we didn't get to because we kept getting crashes from the previous issue. When we are running with cgroups that are rapidly changing, with CFS bandwidth control, and in addition using the cpusets cgroup, we see this crash. Specifically, it seems to occur with cgroups that are throttled and we change the allowed cpuset." The algorithm used to order cfs_rq in rq->leaf_cfs_rq_list assumes that it will walk down to root the 1st time a cfs_rq is used and we will finish to add either a cfs_rq without parent or a cfs_rq with a parent that is already on the list. But this is not always true in presence of throttling. Because a cfs_rq can be throttled even if it has never been used but other CPUs of the cgroup have already used all the bandwdith, we are not sure to go down to the root and add all cfs_rq in the list. Ensure that all cfs_rq will be added in the list even if they are throttled. [ mingo: Fix !CGROUPS build. ] Reported-by: Sargun Dhillon <sargun@sargun.me> Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: tj@kernel.org Fixes: 9c2791f936ef ("Fix hierarchical order in rq->leaf_cfs_rq_list") Link: https://lkml.kernel.org/r/1548825767-10799-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-01-30 05:22:47 +00:00
if (cfs_bandwidth_used()) {
/*
* When bandwidth control is enabled; the cfs_rq_throttled()
* breaks in the above iteration can result in incomplete
* leaf list maintenance, resulting in triggering the assertion
* below.
*/
for_each_sched_entity(se) {
cfs_rq = cfs_rq_of(se);
if (list_add_leaf_cfs_rq(cfs_rq))
break;
}
}
assert_list_leaf_cfs_rq(rq);
hrtick_update(rq);
}
sched: Next buddy hint on sleep and preempt path When a task in a taskgroup sleeps, pick_next_task starts all the way back at the root and picks the task/taskgroup with the min vruntime across all runnable tasks. But when there are many frequently sleeping tasks across different taskgroups, it makes better sense to stay with same taskgroup for its slice period (or until all tasks in the taskgroup sleeps) instead of switching cross taskgroup on each sleep after a short runtime. This helps specifically where taskgroups corresponds to a process with multiple threads. The change reduces the number of CR3 switches in this case. Example: Two taskgroups with 2 threads each which are running for 2ms and sleeping for 1ms. Looking at sched:sched_switch shows: BEFORE: taskgroup_1 threads [5004, 5005], taskgroup_2 threads [5016, 5017] cpu-soaker-5004 [003] 3683.391089 cpu-soaker-5016 [003] 3683.393106 cpu-soaker-5005 [003] 3683.395119 cpu-soaker-5017 [003] 3683.397130 cpu-soaker-5004 [003] 3683.399143 cpu-soaker-5016 [003] 3683.401155 cpu-soaker-5005 [003] 3683.403168 cpu-soaker-5017 [003] 3683.405170 AFTER: taskgroup_1 threads [21890, 21891], taskgroup_2 threads [21934, 21935] cpu-soaker-21890 [003] 865.895494 cpu-soaker-21935 [003] 865.897506 cpu-soaker-21934 [003] 865.899520 cpu-soaker-21935 [003] 865.901532 cpu-soaker-21934 [003] 865.903543 cpu-soaker-21935 [003] 865.905546 cpu-soaker-21891 [003] 865.907548 cpu-soaker-21890 [003] 865.909560 cpu-soaker-21891 [003] 865.911571 cpu-soaker-21890 [003] 865.913582 cpu-soaker-21891 [003] 865.915594 cpu-soaker-21934 [003] 865.917606 Similar problem is there when there are multiple taskgroups and say a task A preempts currently running task B of taskgroup_1. On schedule, pick_next_task can pick an unrelated task on taskgroup_2. Here it would be better to give some preference to task B on pick_next_task. A simple (may be extreme case) benchmark I tried was tbench with 2 tbench client processes with 2 threads each running on a single CPU. Avg throughput across 5 50 sec runs was: BEFORE: 105.84 MB/sec AFTER: 112.42 MB/sec Signed-off-by: Venkatesh Pallipadi <venki@google.com> Acked-by: Rik van Riel <riel@redhat.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/1302802253-25760-1-git-send-email-venki@google.com Signed-off-by: Ingo Molnar <mingo@elte.hu>
2011-04-14 17:30:53 +00:00
static void set_next_buddy(struct sched_entity *se);
/*
* The dequeue_task method is called before nr_running is
* decreased. We remove the task from the rbtree and
* update the fair scheduling stats:
*/
static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
{
struct cfs_rq *cfs_rq;
struct sched_entity *se = &p->se;
sched: Next buddy hint on sleep and preempt path When a task in a taskgroup sleeps, pick_next_task starts all the way back at the root and picks the task/taskgroup with the min vruntime across all runnable tasks. But when there are many frequently sleeping tasks across different taskgroups, it makes better sense to stay with same taskgroup for its slice period (or until all tasks in the taskgroup sleeps) instead of switching cross taskgroup on each sleep after a short runtime. This helps specifically where taskgroups corresponds to a process with multiple threads. The change reduces the number of CR3 switches in this case. Example: Two taskgroups with 2 threads each which are running for 2ms and sleeping for 1ms. Looking at sched:sched_switch shows: BEFORE: taskgroup_1 threads [5004, 5005], taskgroup_2 threads [5016, 5017] cpu-soaker-5004 [003] 3683.391089 cpu-soaker-5016 [003] 3683.393106 cpu-soaker-5005 [003] 3683.395119 cpu-soaker-5017 [003] 3683.397130 cpu-soaker-5004 [003] 3683.399143 cpu-soaker-5016 [003] 3683.401155 cpu-soaker-5005 [003] 3683.403168 cpu-soaker-5017 [003] 3683.405170 AFTER: taskgroup_1 threads [21890, 21891], taskgroup_2 threads [21934, 21935] cpu-soaker-21890 [003] 865.895494 cpu-soaker-21935 [003] 865.897506 cpu-soaker-21934 [003] 865.899520 cpu-soaker-21935 [003] 865.901532 cpu-soaker-21934 [003] 865.903543 cpu-soaker-21935 [003] 865.905546 cpu-soaker-21891 [003] 865.907548 cpu-soaker-21890 [003] 865.909560 cpu-soaker-21891 [003] 865.911571 cpu-soaker-21890 [003] 865.913582 cpu-soaker-21891 [003] 865.915594 cpu-soaker-21934 [003] 865.917606 Similar problem is there when there are multiple taskgroups and say a task A preempts currently running task B of taskgroup_1. On schedule, pick_next_task can pick an unrelated task on taskgroup_2. Here it would be better to give some preference to task B on pick_next_task. A simple (may be extreme case) benchmark I tried was tbench with 2 tbench client processes with 2 threads each running on a single CPU. Avg throughput across 5 50 sec runs was: BEFORE: 105.84 MB/sec AFTER: 112.42 MB/sec Signed-off-by: Venkatesh Pallipadi <venki@google.com> Acked-by: Rik van Riel <riel@redhat.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/1302802253-25760-1-git-send-email-venki@google.com Signed-off-by: Ingo Molnar <mingo@elte.hu>
2011-04-14 17:30:53 +00:00
int task_sleep = flags & DEQUEUE_SLEEP;
int idle_h_nr_running = task_has_idle_policy(p);
sched/fair: Load balance aggressively for SCHED_IDLE CPUs The fair scheduler performs periodic load balance on every CPU to check if it can pull some tasks from other busy CPUs. The duration of this periodic load balance is set to sd->balance_interval for the idle CPUs and is calculated by multiplying the sd->balance_interval with the sd->busy_factor (set to 32 by default) for the busy CPUs. The multiplication is done for busy CPUs to avoid doing load balance too often and rather spend more time executing actual task. While that is the right thing to do for the CPUs busy with SCHED_OTHER or SCHED_BATCH tasks, it may not be the optimal thing for CPUs running only SCHED_IDLE tasks. With the recent enhancements in the fair scheduler around SCHED_IDLE CPUs, we now prefer to enqueue a newly-woken task to a SCHED_IDLE CPU instead of other busy or idle CPUs. The same reasoning should be applied to the load balancer as well to make it migrate tasks more aggressively to a SCHED_IDLE CPU, as that will reduce the scheduling latency of the migrated (SCHED_OTHER) tasks. This patch makes minimal changes to the fair scheduler to do the next load balance soon after the last non SCHED_IDLE task is dequeued from a runqueue, i.e. making the CPU SCHED_IDLE. Also the sd->busy_factor is ignored while calculating the balance_interval for such CPUs. This is done to avoid delaying the periodic load balance by few hundred milliseconds for SCHED_IDLE CPUs. This is tested on ARM64 Hikey620 platform (octa-core) with the help of rt-app and it is verified, using kernel traces, that the newly SCHED_IDLE CPU does load balancing shortly after it becomes SCHED_IDLE and pulls tasks from other busy CPUs. Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org> Link: https://lkml.kernel.org/r/e485827eb8fe7db0943d6f3f6e0f5a4a70272781.1578471925.git.viresh.kumar@linaro.org
2020-01-08 08:27:04 +00:00
bool was_sched_idle = sched_idle_rq(rq);
for_each_sched_entity(se) {
cfs_rq = cfs_rq_of(se);
dequeue_entity(cfs_rq, se, flags);
cfs_rq->h_nr_running--;
cfs_rq->idle_h_nr_running -= idle_h_nr_running;
/* end evaluation on encountering a throttled cfs_rq */
if (cfs_rq_throttled(cfs_rq))
goto dequeue_throttle;
/* Don't dequeue parent if it has other entities besides us */
sched: Next buddy hint on sleep and preempt path When a task in a taskgroup sleeps, pick_next_task starts all the way back at the root and picks the task/taskgroup with the min vruntime across all runnable tasks. But when there are many frequently sleeping tasks across different taskgroups, it makes better sense to stay with same taskgroup for its slice period (or until all tasks in the taskgroup sleeps) instead of switching cross taskgroup on each sleep after a short runtime. This helps specifically where taskgroups corresponds to a process with multiple threads. The change reduces the number of CR3 switches in this case. Example: Two taskgroups with 2 threads each which are running for 2ms and sleeping for 1ms. Looking at sched:sched_switch shows: BEFORE: taskgroup_1 threads [5004, 5005], taskgroup_2 threads [5016, 5017] cpu-soaker-5004 [003] 3683.391089 cpu-soaker-5016 [003] 3683.393106 cpu-soaker-5005 [003] 3683.395119 cpu-soaker-5017 [003] 3683.397130 cpu-soaker-5004 [003] 3683.399143 cpu-soaker-5016 [003] 3683.401155 cpu-soaker-5005 [003] 3683.403168 cpu-soaker-5017 [003] 3683.405170 AFTER: taskgroup_1 threads [21890, 21891], taskgroup_2 threads [21934, 21935] cpu-soaker-21890 [003] 865.895494 cpu-soaker-21935 [003] 865.897506 cpu-soaker-21934 [003] 865.899520 cpu-soaker-21935 [003] 865.901532 cpu-soaker-21934 [003] 865.903543 cpu-soaker-21935 [003] 865.905546 cpu-soaker-21891 [003] 865.907548 cpu-soaker-21890 [003] 865.909560 cpu-soaker-21891 [003] 865.911571 cpu-soaker-21890 [003] 865.913582 cpu-soaker-21891 [003] 865.915594 cpu-soaker-21934 [003] 865.917606 Similar problem is there when there are multiple taskgroups and say a task A preempts currently running task B of taskgroup_1. On schedule, pick_next_task can pick an unrelated task on taskgroup_2. Here it would be better to give some preference to task B on pick_next_task. A simple (may be extreme case) benchmark I tried was tbench with 2 tbench client processes with 2 threads each running on a single CPU. Avg throughput across 5 50 sec runs was: BEFORE: 105.84 MB/sec AFTER: 112.42 MB/sec Signed-off-by: Venkatesh Pallipadi <venki@google.com> Acked-by: Rik van Riel <riel@redhat.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/1302802253-25760-1-git-send-email-venki@google.com Signed-off-by: Ingo Molnar <mingo@elte.hu>
2011-04-14 17:30:53 +00:00
if (cfs_rq->load.weight) {
/* Avoid re-evaluating load for this entity: */
se = parent_entity(se);
sched: Next buddy hint on sleep and preempt path When a task in a taskgroup sleeps, pick_next_task starts all the way back at the root and picks the task/taskgroup with the min vruntime across all runnable tasks. But when there are many frequently sleeping tasks across different taskgroups, it makes better sense to stay with same taskgroup for its slice period (or until all tasks in the taskgroup sleeps) instead of switching cross taskgroup on each sleep after a short runtime. This helps specifically where taskgroups corresponds to a process with multiple threads. The change reduces the number of CR3 switches in this case. Example: Two taskgroups with 2 threads each which are running for 2ms and sleeping for 1ms. Looking at sched:sched_switch shows: BEFORE: taskgroup_1 threads [5004, 5005], taskgroup_2 threads [5016, 5017] cpu-soaker-5004 [003] 3683.391089 cpu-soaker-5016 [003] 3683.393106 cpu-soaker-5005 [003] 3683.395119 cpu-soaker-5017 [003] 3683.397130 cpu-soaker-5004 [003] 3683.399143 cpu-soaker-5016 [003] 3683.401155 cpu-soaker-5005 [003] 3683.403168 cpu-soaker-5017 [003] 3683.405170 AFTER: taskgroup_1 threads [21890, 21891], taskgroup_2 threads [21934, 21935] cpu-soaker-21890 [003] 865.895494 cpu-soaker-21935 [003] 865.897506 cpu-soaker-21934 [003] 865.899520 cpu-soaker-21935 [003] 865.901532 cpu-soaker-21934 [003] 865.903543 cpu-soaker-21935 [003] 865.905546 cpu-soaker-21891 [003] 865.907548 cpu-soaker-21890 [003] 865.909560 cpu-soaker-21891 [003] 865.911571 cpu-soaker-21890 [003] 865.913582 cpu-soaker-21891 [003] 865.915594 cpu-soaker-21934 [003] 865.917606 Similar problem is there when there are multiple taskgroups and say a task A preempts currently running task B of taskgroup_1. On schedule, pick_next_task can pick an unrelated task on taskgroup_2. Here it would be better to give some preference to task B on pick_next_task. A simple (may be extreme case) benchmark I tried was tbench with 2 tbench client processes with 2 threads each running on a single CPU. Avg throughput across 5 50 sec runs was: BEFORE: 105.84 MB/sec AFTER: 112.42 MB/sec Signed-off-by: Venkatesh Pallipadi <venki@google.com> Acked-by: Rik van Riel <riel@redhat.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/1302802253-25760-1-git-send-email-venki@google.com Signed-off-by: Ingo Molnar <mingo@elte.hu>
2011-04-14 17:30:53 +00:00
/*
* Bias pick_next to pick a task from this cfs_rq, as
* p is sleeping when it is within its sched_slice.
*/
if (task_sleep && se && !throttled_hierarchy(cfs_rq))
set_next_buddy(se);
break;
sched: Next buddy hint on sleep and preempt path When a task in a taskgroup sleeps, pick_next_task starts all the way back at the root and picks the task/taskgroup with the min vruntime across all runnable tasks. But when there are many frequently sleeping tasks across different taskgroups, it makes better sense to stay with same taskgroup for its slice period (or until all tasks in the taskgroup sleeps) instead of switching cross taskgroup on each sleep after a short runtime. This helps specifically where taskgroups corresponds to a process with multiple threads. The change reduces the number of CR3 switches in this case. Example: Two taskgroups with 2 threads each which are running for 2ms and sleeping for 1ms. Looking at sched:sched_switch shows: BEFORE: taskgroup_1 threads [5004, 5005], taskgroup_2 threads [5016, 5017] cpu-soaker-5004 [003] 3683.391089 cpu-soaker-5016 [003] 3683.393106 cpu-soaker-5005 [003] 3683.395119 cpu-soaker-5017 [003] 3683.397130 cpu-soaker-5004 [003] 3683.399143 cpu-soaker-5016 [003] 3683.401155 cpu-soaker-5005 [003] 3683.403168 cpu-soaker-5017 [003] 3683.405170 AFTER: taskgroup_1 threads [21890, 21891], taskgroup_2 threads [21934, 21935] cpu-soaker-21890 [003] 865.895494 cpu-soaker-21935 [003] 865.897506 cpu-soaker-21934 [003] 865.899520 cpu-soaker-21935 [003] 865.901532 cpu-soaker-21934 [003] 865.903543 cpu-soaker-21935 [003] 865.905546 cpu-soaker-21891 [003] 865.907548 cpu-soaker-21890 [003] 865.909560 cpu-soaker-21891 [003] 865.911571 cpu-soaker-21890 [003] 865.913582 cpu-soaker-21891 [003] 865.915594 cpu-soaker-21934 [003] 865.917606 Similar problem is there when there are multiple taskgroups and say a task A preempts currently running task B of taskgroup_1. On schedule, pick_next_task can pick an unrelated task on taskgroup_2. Here it would be better to give some preference to task B on pick_next_task. A simple (may be extreme case) benchmark I tried was tbench with 2 tbench client processes with 2 threads each running on a single CPU. Avg throughput across 5 50 sec runs was: BEFORE: 105.84 MB/sec AFTER: 112.42 MB/sec Signed-off-by: Venkatesh Pallipadi <venki@google.com> Acked-by: Rik van Riel <riel@redhat.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/1302802253-25760-1-git-send-email-venki@google.com Signed-off-by: Ingo Molnar <mingo@elte.hu>
2011-04-14 17:30:53 +00:00
}
flags |= DEQUEUE_SLEEP;
}
for_each_sched_entity(se) {
cfs_rq = cfs_rq_of(se);
update_load_avg(cfs_rq, se, UPDATE_TG);
sched/pelt: Add a new runnable average signal Now that runnable_load_avg has been removed, we can replace it by a new signal that will highlight the runnable pressure on a cfs_rq. This signal track the waiting time of tasks on rq and can help to better define the state of rqs. At now, only util_avg is used to define the state of a rq: A rq with more that around 80% of utilization and more than 1 tasks is considered as overloaded. But the util_avg signal of a rq can become temporaly low after that a task migrated onto another rq which can bias the classification of the rq. When tasks compete for the same rq, their runnable average signal will be higher than util_avg as it will include the waiting time and we can use this signal to better classify cfs_rqs. The new runnable_avg will track the runnable time of a task which simply adds the waiting time to the running time. The runnable _avg of cfs_rq will be the /Sum of se's runnable_avg and the runnable_avg of group entity will follow the one of the rq similarly to util_avg. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: "Dietmar Eggemann <dietmar.eggemann@arm.com>" Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Valentin Schneider <valentin.schneider@arm.com> Cc: Phil Auld <pauld@redhat.com> Cc: Hillf Danton <hdanton@sina.com> Link: https://lore.kernel.org/r/20200224095223.13361-9-mgorman@techsingularity.net
2020-02-24 09:52:18 +00:00
se_update_runnable(se);
update_cfs_group(se);
cfs_rq->h_nr_running--;
cfs_rq->idle_h_nr_running -= idle_h_nr_running;
/* end evaluation on encountering a throttled cfs_rq */
if (cfs_rq_throttled(cfs_rq))
goto dequeue_throttle;
}
dequeue_throttle:
if (!se)
sub_nr_running(rq, 1);
sched/fair: Load balance aggressively for SCHED_IDLE CPUs The fair scheduler performs periodic load balance on every CPU to check if it can pull some tasks from other busy CPUs. The duration of this periodic load balance is set to sd->balance_interval for the idle CPUs and is calculated by multiplying the sd->balance_interval with the sd->busy_factor (set to 32 by default) for the busy CPUs. The multiplication is done for busy CPUs to avoid doing load balance too often and rather spend more time executing actual task. While that is the right thing to do for the CPUs busy with SCHED_OTHER or SCHED_BATCH tasks, it may not be the optimal thing for CPUs running only SCHED_IDLE tasks. With the recent enhancements in the fair scheduler around SCHED_IDLE CPUs, we now prefer to enqueue a newly-woken task to a SCHED_IDLE CPU instead of other busy or idle CPUs. The same reasoning should be applied to the load balancer as well to make it migrate tasks more aggressively to a SCHED_IDLE CPU, as that will reduce the scheduling latency of the migrated (SCHED_OTHER) tasks. This patch makes minimal changes to the fair scheduler to do the next load balance soon after the last non SCHED_IDLE task is dequeued from a runqueue, i.e. making the CPU SCHED_IDLE. Also the sd->busy_factor is ignored while calculating the balance_interval for such CPUs. This is done to avoid delaying the periodic load balance by few hundred milliseconds for SCHED_IDLE CPUs. This is tested on ARM64 Hikey620 platform (octa-core) with the help of rt-app and it is verified, using kernel traces, that the newly SCHED_IDLE CPU does load balancing shortly after it becomes SCHED_IDLE and pulls tasks from other busy CPUs. Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org> Link: https://lkml.kernel.org/r/e485827eb8fe7db0943d6f3f6e0f5a4a70272781.1578471925.git.viresh.kumar@linaro.org
2020-01-08 08:27:04 +00:00
/* balance early to pull high priority tasks */
if (unlikely(!was_sched_idle && sched_idle_rq(rq)))
rq->next_balance = jiffies;
sched/fair: Add util_est on top of PELT The util_avg signal computed by PELT is too variable for some use-cases. For example, a big task waking up after a long sleep period will have its utilization almost completely decayed. This introduces some latency before schedutil will be able to pick the best frequency to run a task. The same issue can affect task placement. Indeed, since the task utilization is already decayed at wakeup, when the task is enqueued in a CPU, this can result in a CPU running a big task as being temporarily represented as being almost empty. This leads to a race condition where other tasks can be potentially allocated on a CPU which just started to run a big task which slept for a relatively long period. Moreover, the PELT utilization of a task can be updated every [ms], thus making it a continuously changing value for certain longer running tasks. This means that the instantaneous PELT utilization of a RUNNING task is not really meaningful to properly support scheduler decisions. For all these reasons, a more stable signal can do a better job of representing the expected/estimated utilization of a task/cfs_rq. Such a signal can be easily created on top of PELT by still using it as an estimator which produces values to be aggregated on meaningful events. This patch adds a simple implementation of util_est, a new signal built on top of PELT's util_avg where: util_est(task) = max(task::util_avg, f(task::util_avg@dequeue)) This allows to remember how big a task has been reported by PELT in its previous activations via f(task::util_avg@dequeue), which is the new _task_util_est(struct task_struct*) function added by this patch. If a task should change its behavior and it runs longer in a new activation, after a certain time its util_est will just track the original PELT signal (i.e. task::util_avg). The estimated utilization of cfs_rq is defined only for root ones. That's because the only sensible consumer of this signal are the scheduler and schedutil when looking for the overall CPU utilization due to FAIR tasks. For this reason, the estimated utilization of a root cfs_rq is simply defined as: util_est(cfs_rq) = max(cfs_rq::util_avg, cfs_rq::util_est::enqueued) where: cfs_rq::util_est::enqueued = sum(_task_util_est(task)) for each RUNNABLE task on that root cfs_rq It's worth noting that the estimated utilization is tracked only for objects of interests, specifically: - Tasks: to better support tasks placement decisions - root cfs_rqs: to better support both tasks placement decisions as well as frequencies selection Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Joel Fernandes <joelaf@google.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Rafael J . Wysocki <rafael.j.wysocki@intel.com> Cc: Steve Muckle <smuckle@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Todd Kjos <tkjos@android.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Link: http://lkml.kernel.org/r/20180309095245.11071-2-patrick.bellasi@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-03-09 09:52:42 +00:00
util_est_dequeue(&rq->cfs, p, task_sleep);
hrtick_update(rq);
}
#ifdef CONFIG_SMP
sched/core: Rewrite and improve select_idle_siblings() select_idle_siblings() is a known pain point for a number of workloads; it either does too much or not enough and sometimes just does plain wrong. This rewrite attempts to address a number of issues (but sadly not all). The current code does an unconditional sched_domain iteration; with the intent of finding an idle core (on SMT hardware). The problems which this patch tries to address are: - its pointless to look for idle cores if the machine is real busy; at which point you're just wasting cycles. - it's behaviour is inconsistent between SMT and !SMT hardware in that !SMT hardware ends up doing a scan for any idle CPU in the LLC domain, while SMT hardware does a scan for idle cores and if that fails, falls back to a scan for idle threads on the 'target' core. The new code replaces the sched_domain scan with 3 explicit scans: 1) search for an idle core in the LLC 2) search for an idle CPU in the LLC 3) search for an idle thread in the 'target' core where 1 and 3 are conditional on SMT support and 1 and 2 have runtime heuristics to skip the step. Step 1) is conditional on sd_llc_shared->has_idle_cores; when a cpu goes idle and sd_llc_shared->has_idle_cores is false, we scan all SMT siblings of the CPU going idle. Similarly, we clear sd_llc_shared->has_idle_cores when we fail to find an idle core. Step 2) tracks the average cost of the scan and compares this to the average idle time guestimate for the CPU doing the wakeup. There is a significant fudge factor involved to deal with the variability of the averages. Esp. hackbench was sensitive to this. Step 3) is unconditional; we assume (also per step 1) that scanning all SMT siblings in a core is 'cheap'. With this; SMT systems gain step 2, which cures a few benchmarks -- notably one from Facebook. One 'feature' of the sched_domain iteration, which we preserve in the new code, is that it would start scanning from the 'target' CPU, instead of scanning the cpumask in cpu id order. This avoids multiple CPUs in the LLC scanning for idle to gang up and find the same CPU quite as much. The down side is that tasks can end up hopping across the LLC for no apparent reason. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-05-09 08:38:05 +00:00
/* Working cpumask for: load_balance, load_balance_newidle. */
DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
#ifdef CONFIG_NO_HZ_COMMON
static struct {
cpumask_var_t idle_cpus_mask;
atomic_t nr_cpus;
int has_blocked; /* Idle CPUS has blocked load */
unsigned long next_balance; /* in jiffy units */
unsigned long next_blocked; /* Next update of blocked load in jiffies */
} nohz ____cacheline_aligned;
#endif /* CONFIG_NO_HZ_COMMON */
static unsigned long cpu_load(struct rq *rq)
{
return cfs_rq_load_avg(&rq->cfs);
}
sched/fair: Fix rework of find_idlest_group() The task, for which the scheduler looks for the idlest group of CPUs, must be discounted from all statistics in order to get a fair comparison between groups. This includes utilization, load, nr_running and idle_cpus. Such unfairness can be easily highlighted with the unixbench execl 1 task. This test continuously call execve() and the scheduler looks for the idlest group/CPU on which it should place the task. Because the task runs on the local group/CPU, the latter seems already busy even if there is nothing else running on it. As a result, the scheduler will always select another group/CPU than the local one. This recovers most of the performance regression on my system from the recent load-balancer rewrite. [ mingo: Minor cleanups. ] Reported-by: kernel test robot <rong.a.chen@intel.com> Tested-by: kernel test robot <rong.a.chen@intel.com> Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Fixes: 57abff067a08 ("sched/fair: Rework find_idlest_group()") Link: https://lkml.kernel.org/r/1571762798-25900-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-22 16:46:38 +00:00
/*
* cpu_load_without - compute CPU load without any contributions from *p
* @cpu: the CPU which load is requested
* @p: the task which load should be discounted
*
* The load of a CPU is defined by the load of tasks currently enqueued on that
* CPU as well as tasks which are currently sleeping after an execution on that
* CPU.
*
* This method returns the load of the specified CPU by discounting the load of
* the specified task, whenever the task is currently contributing to the CPU
* load.
*/
static unsigned long cpu_load_without(struct rq *rq, struct task_struct *p)
{
struct cfs_rq *cfs_rq;
unsigned int load;
/* Task has no contribution or is new */
if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
return cpu_load(rq);
cfs_rq = &rq->cfs;
load = READ_ONCE(cfs_rq->avg.load_avg);
/* Discount task's util from CPU's util */
lsub_positive(&load, task_h_load(p));
return load;
}
sched/pelt: Add a new runnable average signal Now that runnable_load_avg has been removed, we can replace it by a new signal that will highlight the runnable pressure on a cfs_rq. This signal track the waiting time of tasks on rq and can help to better define the state of rqs. At now, only util_avg is used to define the state of a rq: A rq with more that around 80% of utilization and more than 1 tasks is considered as overloaded. But the util_avg signal of a rq can become temporaly low after that a task migrated onto another rq which can bias the classification of the rq. When tasks compete for the same rq, their runnable average signal will be higher than util_avg as it will include the waiting time and we can use this signal to better classify cfs_rqs. The new runnable_avg will track the runnable time of a task which simply adds the waiting time to the running time. The runnable _avg of cfs_rq will be the /Sum of se's runnable_avg and the runnable_avg of group entity will follow the one of the rq similarly to util_avg. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: "Dietmar Eggemann <dietmar.eggemann@arm.com>" Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Valentin Schneider <valentin.schneider@arm.com> Cc: Phil Auld <pauld@redhat.com> Cc: Hillf Danton <hdanton@sina.com> Link: https://lore.kernel.org/r/20200224095223.13361-9-mgorman@techsingularity.net
2020-02-24 09:52:18 +00:00
static unsigned long cpu_runnable(struct rq *rq)
{
return cfs_rq_runnable_avg(&rq->cfs);
}
static unsigned long cpu_runnable_without(struct rq *rq, struct task_struct *p)
{
struct cfs_rq *cfs_rq;
unsigned int runnable;
/* Task has no contribution or is new */
if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
return cpu_runnable(rq);
cfs_rq = &rq->cfs;
runnable = READ_ONCE(cfs_rq->avg.runnable_avg);
/* Discount task's runnable from CPU's runnable */
lsub_positive(&runnable, p->se.avg.runnable_avg);
return runnable;
}
static unsigned long capacity_of(int cpu)
{
return cpu_rq(cpu)->cpu_capacity;
}
static void record_wakee(struct task_struct *p)
{
/*
* Only decay a single time; tasks that have less then 1 wakeup per
* jiffy will not have built up many flips.
*/
if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
current->wakee_flips >>= 1;
current->wakee_flip_decay_ts = jiffies;
}
if (current->last_wakee != p) {
current->last_wakee = p;
current->wakee_flips++;
}
}
/*
* Detect M:N waker/wakee relationships via a switching-frequency heuristic.
*
* A waker of many should wake a different task than the one last awakened
* at a frequency roughly N times higher than one of its wakees.
*
* In order to determine whether we should let the load spread vs consolidating
* to shared cache, we look for a minimum 'flip' frequency of llc_size in one
* partner, and a factor of lls_size higher frequency in the other.
*
* With both conditions met, we can be relatively sure that the relationship is
* non-monogamous, with partner count exceeding socket size.
*
* Waker/wakee being client/server, worker/dispatcher, interrupt source or
* whatever is irrelevant, spread criteria is apparent partner count exceeds
* socket size.
*/
sched: Implement smarter wake-affine logic The wake-affine scheduler feature is currently always trying to pull the wakee close to the waker. In theory this should be beneficial if the waker's CPU caches hot data for the wakee, and it's also beneficial in the extreme ping-pong high context switch rate case. Testing shows it can benefit hackbench up to 15%. However, the feature is somewhat blind, from which some workloads such as pgbench suffer. It's also time-consuming algorithmically. Testing shows it can damage pgbench up to 50% - far more than the benefit it brings in the best case. So wake-affine should be smarter and it should realize when to stop its thankless effort at trying to find a suitable CPU to wake on. This patch introduces 'wakee_flips', which will be increased each time the task flips (switches) its wakee target. So a high 'wakee_flips' value means the task has more than one wakee, and the bigger the number, the higher the wakeup frequency. Now when making the decision on whether to pull or not, pay attention to the wakee with a high 'wakee_flips', pulling such a task may benefit the wakee. Also imply that the waker will face cruel competition later, it could be very cruel or very fast depends on the story behind 'wakee_flips', waker therefore suffers. Furthermore, if waker also has a high 'wakee_flips', that implies that multiple tasks rely on it, then waker's higher latency will damage all of them, so pulling wakee seems to be a bad deal. Thus, when 'waker->wakee_flips / wakee->wakee_flips' becomes higher and higher, the cost of pulling seems to be worse and worse. The patch therefore helps the wake-affine feature to stop its pulling work when: wakee->wakee_flips > factor && waker->wakee_flips > (factor * wakee->wakee_flips) The 'factor' here is the number of CPUs in the current CPU's NUMA node, so a bigger node will lead to more pulling since the trial becomes more severe. After applying the patch, pgbench shows up to 40% improvements and no regressions. Tested with 12 cpu x86 server and tip 3.10.0-rc7. The percentages in the final column highlight the areas with the biggest wins, all other areas improved as well: pgbench base smart | db_size | clients | tps | | tps | +---------+---------+-------+ +-------+ | 22 MB | 1 | 10598 | | 10796 | | 22 MB | 2 | 21257 | | 21336 | | 22 MB | 4 | 41386 | | 41622 | | 22 MB | 8 | 51253 | | 57932 | | 22 MB | 12 | 48570 | | 54000 | | 22 MB | 16 | 46748 | | 55982 | +19.75% | 22 MB | 24 | 44346 | | 55847 | +25.93% | 22 MB | 32 | 43460 | | 54614 | +25.66% | 7484 MB | 1 | 8951 | | 9193 | | 7484 MB | 2 | 19233 | | 19240 | | 7484 MB | 4 | 37239 | | 37302 | | 7484 MB | 8 | 46087 | | 50018 | | 7484 MB | 12 | 42054 | | 48763 | | 7484 MB | 16 | 40765 | | 51633 | +26.66% | 7484 MB | 24 | 37651 | | 52377 | +39.11% | 7484 MB | 32 | 37056 | | 51108 | +37.92% | 15 GB | 1 | 8845 | | 9104 | | 15 GB | 2 | 19094 | | 19162 | | 15 GB | 4 | 36979 | | 36983 | | 15 GB | 8 | 46087 | | 49977 | | 15 GB | 12 | 41901 | | 48591 | | 15 GB | 16 | 40147 | | 50651 | +26.16% | 15 GB | 24 | 37250 | | 52365 | +40.58% | 15 GB | 32 | 36470 | | 50015 | +37.14% Signed-off-by: Michael Wang <wangyun@linux.vnet.ibm.com> Cc: Mike Galbraith <efault@gmx.de> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/51D50057.9000809@linux.vnet.ibm.com [ Improved the changelog. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-07-04 04:55:51 +00:00
static int wake_wide(struct task_struct *p)
{
unsigned int master = current->wakee_flips;
unsigned int slave = p->wakee_flips;
int factor = this_cpu_read(sd_llc_size);
sched: Implement smarter wake-affine logic The wake-affine scheduler feature is currently always trying to pull the wakee close to the waker. In theory this should be beneficial if the waker's CPU caches hot data for the wakee, and it's also beneficial in the extreme ping-pong high context switch rate case. Testing shows it can benefit hackbench up to 15%. However, the feature is somewhat blind, from which some workloads such as pgbench suffer. It's also time-consuming algorithmically. Testing shows it can damage pgbench up to 50% - far more than the benefit it brings in the best case. So wake-affine should be smarter and it should realize when to stop its thankless effort at trying to find a suitable CPU to wake on. This patch introduces 'wakee_flips', which will be increased each time the task flips (switches) its wakee target. So a high 'wakee_flips' value means the task has more than one wakee, and the bigger the number, the higher the wakeup frequency. Now when making the decision on whether to pull or not, pay attention to the wakee with a high 'wakee_flips', pulling such a task may benefit the wakee. Also imply that the waker will face cruel competition later, it could be very cruel or very fast depends on the story behind 'wakee_flips', waker therefore suffers. Furthermore, if waker also has a high 'wakee_flips', that implies that multiple tasks rely on it, then waker's higher latency will damage all of them, so pulling wakee seems to be a bad deal. Thus, when 'waker->wakee_flips / wakee->wakee_flips' becomes higher and higher, the cost of pulling seems to be worse and worse. The patch therefore helps the wake-affine feature to stop its pulling work when: wakee->wakee_flips > factor && waker->wakee_flips > (factor * wakee->wakee_flips) The 'factor' here is the number of CPUs in the current CPU's NUMA node, so a bigger node will lead to more pulling since the trial becomes more severe. After applying the patch, pgbench shows up to 40% improvements and no regressions. Tested with 12 cpu x86 server and tip 3.10.0-rc7. The percentages in the final column highlight the areas with the biggest wins, all other areas improved as well: pgbench base smart | db_size | clients | tps | | tps | +---------+---------+-------+ +-------+ | 22 MB | 1 | 10598 | | 10796 | | 22 MB | 2 | 21257 | | 21336 | | 22 MB | 4 | 41386 | | 41622 | | 22 MB | 8 | 51253 | | 57932 | | 22 MB | 12 | 48570 | | 54000 | | 22 MB | 16 | 46748 | | 55982 | +19.75% | 22 MB | 24 | 44346 | | 55847 | +25.93% | 22 MB | 32 | 43460 | | 54614 | +25.66% | 7484 MB | 1 | 8951 | | 9193 | | 7484 MB | 2 | 19233 | | 19240 | | 7484 MB | 4 | 37239 | | 37302 | | 7484 MB | 8 | 46087 | | 50018 | | 7484 MB | 12 | 42054 | | 48763 | | 7484 MB | 16 | 40765 | | 51633 | +26.66% | 7484 MB | 24 | 37651 | | 52377 | +39.11% | 7484 MB | 32 | 37056 | | 51108 | +37.92% | 15 GB | 1 | 8845 | | 9104 | | 15 GB | 2 | 19094 | | 19162 | | 15 GB | 4 | 36979 | | 36983 | | 15 GB | 8 | 46087 | | 49977 | | 15 GB | 12 | 41901 | | 48591 | | 15 GB | 16 | 40147 | | 50651 | +26.16% | 15 GB | 24 | 37250 | | 52365 | +40.58% | 15 GB | 32 | 36470 | | 50015 | +37.14% Signed-off-by: Michael Wang <wangyun@linux.vnet.ibm.com> Cc: Mike Galbraith <efault@gmx.de> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/51D50057.9000809@linux.vnet.ibm.com [ Improved the changelog. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-07-04 04:55:51 +00:00
if (master < slave)
swap(master, slave);
if (slave < factor || master < slave * factor)
return 0;
return 1;
sched: Implement smarter wake-affine logic The wake-affine scheduler feature is currently always trying to pull the wakee close to the waker. In theory this should be beneficial if the waker's CPU caches hot data for the wakee, and it's also beneficial in the extreme ping-pong high context switch rate case. Testing shows it can benefit hackbench up to 15%. However, the feature is somewhat blind, from which some workloads such as pgbench suffer. It's also time-consuming algorithmically. Testing shows it can damage pgbench up to 50% - far more than the benefit it brings in the best case. So wake-affine should be smarter and it should realize when to stop its thankless effort at trying to find a suitable CPU to wake on. This patch introduces 'wakee_flips', which will be increased each time the task flips (switches) its wakee target. So a high 'wakee_flips' value means the task has more than one wakee, and the bigger the number, the higher the wakeup frequency. Now when making the decision on whether to pull or not, pay attention to the wakee with a high 'wakee_flips', pulling such a task may benefit the wakee. Also imply that the waker will face cruel competition later, it could be very cruel or very fast depends on the story behind 'wakee_flips', waker therefore suffers. Furthermore, if waker also has a high 'wakee_flips', that implies that multiple tasks rely on it, then waker's higher latency will damage all of them, so pulling wakee seems to be a bad deal. Thus, when 'waker->wakee_flips / wakee->wakee_flips' becomes higher and higher, the cost of pulling seems to be worse and worse. The patch therefore helps the wake-affine feature to stop its pulling work when: wakee->wakee_flips > factor && waker->wakee_flips > (factor * wakee->wakee_flips) The 'factor' here is the number of CPUs in the current CPU's NUMA node, so a bigger node will lead to more pulling since the trial becomes more severe. After applying the patch, pgbench shows up to 40% improvements and no regressions. Tested with 12 cpu x86 server and tip 3.10.0-rc7. The percentages in the final column highlight the areas with the biggest wins, all other areas improved as well: pgbench base smart | db_size | clients | tps | | tps | +---------+---------+-------+ +-------+ | 22 MB | 1 | 10598 | | 10796 | | 22 MB | 2 | 21257 | | 21336 | | 22 MB | 4 | 41386 | | 41622 | | 22 MB | 8 | 51253 | | 57932 | | 22 MB | 12 | 48570 | | 54000 | | 22 MB | 16 | 46748 | | 55982 | +19.75% | 22 MB | 24 | 44346 | | 55847 | +25.93% | 22 MB | 32 | 43460 | | 54614 | +25.66% | 7484 MB | 1 | 8951 | | 9193 | | 7484 MB | 2 | 19233 | | 19240 | | 7484 MB | 4 | 37239 | | 37302 | | 7484 MB | 8 | 46087 | | 50018 | | 7484 MB | 12 | 42054 | | 48763 | | 7484 MB | 16 | 40765 | | 51633 | +26.66% | 7484 MB | 24 | 37651 | | 52377 | +39.11% | 7484 MB | 32 | 37056 | | 51108 | +37.92% | 15 GB | 1 | 8845 | | 9104 | | 15 GB | 2 | 19094 | | 19162 | | 15 GB | 4 | 36979 | | 36983 | | 15 GB | 8 | 46087 | | 49977 | | 15 GB | 12 | 41901 | | 48591 | | 15 GB | 16 | 40147 | | 50651 | +26.16% | 15 GB | 24 | 37250 | | 52365 | +40.58% | 15 GB | 32 | 36470 | | 50015 | +37.14% Signed-off-by: Michael Wang <wangyun@linux.vnet.ibm.com> Cc: Mike Galbraith <efault@gmx.de> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/51D50057.9000809@linux.vnet.ibm.com [ Improved the changelog. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-07-04 04:55:51 +00:00
}
sched/fair: Fix wake_affine() for !NUMA_BALANCING In commit: 3fed382b46ba ("sched/numa: Implement NUMA node level wake_affine()") Rik changed wake_affine to consider NUMA information when balancing between LLC domains. There are a number of problems here which this patch tries to address: - LLC < NODE; in this case we'd use the wrong information to balance - !NUMA_BALANCING: in this case, the new code doesn't do any balancing at all - re-computes the NUMA data for every wakeup, this can mean iterating up to 64 CPUs for every wakeup. - default affine wakeups inside a cache We address these by saving the load/capacity values for each sched_domain during regular load-balance and using these values in wake_affine_llc(). The obvious down-side to using cached values is that they can be too old and poorly reflect reality. But this way we can use LLC wide information and thus not rely on assuming LLC matches NODE. We also don't rely on NUMA_BALANCING nor do we have to aggegate two nodes (or even cache domains) worth of CPUs for each wakeup. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Fixes: 3fed382b46ba ("sched/numa: Implement NUMA node level wake_affine()") [ Minor readability improvements. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-07-31 15:50:05 +00:00
/*
sched/core: Fix wake_affine() performance regression Eric reported a sysbench regression against commit: 3fed382b46ba ("sched/numa: Implement NUMA node level wake_affine()") Similarly, Rik was looking at the NAS-lu.C benchmark, which regressed against his v3.10 enterprise kernel. PRE (current tip/master): ivb-ep sysbench: 2: [30 secs] transactions: 64110 (2136.94 per sec.) 5: [30 secs] transactions: 143644 (4787.99 per sec.) 10: [30 secs] transactions: 274298 (9142.93 per sec.) 20: [30 secs] transactions: 418683 (13955.45 per sec.) 40: [30 secs] transactions: 320731 (10690.15 per sec.) 80: [30 secs] transactions: 355096 (11834.28 per sec.) hsw-ex NAS: OMP_PROC_BIND/lu.C.x_threads_144_run_1.log: Time in seconds = 18.01 OMP_PROC_BIND/lu.C.x_threads_144_run_2.log: Time in seconds = 17.89 OMP_PROC_BIND/lu.C.x_threads_144_run_3.log: Time in seconds = 17.93 lu.C.x_threads_144_run_1.log: Time in seconds = 434.68 lu.C.x_threads_144_run_2.log: Time in seconds = 405.36 lu.C.x_threads_144_run_3.log: Time in seconds = 433.83 POST (+patch): ivb-ep sysbench: 2: [30 secs] transactions: 64494 (2149.75 per sec.) 5: [30 secs] transactions: 145114 (4836.99 per sec.) 10: [30 secs] transactions: 278311 (9276.69 per sec.) 20: [30 secs] transactions: 437169 (14571.60 per sec.) 40: [30 secs] transactions: 669837 (22326.73 per sec.) 80: [30 secs] transactions: 631739 (21055.88 per sec.) hsw-ex NAS: lu.C.x_threads_144_run_1.log: Time in seconds = 23.36 lu.C.x_threads_144_run_2.log: Time in seconds = 22.96 lu.C.x_threads_144_run_3.log: Time in seconds = 22.52 This patch takes out all the shiny wake_affine() stuff and goes back to utter basics. Between the two CPUs involved with the wakeup (the CPU doing the wakeup and the CPU we ran on previously) pick the CPU we can run on _now_. This restores much of the regressions against the older kernels, but leaves some ground in the overloaded case. The default-enabled WA_WEIGHT (which will be introduced in the next patch) is an attempt to address the overloaded situation. Reported-by: Eric Farman <farman@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Matthew Rosato <mjrosato@linux.vnet.ibm.com> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: jinpuwang@gmail.com Cc: vcaputo@pengaru.com Fixes: 3fed382b46ba ("sched/numa: Implement NUMA node level wake_affine()") Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-09-27 09:35:30 +00:00
* The purpose of wake_affine() is to quickly determine on which CPU we can run
* soonest. For the purpose of speed we only consider the waking and previous
* CPU.
sched/fair: Fix wake_affine() for !NUMA_BALANCING In commit: 3fed382b46ba ("sched/numa: Implement NUMA node level wake_affine()") Rik changed wake_affine to consider NUMA information when balancing between LLC domains. There are a number of problems here which this patch tries to address: - LLC < NODE; in this case we'd use the wrong information to balance - !NUMA_BALANCING: in this case, the new code doesn't do any balancing at all - re-computes the NUMA data for every wakeup, this can mean iterating up to 64 CPUs for every wakeup. - default affine wakeups inside a cache We address these by saving the load/capacity values for each sched_domain during regular load-balance and using these values in wake_affine_llc(). The obvious down-side to using cached values is that they can be too old and poorly reflect reality. But this way we can use LLC wide information and thus not rely on assuming LLC matches NODE. We also don't rely on NUMA_BALANCING nor do we have to aggegate two nodes (or even cache domains) worth of CPUs for each wakeup. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Fixes: 3fed382b46ba ("sched/numa: Implement NUMA node level wake_affine()") [ Minor readability improvements. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-07-31 15:50:05 +00:00
*
sched/fair: Only immediately migrate tasks due to interrupts if prev and target CPUs share cache If waking from an idle CPU due to an interrupt then it's possible that the waker task will be pulled to wake on the current CPU. Unfortunately, depending on the type of interrupt and IRQ configuration, there may not be a strong relationship between the CPU an interrupt was delivered on and the CPU a task was running on. For example, the interrupts could all be delivered to CPUs on one particular node due to the machine topology or IRQ affinity configuration. Another example is an interrupt for an IO completion which can be delivered to any CPU where there is no guarantee the data is either cache hot or even local. This patch was motivated by the observation that an IO workload was being pulled cross-node on a frequent basis when IO completed. From a wakeup latency perspective, it's still useful to know that an idle CPU is immediately available for use but lets only consider an automatic migration if the CPUs share cache to limit damage due to NUMA migrations. Migrations may still occur if wake_affine_weight determines it's appropriate. These are the throughput results for dbench running on ext4 comparing 4.15-rc3 and this patch on a 2-socket machine where interrupts due to IO completions can happen on any CPU. 4.15.0-rc3 4.15.0-rc3 vanilla lessmigrate Hmean 1 854.64 ( 0.00%) 865.01 ( 1.21%) Hmean 2 1229.60 ( 0.00%) 1274.44 ( 3.65%) Hmean 4 1591.81 ( 0.00%) 1628.08 ( 2.28%) Hmean 8 1845.04 ( 0.00%) 1831.80 ( -0.72%) Hmean 16 2038.61 ( 0.00%) 2091.44 ( 2.59%) Hmean 32 2327.19 ( 0.00%) 2430.29 ( 4.43%) Hmean 64 2570.61 ( 0.00%) 2568.54 ( -0.08%) Hmean 128 2481.89 ( 0.00%) 2499.28 ( 0.70%) Stddev 1 14.31 ( 0.00%) 5.35 ( 62.65%) Stddev 2 21.29 ( 0.00%) 11.09 ( 47.92%) Stddev 4 7.22 ( 0.00%) 6.80 ( 5.92%) Stddev 8 26.70 ( 0.00%) 9.41 ( 64.76%) Stddev 16 22.40 ( 0.00%) 20.01 ( 10.70%) Stddev 32 45.13 ( 0.00%) 44.74 ( 0.85%) Stddev 64 93.10 ( 0.00%) 93.18 ( -0.09%) Stddev 128 184.28 ( 0.00%) 177.85 ( 3.49%) Note the small increase in throughput for low thread counts but also note that the standard deviation for each sample during the test run is lower. The throughput figures for dbench can be misleading so the benchmark is actually modified to time the latency of the processing of one load file with many samples taken. The difference in latency is 4.15.0-rc3 4.15.0-rc3 vanilla lessmigrate Amean 1 21.71 ( 0.00%) 21.47 ( 1.08%) Amean 2 30.89 ( 0.00%) 29.58 ( 4.26%) Amean 4 47.54 ( 0.00%) 46.61 ( 1.97%) Amean 8 82.71 ( 0.00%) 82.81 ( -0.12%) Amean 16 149.45 ( 0.00%) 145.01 ( 2.97%) Amean 32 265.49 ( 0.00%) 248.43 ( 6.42%) Amean 64 463.23 ( 0.00%) 463.55 ( -0.07%) Amean 128 933.97 ( 0.00%) 935.50 ( -0.16%) Stddev 1 1.58 ( 0.00%) 1.54 ( 2.26%) Stddev 2 2.84 ( 0.00%) 2.95 ( -4.15%) Stddev 4 6.78 ( 0.00%) 6.85 ( -0.99%) Stddev 8 16.85 ( 0.00%) 16.37 ( 2.85%) Stddev 16 41.59 ( 0.00%) 41.04 ( 1.32%) Stddev 32 111.05 ( 0.00%) 105.11 ( 5.35%) Stddev 64 285.94 ( 0.00%) 288.01 ( -0.72%) Stddev 128 803.39 ( 0.00%) 809.73 ( -0.79%) It's a small improvement which is not surprising given that migrations that migrate to a different node as not that common. However, it is noticeable in the CPU migration statistics which are reduced by 24%. There was a query for v1 of this patch about NAS so here are the results for C-class using MPI for parallelisation on the same machine nas-mpi 4.15.0-rc3 4.15.0-rc3 vanilla noirq Time cg.C 24.25 ( 0.00%) 23.17 ( 4.45%) Time ep.C 8.22 ( 0.00%) 8.29 ( -0.85%) Time ft.C 22.67 ( 0.00%) 20.34 ( 10.28%) Time is.C 1.42 ( 0.00%) 1.47 ( -3.52%) Time lu.C 55.62 ( 0.00%) 54.81 ( 1.46%) Time mg.C 7.93 ( 0.00%) 7.91 ( 0.25%) 4.15.0-rc3 4.15.0-rc3 vanilla noirq-v1r1 User 3799.96 3748.34 System 672.10 626.15 Elapsed 91.91 79.49 lu.C sees a small gain, ft.C a large gain and ep.C and is.C see small regressions but in terms of absolute time, the difference is small and likely within run-to-run variance. System CPU usage is slightly reduced. schbench from Facebook was also requested. This is a bit of a mixed bag but it's important to note that this workload should not be heavily impacted by wakeups from interrupt context. 4.15.0-rc3 4.15.0-rc3 vanilla noirq-v1r1 Lat 50.00th-qrtle-1 41.00 ( 0.00%) 41.00 ( 0.00%) Lat 75.00th-qrtle-1 42.00 ( 0.00%) 42.00 ( 0.00%) Lat 90.00th-qrtle-1 43.00 ( 0.00%) 44.00 ( -2.33%) Lat 95.00th-qrtle-1 44.00 ( 0.00%) 46.00 ( -4.55%) Lat 99.00th-qrtle-1 57.00 ( 0.00%) 58.00 ( -1.75%) Lat 99.50th-qrtle-1 59.00 ( 0.00%) 59.00 ( 0.00%) Lat 99.90th-qrtle-1 67.00 ( 0.00%) 78.00 ( -16.42%) Lat 50.00th-qrtle-2 40.00 ( 0.00%) 51.00 ( -27.50%) Lat 75.00th-qrtle-2 45.00 ( 0.00%) 56.00 ( -24.44%) Lat 90.00th-qrtle-2 53.00 ( 0.00%) 59.00 ( -11.32%) Lat 95.00th-qrtle-2 57.00 ( 0.00%) 61.00 ( -7.02%) Lat 99.00th-qrtle-2 67.00 ( 0.00%) 71.00 ( -5.97%) Lat 99.50th-qrtle-2 69.00 ( 0.00%) 74.00 ( -7.25%) Lat 99.90th-qrtle-2 83.00 ( 0.00%) 77.00 ( 7.23%) Lat 50.00th-qrtle-4 51.00 ( 0.00%) 51.00 ( 0.00%) Lat 75.00th-qrtle-4 57.00 ( 0.00%) 56.00 ( 1.75%) Lat 90.00th-qrtle-4 60.00 ( 0.00%) 59.00 ( 1.67%) Lat 95.00th-qrtle-4 62.00 ( 0.00%) 62.00 ( 0.00%) Lat 99.00th-qrtle-4 73.00 ( 0.00%) 72.00 ( 1.37%) Lat 99.50th-qrtle-4 76.00 ( 0.00%) 74.00 ( 2.63%) Lat 99.90th-qrtle-4 85.00 ( 0.00%) 78.00 ( 8.24%) Lat 50.00th-qrtle-8 54.00 ( 0.00%) 58.00 ( -7.41%) Lat 75.00th-qrtle-8 59.00 ( 0.00%) 62.00 ( -5.08%) Lat 90.00th-qrtle-8 65.00 ( 0.00%) 66.00 ( -1.54%) Lat 95.00th-qrtle-8 67.00 ( 0.00%) 70.00 ( -4.48%) Lat 99.00th-qrtle-8 78.00 ( 0.00%) 79.00 ( -1.28%) Lat 99.50th-qrtle-8 81.00 ( 0.00%) 80.00 ( 1.23%) Lat 99.90th-qrtle-8 116.00 ( 0.00%) 83.00 ( 28.45%) Lat 50.00th-qrtle-16 65.00 ( 0.00%) 64.00 ( 1.54%) Lat 75.00th-qrtle-16 77.00 ( 0.00%) 71.00 ( 7.79%) Lat 90.00th-qrtle-16 83.00 ( 0.00%) 82.00 ( 1.20%) Lat 95.00th-qrtle-16 87.00 ( 0.00%) 87.00 ( 0.00%) Lat 99.00th-qrtle-16 95.00 ( 0.00%) 96.00 ( -1.05%) Lat 99.50th-qrtle-16 99.00 ( 0.00%) 103.00 ( -4.04%) Lat 99.90th-qrtle-16 104.00 ( 0.00%) 122.00 ( -17.31%) Lat 50.00th-qrtle-32 71.00 ( 0.00%) 73.00 ( -2.82%) Lat 75.00th-qrtle-32 91.00 ( 0.00%) 92.00 ( -1.10%) Lat 90.00th-qrtle-32 108.00 ( 0.00%) 107.00 ( 0.93%) Lat 95.00th-qrtle-32 118.00 ( 0.00%) 115.00 ( 2.54%) Lat 99.00th-qrtle-32 134.00 ( 0.00%) 129.00 ( 3.73%) Lat 99.50th-qrtle-32 138.00 ( 0.00%) 133.00 ( 3.62%) Lat 99.90th-qrtle-32 149.00 ( 0.00%) 146.00 ( 2.01%) Lat 50.00th-qrtle-39 83.00 ( 0.00%) 81.00 ( 2.41%) Lat 75.00th-qrtle-39 105.00 ( 0.00%) 102.00 ( 2.86%) Lat 90.00th-qrtle-39 120.00 ( 0.00%) 119.00 ( 0.83%) Lat 95.00th-qrtle-39 129.00 ( 0.00%) 128.00 ( 0.78%) Lat 99.00th-qrtle-39 153.00 ( 0.00%) 149.00 ( 2.61%) Lat 99.50th-qrtle-39 166.00 ( 0.00%) 156.00 ( 6.02%) Lat 99.90th-qrtle-39 12304.00 ( 0.00%) 12848.00 ( -4.42%) When heavily loaded (e.g. 99.50th-qrtle-39 indicates 39 threads), there are small gains in many cases. Otherwise it depends on the quartile used where it can be bad -- e.g. 75.00th-qrtle-2. However, even these results are probably a co-incidence. For this workload, much depends on what node the threads get placed on and their relative locality and not wakeups from interrupt context. A larger component on how it behaves would be automatic NUMA balancing where a fault incurred to measure locality would be a much larger contributer to latency than the wakeup path. This is the results from an almost identical machine that happened to run the same test. They only differ in terms of storage which is irrelevant for this test. 4.15.0-rc3 4.15.0-rc3 vanilla noirq-v1r1 Lat 50.00th-qrtle-1 41.00 ( 0.00%) 41.00 ( 0.00%) Lat 75.00th-qrtle-1 42.00 ( 0.00%) 42.00 ( 0.00%) Lat 90.00th-qrtle-1 44.00 ( 0.00%) 43.00 ( 2.27%) Lat 95.00th-qrtle-1 53.00 ( 0.00%) 45.00 ( 15.09%) Lat 99.00th-qrtle-1 59.00 ( 0.00%) 58.00 ( 1.69%) Lat 99.50th-qrtle-1 60.00 ( 0.00%) 59.00 ( 1.67%) Lat 99.90th-qrtle-1 86.00 ( 0.00%) 61.00 ( 29.07%) Lat 50.00th-qrtle-2 52.00 ( 0.00%) 41.00 ( 21.15%) Lat 75.00th-qrtle-2 57.00 ( 0.00%) 46.00 ( 19.30%) Lat 90.00th-qrtle-2 60.00 ( 0.00%) 53.00 ( 11.67%) Lat 95.00th-qrtle-2 62.00 ( 0.00%) 57.00 ( 8.06%) Lat 99.00th-qrtle-2 73.00 ( 0.00%) 68.00 ( 6.85%) Lat 99.50th-qrtle-2 74.00 ( 0.00%) 71.00 ( 4.05%) Lat 99.90th-qrtle-2 90.00 ( 0.00%) 75.00 ( 16.67%) Lat 50.00th-qrtle-4 57.00 ( 0.00%) 52.00 ( 8.77%) Lat 75.00th-qrtle-4 60.00 ( 0.00%) 58.00 ( 3.33%) Lat 90.00th-qrtle-4 62.00 ( 0.00%) 62.00 ( 0.00%) Lat 95.00th-qrtle-4 65.00 ( 0.00%) 65.00 ( 0.00%) Lat 99.00th-qrtle-4 76.00 ( 0.00%) 75.00 ( 1.32%) Lat 99.50th-qrtle-4 77.00 ( 0.00%) 77.00 ( 0.00%) Lat 99.90th-qrtle-4 87.00 ( 0.00%) 81.00 ( 6.90%) Lat 50.00th-qrtle-8 59.00 ( 0.00%) 57.00 ( 3.39%) Lat 75.00th-qrtle-8 63.00 ( 0.00%) 62.00 ( 1.59%) Lat 90.00th-qrtle-8 66.00 ( 0.00%) 67.00 ( -1.52%) Lat 95.00th-qrtle-8 68.00 ( 0.00%) 70.00 ( -2.94%) Lat 99.00th-qrtle-8 79.00 ( 0.00%) 80.00 ( -1.27%) Lat 99.50th-qrtle-8 80.00 ( 0.00%) 84.00 ( -5.00%) Lat 99.90th-qrtle-8 84.00 ( 0.00%) 90.00 ( -7.14%) Lat 50.00th-qrtle-16 65.00 ( 0.00%) 65.00 ( 0.00%) Lat 75.00th-qrtle-16 77.00 ( 0.00%) 75.00 ( 2.60%) Lat 90.00th-qrtle-16 84.00 ( 0.00%) 83.00 ( 1.19%) Lat 95.00th-qrtle-16 88.00 ( 0.00%) 87.00 ( 1.14%) Lat 99.00th-qrtle-16 97.00 ( 0.00%) 96.00 ( 1.03%) Lat 99.50th-qrtle-16 100.00 ( 0.00%) 104.00 ( -4.00%) Lat 99.90th-qrtle-16 110.00 ( 0.00%) 126.00 ( -14.55%) Lat 50.00th-qrtle-32 70.00 ( 0.00%) 71.00 ( -1.43%) Lat 75.00th-qrtle-32 92.00 ( 0.00%) 94.00 ( -2.17%) Lat 90.00th-qrtle-32 110.00 ( 0.00%) 110.00 ( 0.00%) Lat 95.00th-qrtle-32 121.00 ( 0.00%) 118.00 ( 2.48%) Lat 99.00th-qrtle-32 135.00 ( 0.00%) 137.00 ( -1.48%) Lat 99.50th-qrtle-32 140.00 ( 0.00%) 146.00 ( -4.29%) Lat 99.90th-qrtle-32 150.00 ( 0.00%) 160.00 ( -6.67%) Lat 50.00th-qrtle-39 80.00 ( 0.00%) 71.00 ( 11.25%) Lat 75.00th-qrtle-39 102.00 ( 0.00%) 91.00 ( 10.78%) Lat 90.00th-qrtle-39 118.00 ( 0.00%) 108.00 ( 8.47%) Lat 95.00th-qrtle-39 128.00 ( 0.00%) 117.00 ( 8.59%) Lat 99.00th-qrtle-39 149.00 ( 0.00%) 133.00 ( 10.74%) Lat 99.50th-qrtle-39 160.00 ( 0.00%) 139.00 ( 13.12%) Lat 99.90th-qrtle-39 13808.00 ( 0.00%) 4920.00 ( 64.37%) Despite being nearly identical, it showed a variety of major gains so I'm not convinced that heavy emphasis should be placed on this particular workload in terms of evaluating this particular patch. Further evidence of this is the fact that testing on a UMA machine showed small gains/losses even though the patch should be a no-op on UMA. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Matt Fleming <matt@codeblueprint.co.uk> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/20171219085947.13136-2-mgorman@techsingularity.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-12-19 08:59:47 +00:00
* wake_affine_idle() - only considers 'now', it check if the waking CPU is
* cache-affine and is (or will be) idle.
sched/core: Address more wake_affine() regressions The trivial wake_affine_idle() implementation is very good for a number of workloads, but it comes apart at the moment there are no idle CPUs left, IOW. the overloaded case. hackbench: NO_WA_WEIGHT WA_WEIGHT hackbench-20 : 7.362717561 seconds 6.450509391 seconds (win) netperf: NO_WA_WEIGHT WA_WEIGHT TCP_SENDFILE-1 : Avg: 54524.6 Avg: 52224.3 TCP_SENDFILE-10 : Avg: 48185.2 Avg: 46504.3 TCP_SENDFILE-20 : Avg: 29031.2 Avg: 28610.3 TCP_SENDFILE-40 : Avg: 9819.72 Avg: 9253.12 TCP_SENDFILE-80 : Avg: 5355.3 Avg: 4687.4 TCP_STREAM-1 : Avg: 41448.3 Avg: 42254 TCP_STREAM-10 : Avg: 24123.2 Avg: 25847.9 TCP_STREAM-20 : Avg: 15834.5 Avg: 18374.4 TCP_STREAM-40 : Avg: 5583.91 Avg: 5599.57 TCP_STREAM-80 : Avg: 2329.66 Avg: 2726.41 TCP_RR-1 : Avg: 80473.5 Avg: 82638.8 TCP_RR-10 : Avg: 72660.5 Avg: 73265.1 TCP_RR-20 : Avg: 52607.1 Avg: 52634.5 TCP_RR-40 : Avg: 57199.2 Avg: 56302.3 TCP_RR-80 : Avg: 25330.3 Avg: 26867.9 UDP_RR-1 : Avg: 108266 Avg: 107844 UDP_RR-10 : Avg: 95480 Avg: 95245.2 UDP_RR-20 : Avg: 68770.8 Avg: 68673.7 UDP_RR-40 : Avg: 76231 Avg: 75419.1 UDP_RR-80 : Avg: 34578.3 Avg: 35639.1 UDP_STREAM-1 : Avg: 64684.3 Avg: 66606 UDP_STREAM-10 : Avg: 52701.2 Avg: 52959.5 UDP_STREAM-20 : Avg: 30376.4 Avg: 29704 UDP_STREAM-40 : Avg: 15685.8 Avg: 15266.5 UDP_STREAM-80 : Avg: 8415.13 Avg: 7388.97 (wins and losses) sysbench: NO_WA_WEIGHT WA_WEIGHT sysbench-mysql-2 : 2135.17 per sec. 2142.51 per sec. sysbench-mysql-5 : 4809.68 per sec. 4800.19 per sec. sysbench-mysql-10 : 9158.59 per sec. 9157.05 per sec. sysbench-mysql-20 : 14570.70 per sec. 14543.55 per sec. sysbench-mysql-40 : 22130.56 per sec. 22184.82 per sec. sysbench-mysql-80 : 20995.56 per sec. 21904.18 per sec. sysbench-psql-2 : 1679.58 per sec. 1705.06 per sec. sysbench-psql-5 : 3797.69 per sec. 3879.93 per sec. sysbench-psql-10 : 7253.22 per sec. 7258.06 per sec. sysbench-psql-20 : 11166.75 per sec. 11220.00 per sec. sysbench-psql-40 : 17277.28 per sec. 17359.78 per sec. sysbench-psql-80 : 17112.44 per sec. 17221.16 per sec. (increase on the top end) tbench: NO_WA_WEIGHT Throughput 685.211 MB/sec 2 clients 2 procs max_latency=0.123 ms Throughput 1596.64 MB/sec 5 clients 5 procs max_latency=0.119 ms Throughput 2985.47 MB/sec 10 clients 10 procs max_latency=0.262 ms Throughput 4521.15 MB/sec 20 clients 20 procs max_latency=0.506 ms Throughput 9438.1 MB/sec 40 clients 40 procs max_latency=2.052 ms Throughput 8210.5 MB/sec 80 clients 80 procs max_latency=8.310 ms WA_WEIGHT Throughput 697.292 MB/sec 2 clients 2 procs max_latency=0.127 ms Throughput 1596.48 MB/sec 5 clients 5 procs max_latency=0.080 ms Throughput 2975.22 MB/sec 10 clients 10 procs max_latency=0.254 ms Throughput 4575.14 MB/sec 20 clients 20 procs max_latency=0.502 ms Throughput 9468.65 MB/sec 40 clients 40 procs max_latency=2.069 ms Throughput 8631.73 MB/sec 80 clients 80 procs max_latency=8.605 ms (increase on the top end) Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Rik van Riel <riel@redhat.com> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-10-06 07:23:24 +00:00
*
* wake_affine_weight() - considers the weight to reflect the average
* scheduling latency of the CPUs. This seems to work
* for the overloaded case.
sched/fair: Fix wake_affine() for !NUMA_BALANCING In commit: 3fed382b46ba ("sched/numa: Implement NUMA node level wake_affine()") Rik changed wake_affine to consider NUMA information when balancing between LLC domains. There are a number of problems here which this patch tries to address: - LLC < NODE; in this case we'd use the wrong information to balance - !NUMA_BALANCING: in this case, the new code doesn't do any balancing at all - re-computes the NUMA data for every wakeup, this can mean iterating up to 64 CPUs for every wakeup. - default affine wakeups inside a cache We address these by saving the load/capacity values for each sched_domain during regular load-balance and using these values in wake_affine_llc(). The obvious down-side to using cached values is that they can be too old and poorly reflect reality. But this way we can use LLC wide information and thus not rely on assuming LLC matches NODE. We also don't rely on NUMA_BALANCING nor do we have to aggegate two nodes (or even cache domains) worth of CPUs for each wakeup. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Fixes: 3fed382b46ba ("sched/numa: Implement NUMA node level wake_affine()") [ Minor readability improvements. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-07-31 15:50:05 +00:00
*/
static int
wake_affine_idle(int this_cpu, int prev_cpu, int sync)
sched/fair: Fix wake_affine() for !NUMA_BALANCING In commit: 3fed382b46ba ("sched/numa: Implement NUMA node level wake_affine()") Rik changed wake_affine to consider NUMA information when balancing between LLC domains. There are a number of problems here which this patch tries to address: - LLC < NODE; in this case we'd use the wrong information to balance - !NUMA_BALANCING: in this case, the new code doesn't do any balancing at all - re-computes the NUMA data for every wakeup, this can mean iterating up to 64 CPUs for every wakeup. - default affine wakeups inside a cache We address these by saving the load/capacity values for each sched_domain during regular load-balance and using these values in wake_affine_llc(). The obvious down-side to using cached values is that they can be too old and poorly reflect reality. But this way we can use LLC wide information and thus not rely on assuming LLC matches NODE. We also don't rely on NUMA_BALANCING nor do we have to aggegate two nodes (or even cache domains) worth of CPUs for each wakeup. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Fixes: 3fed382b46ba ("sched/numa: Implement NUMA node level wake_affine()") [ Minor readability improvements. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-07-31 15:50:05 +00:00
{
sched/fair: Only immediately migrate tasks due to interrupts if prev and target CPUs share cache If waking from an idle CPU due to an interrupt then it's possible that the waker task will be pulled to wake on the current CPU. Unfortunately, depending on the type of interrupt and IRQ configuration, there may not be a strong relationship between the CPU an interrupt was delivered on and the CPU a task was running on. For example, the interrupts could all be delivered to CPUs on one particular node due to the machine topology or IRQ affinity configuration. Another example is an interrupt for an IO completion which can be delivered to any CPU where there is no guarantee the data is either cache hot or even local. This patch was motivated by the observation that an IO workload was being pulled cross-node on a frequent basis when IO completed. From a wakeup latency perspective, it's still useful to know that an idle CPU is immediately available for use but lets only consider an automatic migration if the CPUs share cache to limit damage due to NUMA migrations. Migrations may still occur if wake_affine_weight determines it's appropriate. These are the throughput results for dbench running on ext4 comparing 4.15-rc3 and this patch on a 2-socket machine where interrupts due to IO completions can happen on any CPU. 4.15.0-rc3 4.15.0-rc3 vanilla lessmigrate Hmean 1 854.64 ( 0.00%) 865.01 ( 1.21%) Hmean 2 1229.60 ( 0.00%) 1274.44 ( 3.65%) Hmean 4 1591.81 ( 0.00%) 1628.08 ( 2.28%) Hmean 8 1845.04 ( 0.00%) 1831.80 ( -0.72%) Hmean 16 2038.61 ( 0.00%) 2091.44 ( 2.59%) Hmean 32 2327.19 ( 0.00%) 2430.29 ( 4.43%) Hmean 64 2570.61 ( 0.00%) 2568.54 ( -0.08%) Hmean 128 2481.89 ( 0.00%) 2499.28 ( 0.70%) Stddev 1 14.31 ( 0.00%) 5.35 ( 62.65%) Stddev 2 21.29 ( 0.00%) 11.09 ( 47.92%) Stddev 4 7.22 ( 0.00%) 6.80 ( 5.92%) Stddev 8 26.70 ( 0.00%) 9.41 ( 64.76%) Stddev 16 22.40 ( 0.00%) 20.01 ( 10.70%) Stddev 32 45.13 ( 0.00%) 44.74 ( 0.85%) Stddev 64 93.10 ( 0.00%) 93.18 ( -0.09%) Stddev 128 184.28 ( 0.00%) 177.85 ( 3.49%) Note the small increase in throughput for low thread counts but also note that the standard deviation for each sample during the test run is lower. The throughput figures for dbench can be misleading so the benchmark is actually modified to time the latency of the processing of one load file with many samples taken. The difference in latency is 4.15.0-rc3 4.15.0-rc3 vanilla lessmigrate Amean 1 21.71 ( 0.00%) 21.47 ( 1.08%) Amean 2 30.89 ( 0.00%) 29.58 ( 4.26%) Amean 4 47.54 ( 0.00%) 46.61 ( 1.97%) Amean 8 82.71 ( 0.00%) 82.81 ( -0.12%) Amean 16 149.45 ( 0.00%) 145.01 ( 2.97%) Amean 32 265.49 ( 0.00%) 248.43 ( 6.42%) Amean 64 463.23 ( 0.00%) 463.55 ( -0.07%) Amean 128 933.97 ( 0.00%) 935.50 ( -0.16%) Stddev 1 1.58 ( 0.00%) 1.54 ( 2.26%) Stddev 2 2.84 ( 0.00%) 2.95 ( -4.15%) Stddev 4 6.78 ( 0.00%) 6.85 ( -0.99%) Stddev 8 16.85 ( 0.00%) 16.37 ( 2.85%) Stddev 16 41.59 ( 0.00%) 41.04 ( 1.32%) Stddev 32 111.05 ( 0.00%) 105.11 ( 5.35%) Stddev 64 285.94 ( 0.00%) 288.01 ( -0.72%) Stddev 128 803.39 ( 0.00%) 809.73 ( -0.79%) It's a small improvement which is not surprising given that migrations that migrate to a different node as not that common. However, it is noticeable in the CPU migration statistics which are reduced by 24%. There was a query for v1 of this patch about NAS so here are the results for C-class using MPI for parallelisation on the same machine nas-mpi 4.15.0-rc3 4.15.0-rc3 vanilla noirq Time cg.C 24.25 ( 0.00%) 23.17 ( 4.45%) Time ep.C 8.22 ( 0.00%) 8.29 ( -0.85%) Time ft.C 22.67 ( 0.00%) 20.34 ( 10.28%) Time is.C 1.42 ( 0.00%) 1.47 ( -3.52%) Time lu.C 55.62 ( 0.00%) 54.81 ( 1.46%) Time mg.C 7.93 ( 0.00%) 7.91 ( 0.25%) 4.15.0-rc3 4.15.0-rc3 vanilla noirq-v1r1 User 3799.96 3748.34 System 672.10 626.15 Elapsed 91.91 79.49 lu.C sees a small gain, ft.C a large gain and ep.C and is.C see small regressions but in terms of absolute time, the difference is small and likely within run-to-run variance. System CPU usage is slightly reduced. schbench from Facebook was also requested. This is a bit of a mixed bag but it's important to note that this workload should not be heavily impacted by wakeups from interrupt context. 4.15.0-rc3 4.15.0-rc3 vanilla noirq-v1r1 Lat 50.00th-qrtle-1 41.00 ( 0.00%) 41.00 ( 0.00%) Lat 75.00th-qrtle-1 42.00 ( 0.00%) 42.00 ( 0.00%) Lat 90.00th-qrtle-1 43.00 ( 0.00%) 44.00 ( -2.33%) Lat 95.00th-qrtle-1 44.00 ( 0.00%) 46.00 ( -4.55%) Lat 99.00th-qrtle-1 57.00 ( 0.00%) 58.00 ( -1.75%) Lat 99.50th-qrtle-1 59.00 ( 0.00%) 59.00 ( 0.00%) Lat 99.90th-qrtle-1 67.00 ( 0.00%) 78.00 ( -16.42%) Lat 50.00th-qrtle-2 40.00 ( 0.00%) 51.00 ( -27.50%) Lat 75.00th-qrtle-2 45.00 ( 0.00%) 56.00 ( -24.44%) Lat 90.00th-qrtle-2 53.00 ( 0.00%) 59.00 ( -11.32%) Lat 95.00th-qrtle-2 57.00 ( 0.00%) 61.00 ( -7.02%) Lat 99.00th-qrtle-2 67.00 ( 0.00%) 71.00 ( -5.97%) Lat 99.50th-qrtle-2 69.00 ( 0.00%) 74.00 ( -7.25%) Lat 99.90th-qrtle-2 83.00 ( 0.00%) 77.00 ( 7.23%) Lat 50.00th-qrtle-4 51.00 ( 0.00%) 51.00 ( 0.00%) Lat 75.00th-qrtle-4 57.00 ( 0.00%) 56.00 ( 1.75%) Lat 90.00th-qrtle-4 60.00 ( 0.00%) 59.00 ( 1.67%) Lat 95.00th-qrtle-4 62.00 ( 0.00%) 62.00 ( 0.00%) Lat 99.00th-qrtle-4 73.00 ( 0.00%) 72.00 ( 1.37%) Lat 99.50th-qrtle-4 76.00 ( 0.00%) 74.00 ( 2.63%) Lat 99.90th-qrtle-4 85.00 ( 0.00%) 78.00 ( 8.24%) Lat 50.00th-qrtle-8 54.00 ( 0.00%) 58.00 ( -7.41%) Lat 75.00th-qrtle-8 59.00 ( 0.00%) 62.00 ( -5.08%) Lat 90.00th-qrtle-8 65.00 ( 0.00%) 66.00 ( -1.54%) Lat 95.00th-qrtle-8 67.00 ( 0.00%) 70.00 ( -4.48%) Lat 99.00th-qrtle-8 78.00 ( 0.00%) 79.00 ( -1.28%) Lat 99.50th-qrtle-8 81.00 ( 0.00%) 80.00 ( 1.23%) Lat 99.90th-qrtle-8 116.00 ( 0.00%) 83.00 ( 28.45%) Lat 50.00th-qrtle-16 65.00 ( 0.00%) 64.00 ( 1.54%) Lat 75.00th-qrtle-16 77.00 ( 0.00%) 71.00 ( 7.79%) Lat 90.00th-qrtle-16 83.00 ( 0.00%) 82.00 ( 1.20%) Lat 95.00th-qrtle-16 87.00 ( 0.00%) 87.00 ( 0.00%) Lat 99.00th-qrtle-16 95.00 ( 0.00%) 96.00 ( -1.05%) Lat 99.50th-qrtle-16 99.00 ( 0.00%) 103.00 ( -4.04%) Lat 99.90th-qrtle-16 104.00 ( 0.00%) 122.00 ( -17.31%) Lat 50.00th-qrtle-32 71.00 ( 0.00%) 73.00 ( -2.82%) Lat 75.00th-qrtle-32 91.00 ( 0.00%) 92.00 ( -1.10%) Lat 90.00th-qrtle-32 108.00 ( 0.00%) 107.00 ( 0.93%) Lat 95.00th-qrtle-32 118.00 ( 0.00%) 115.00 ( 2.54%) Lat 99.00th-qrtle-32 134.00 ( 0.00%) 129.00 ( 3.73%) Lat 99.50th-qrtle-32 138.00 ( 0.00%) 133.00 ( 3.62%) Lat 99.90th-qrtle-32 149.00 ( 0.00%) 146.00 ( 2.01%) Lat 50.00th-qrtle-39 83.00 ( 0.00%) 81.00 ( 2.41%) Lat 75.00th-qrtle-39 105.00 ( 0.00%) 102.00 ( 2.86%) Lat 90.00th-qrtle-39 120.00 ( 0.00%) 119.00 ( 0.83%) Lat 95.00th-qrtle-39 129.00 ( 0.00%) 128.00 ( 0.78%) Lat 99.00th-qrtle-39 153.00 ( 0.00%) 149.00 ( 2.61%) Lat 99.50th-qrtle-39 166.00 ( 0.00%) 156.00 ( 6.02%) Lat 99.90th-qrtle-39 12304.00 ( 0.00%) 12848.00 ( -4.42%) When heavily loaded (e.g. 99.50th-qrtle-39 indicates 39 threads), there are small gains in many cases. Otherwise it depends on the quartile used where it can be bad -- e.g. 75.00th-qrtle-2. However, even these results are probably a co-incidence. For this workload, much depends on what node the threads get placed on and their relative locality and not wakeups from interrupt context. A larger component on how it behaves would be automatic NUMA balancing where a fault incurred to measure locality would be a much larger contributer to latency than the wakeup path. This is the results from an almost identical machine that happened to run the same test. They only differ in terms of storage which is irrelevant for this test. 4.15.0-rc3 4.15.0-rc3 vanilla noirq-v1r1 Lat 50.00th-qrtle-1 41.00 ( 0.00%) 41.00 ( 0.00%) Lat 75.00th-qrtle-1 42.00 ( 0.00%) 42.00 ( 0.00%) Lat 90.00th-qrtle-1 44.00 ( 0.00%) 43.00 ( 2.27%) Lat 95.00th-qrtle-1 53.00 ( 0.00%) 45.00 ( 15.09%) Lat 99.00th-qrtle-1 59.00 ( 0.00%) 58.00 ( 1.69%) Lat 99.50th-qrtle-1 60.00 ( 0.00%) 59.00 ( 1.67%) Lat 99.90th-qrtle-1 86.00 ( 0.00%) 61.00 ( 29.07%) Lat 50.00th-qrtle-2 52.00 ( 0.00%) 41.00 ( 21.15%) Lat 75.00th-qrtle-2 57.00 ( 0.00%) 46.00 ( 19.30%) Lat 90.00th-qrtle-2 60.00 ( 0.00%) 53.00 ( 11.67%) Lat 95.00th-qrtle-2 62.00 ( 0.00%) 57.00 ( 8.06%) Lat 99.00th-qrtle-2 73.00 ( 0.00%) 68.00 ( 6.85%) Lat 99.50th-qrtle-2 74.00 ( 0.00%) 71.00 ( 4.05%) Lat 99.90th-qrtle-2 90.00 ( 0.00%) 75.00 ( 16.67%) Lat 50.00th-qrtle-4 57.00 ( 0.00%) 52.00 ( 8.77%) Lat 75.00th-qrtle-4 60.00 ( 0.00%) 58.00 ( 3.33%) Lat 90.00th-qrtle-4 62.00 ( 0.00%) 62.00 ( 0.00%) Lat 95.00th-qrtle-4 65.00 ( 0.00%) 65.00 ( 0.00%) Lat 99.00th-qrtle-4 76.00 ( 0.00%) 75.00 ( 1.32%) Lat 99.50th-qrtle-4 77.00 ( 0.00%) 77.00 ( 0.00%) Lat 99.90th-qrtle-4 87.00 ( 0.00%) 81.00 ( 6.90%) Lat 50.00th-qrtle-8 59.00 ( 0.00%) 57.00 ( 3.39%) Lat 75.00th-qrtle-8 63.00 ( 0.00%) 62.00 ( 1.59%) Lat 90.00th-qrtle-8 66.00 ( 0.00%) 67.00 ( -1.52%) Lat 95.00th-qrtle-8 68.00 ( 0.00%) 70.00 ( -2.94%) Lat 99.00th-qrtle-8 79.00 ( 0.00%) 80.00 ( -1.27%) Lat 99.50th-qrtle-8 80.00 ( 0.00%) 84.00 ( -5.00%) Lat 99.90th-qrtle-8 84.00 ( 0.00%) 90.00 ( -7.14%) Lat 50.00th-qrtle-16 65.00 ( 0.00%) 65.00 ( 0.00%) Lat 75.00th-qrtle-16 77.00 ( 0.00%) 75.00 ( 2.60%) Lat 90.00th-qrtle-16 84.00 ( 0.00%) 83.00 ( 1.19%) Lat 95.00th-qrtle-16 88.00 ( 0.00%) 87.00 ( 1.14%) Lat 99.00th-qrtle-16 97.00 ( 0.00%) 96.00 ( 1.03%) Lat 99.50th-qrtle-16 100.00 ( 0.00%) 104.00 ( -4.00%) Lat 99.90th-qrtle-16 110.00 ( 0.00%) 126.00 ( -14.55%) Lat 50.00th-qrtle-32 70.00 ( 0.00%) 71.00 ( -1.43%) Lat 75.00th-qrtle-32 92.00 ( 0.00%) 94.00 ( -2.17%) Lat 90.00th-qrtle-32 110.00 ( 0.00%) 110.00 ( 0.00%) Lat 95.00th-qrtle-32 121.00 ( 0.00%) 118.00 ( 2.48%) Lat 99.00th-qrtle-32 135.00 ( 0.00%) 137.00 ( -1.48%) Lat 99.50th-qrtle-32 140.00 ( 0.00%) 146.00 ( -4.29%) Lat 99.90th-qrtle-32 150.00 ( 0.00%) 160.00 ( -6.67%) Lat 50.00th-qrtle-39 80.00 ( 0.00%) 71.00 ( 11.25%) Lat 75.00th-qrtle-39 102.00 ( 0.00%) 91.00 ( 10.78%) Lat 90.00th-qrtle-39 118.00 ( 0.00%) 108.00 ( 8.47%) Lat 95.00th-qrtle-39 128.00 ( 0.00%) 117.00 ( 8.59%) Lat 99.00th-qrtle-39 149.00 ( 0.00%) 133.00 ( 10.74%) Lat 99.50th-qrtle-39 160.00 ( 0.00%) 139.00 ( 13.12%) Lat 99.90th-qrtle-39 13808.00 ( 0.00%) 4920.00 ( 64.37%) Despite being nearly identical, it showed a variety of major gains so I'm not convinced that heavy emphasis should be placed on this particular workload in terms of evaluating this particular patch. Further evidence of this is the fact that testing on a UMA machine showed small gains/losses even though the patch should be a no-op on UMA. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Matt Fleming <matt@codeblueprint.co.uk> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/20171219085947.13136-2-mgorman@techsingularity.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-12-19 08:59:47 +00:00
/*
* If this_cpu is idle, it implies the wakeup is from interrupt
* context. Only allow the move if cache is shared. Otherwise an
* interrupt intensive workload could force all tasks onto one
* node depending on the IO topology or IRQ affinity settings.
sched/fair: Do not migrate if the prev_cpu is idle wake_affine_idle() prefers to move a task to the current CPU if the wakeup is due to an interrupt. The expectation is that the interrupt data is cache hot and relevant to the waking task as well as avoiding a search. However, there is no way to determine if there was cache hot data on the previous CPU that may exceed the interrupt data. Furthermore, round-robin delivery of interrupts can migrate tasks around a socket where each CPU is under-utilised. This can interact badly with cpufreq which makes decisions based on per-cpu data. It has been observed on machines with HWP that p-states are not boosted to their maximum levels even though the workload is latency and throughput sensitive. This patch uses the previous CPU for the task if it's idle and cache-affine with the current CPU even if the current CPU is idle due to the wakup being related to the interrupt. This reduces migrations at the cost of the interrupt data not being cache hot when the task wakes. A variety of workloads were tested on various machines and no adverse impact was noticed that was outside noise. dbench on ext4 on UMA showed roughly 10% reduction in the number of CPU migrations and it is a case where interrupts are frequent for IO competions. In most cases, the difference in performance is quite small but variability is often reduced. For example, this is the result for pgbench running on a UMA machine with different numbers of clients. 4.15.0-rc9 4.15.0-rc9 baseline waprev-v1 Hmean 1 22096.28 ( 0.00%) 22734.86 ( 2.89%) Hmean 4 74633.42 ( 0.00%) 75496.77 ( 1.16%) Hmean 7 115017.50 ( 0.00%) 113030.81 ( -1.73%) Hmean 12 126209.63 ( 0.00%) 126613.40 ( 0.32%) Hmean 16 131886.91 ( 0.00%) 130844.35 ( -0.79%) Stddev 1 636.38 ( 0.00%) 417.11 ( 34.46%) Stddev 4 614.64 ( 0.00%) 583.24 ( 5.11%) Stddev 7 542.46 ( 0.00%) 435.45 ( 19.73%) Stddev 12 173.93 ( 0.00%) 171.50 ( 1.40%) Stddev 16 671.42 ( 0.00%) 680.30 ( -1.32%) CoeffVar 1 2.88 ( 0.00%) 1.83 ( 36.26%) Note that the different in performance is marginal but for low utilisation, there is less variability. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Matt Fleming <matt@codeblueprint.co.uk> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/20180130104555.4125-4-mgorman@techsingularity.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-01-30 10:45:54 +00:00
*
* If the prev_cpu is idle and cache affine then avoid a migration.
* There is no guarantee that the cache hot data from an interrupt
* is more important than cache hot data on the prev_cpu and from
* a cpufreq perspective, it's better to have higher utilisation
* on one CPU.
sched/fair: Only immediately migrate tasks due to interrupts if prev and target CPUs share cache If waking from an idle CPU due to an interrupt then it's possible that the waker task will be pulled to wake on the current CPU. Unfortunately, depending on the type of interrupt and IRQ configuration, there may not be a strong relationship between the CPU an interrupt was delivered on and the CPU a task was running on. For example, the interrupts could all be delivered to CPUs on one particular node due to the machine topology or IRQ affinity configuration. Another example is an interrupt for an IO completion which can be delivered to any CPU where there is no guarantee the data is either cache hot or even local. This patch was motivated by the observation that an IO workload was being pulled cross-node on a frequent basis when IO completed. From a wakeup latency perspective, it's still useful to know that an idle CPU is immediately available for use but lets only consider an automatic migration if the CPUs share cache to limit damage due to NUMA migrations. Migrations may still occur if wake_affine_weight determines it's appropriate. These are the throughput results for dbench running on ext4 comparing 4.15-rc3 and this patch on a 2-socket machine where interrupts due to IO completions can happen on any CPU. 4.15.0-rc3 4.15.0-rc3 vanilla lessmigrate Hmean 1 854.64 ( 0.00%) 865.01 ( 1.21%) Hmean 2 1229.60 ( 0.00%) 1274.44 ( 3.65%) Hmean 4 1591.81 ( 0.00%) 1628.08 ( 2.28%) Hmean 8 1845.04 ( 0.00%) 1831.80 ( -0.72%) Hmean 16 2038.61 ( 0.00%) 2091.44 ( 2.59%) Hmean 32 2327.19 ( 0.00%) 2430.29 ( 4.43%) Hmean 64 2570.61 ( 0.00%) 2568.54 ( -0.08%) Hmean 128 2481.89 ( 0.00%) 2499.28 ( 0.70%) Stddev 1 14.31 ( 0.00%) 5.35 ( 62.65%) Stddev 2 21.29 ( 0.00%) 11.09 ( 47.92%) Stddev 4 7.22 ( 0.00%) 6.80 ( 5.92%) Stddev 8 26.70 ( 0.00%) 9.41 ( 64.76%) Stddev 16 22.40 ( 0.00%) 20.01 ( 10.70%) Stddev 32 45.13 ( 0.00%) 44.74 ( 0.85%) Stddev 64 93.10 ( 0.00%) 93.18 ( -0.09%) Stddev 128 184.28 ( 0.00%) 177.85 ( 3.49%) Note the small increase in throughput for low thread counts but also note that the standard deviation for each sample during the test run is lower. The throughput figures for dbench can be misleading so the benchmark is actually modified to time the latency of the processing of one load file with many samples taken. The difference in latency is 4.15.0-rc3 4.15.0-rc3 vanilla lessmigrate Amean 1 21.71 ( 0.00%) 21.47 ( 1.08%) Amean 2 30.89 ( 0.00%) 29.58 ( 4.26%) Amean 4 47.54 ( 0.00%) 46.61 ( 1.97%) Amean 8 82.71 ( 0.00%) 82.81 ( -0.12%) Amean 16 149.45 ( 0.00%) 145.01 ( 2.97%) Amean 32 265.49 ( 0.00%) 248.43 ( 6.42%) Amean 64 463.23 ( 0.00%) 463.55 ( -0.07%) Amean 128 933.97 ( 0.00%) 935.50 ( -0.16%) Stddev 1 1.58 ( 0.00%) 1.54 ( 2.26%) Stddev 2 2.84 ( 0.00%) 2.95 ( -4.15%) Stddev 4 6.78 ( 0.00%) 6.85 ( -0.99%) Stddev 8 16.85 ( 0.00%) 16.37 ( 2.85%) Stddev 16 41.59 ( 0.00%) 41.04 ( 1.32%) Stddev 32 111.05 ( 0.00%) 105.11 ( 5.35%) Stddev 64 285.94 ( 0.00%) 288.01 ( -0.72%) Stddev 128 803.39 ( 0.00%) 809.73 ( -0.79%) It's a small improvement which is not surprising given that migrations that migrate to a different node as not that common. However, it is noticeable in the CPU migration statistics which are reduced by 24%. There was a query for v1 of this patch about NAS so here are the results for C-class using MPI for parallelisation on the same machine nas-mpi 4.15.0-rc3 4.15.0-rc3 vanilla noirq Time cg.C 24.25 ( 0.00%) 23.17 ( 4.45%) Time ep.C 8.22 ( 0.00%) 8.29 ( -0.85%) Time ft.C 22.67 ( 0.00%) 20.34 ( 10.28%) Time is.C 1.42 ( 0.00%) 1.47 ( -3.52%) Time lu.C 55.62 ( 0.00%) 54.81 ( 1.46%) Time mg.C 7.93 ( 0.00%) 7.91 ( 0.25%) 4.15.0-rc3 4.15.0-rc3 vanilla noirq-v1r1 User 3799.96 3748.34 System 672.10 626.15 Elapsed 91.91 79.49 lu.C sees a small gain, ft.C a large gain and ep.C and is.C see small regressions but in terms of absolute time, the difference is small and likely within run-to-run variance. System CPU usage is slightly reduced. schbench from Facebook was also requested. This is a bit of a mixed bag but it's important to note that this workload should not be heavily impacted by wakeups from interrupt context. 4.15.0-rc3 4.15.0-rc3 vanilla noirq-v1r1 Lat 50.00th-qrtle-1 41.00 ( 0.00%) 41.00 ( 0.00%) Lat 75.00th-qrtle-1 42.00 ( 0.00%) 42.00 ( 0.00%) Lat 90.00th-qrtle-1 43.00 ( 0.00%) 44.00 ( -2.33%) Lat 95.00th-qrtle-1 44.00 ( 0.00%) 46.00 ( -4.55%) Lat 99.00th-qrtle-1 57.00 ( 0.00%) 58.00 ( -1.75%) Lat 99.50th-qrtle-1 59.00 ( 0.00%) 59.00 ( 0.00%) Lat 99.90th-qrtle-1 67.00 ( 0.00%) 78.00 ( -16.42%) Lat 50.00th-qrtle-2 40.00 ( 0.00%) 51.00 ( -27.50%) Lat 75.00th-qrtle-2 45.00 ( 0.00%) 56.00 ( -24.44%) Lat 90.00th-qrtle-2 53.00 ( 0.00%) 59.00 ( -11.32%) Lat 95.00th-qrtle-2 57.00 ( 0.00%) 61.00 ( -7.02%) Lat 99.00th-qrtle-2 67.00 ( 0.00%) 71.00 ( -5.97%) Lat 99.50th-qrtle-2 69.00 ( 0.00%) 74.00 ( -7.25%) Lat 99.90th-qrtle-2 83.00 ( 0.00%) 77.00 ( 7.23%) Lat 50.00th-qrtle-4 51.00 ( 0.00%) 51.00 ( 0.00%) Lat 75.00th-qrtle-4 57.00 ( 0.00%) 56.00 ( 1.75%) Lat 90.00th-qrtle-4 60.00 ( 0.00%) 59.00 ( 1.67%) Lat 95.00th-qrtle-4 62.00 ( 0.00%) 62.00 ( 0.00%) Lat 99.00th-qrtle-4 73.00 ( 0.00%) 72.00 ( 1.37%) Lat 99.50th-qrtle-4 76.00 ( 0.00%) 74.00 ( 2.63%) Lat 99.90th-qrtle-4 85.00 ( 0.00%) 78.00 ( 8.24%) Lat 50.00th-qrtle-8 54.00 ( 0.00%) 58.00 ( -7.41%) Lat 75.00th-qrtle-8 59.00 ( 0.00%) 62.00 ( -5.08%) Lat 90.00th-qrtle-8 65.00 ( 0.00%) 66.00 ( -1.54%) Lat 95.00th-qrtle-8 67.00 ( 0.00%) 70.00 ( -4.48%) Lat 99.00th-qrtle-8 78.00 ( 0.00%) 79.00 ( -1.28%) Lat 99.50th-qrtle-8 81.00 ( 0.00%) 80.00 ( 1.23%) Lat 99.90th-qrtle-8 116.00 ( 0.00%) 83.00 ( 28.45%) Lat 50.00th-qrtle-16 65.00 ( 0.00%) 64.00 ( 1.54%) Lat 75.00th-qrtle-16 77.00 ( 0.00%) 71.00 ( 7.79%) Lat 90.00th-qrtle-16 83.00 ( 0.00%) 82.00 ( 1.20%) Lat 95.00th-qrtle-16 87.00 ( 0.00%) 87.00 ( 0.00%) Lat 99.00th-qrtle-16 95.00 ( 0.00%) 96.00 ( -1.05%) Lat 99.50th-qrtle-16 99.00 ( 0.00%) 103.00 ( -4.04%) Lat 99.90th-qrtle-16 104.00 ( 0.00%) 122.00 ( -17.31%) Lat 50.00th-qrtle-32 71.00 ( 0.00%) 73.00 ( -2.82%) Lat 75.00th-qrtle-32 91.00 ( 0.00%) 92.00 ( -1.10%) Lat 90.00th-qrtle-32 108.00 ( 0.00%) 107.00 ( 0.93%) Lat 95.00th-qrtle-32 118.00 ( 0.00%) 115.00 ( 2.54%) Lat 99.00th-qrtle-32 134.00 ( 0.00%) 129.00 ( 3.73%) Lat 99.50th-qrtle-32 138.00 ( 0.00%) 133.00 ( 3.62%) Lat 99.90th-qrtle-32 149.00 ( 0.00%) 146.00 ( 2.01%) Lat 50.00th-qrtle-39 83.00 ( 0.00%) 81.00 ( 2.41%) Lat 75.00th-qrtle-39 105.00 ( 0.00%) 102.00 ( 2.86%) Lat 90.00th-qrtle-39 120.00 ( 0.00%) 119.00 ( 0.83%) Lat 95.00th-qrtle-39 129.00 ( 0.00%) 128.00 ( 0.78%) Lat 99.00th-qrtle-39 153.00 ( 0.00%) 149.00 ( 2.61%) Lat 99.50th-qrtle-39 166.00 ( 0.00%) 156.00 ( 6.02%) Lat 99.90th-qrtle-39 12304.00 ( 0.00%) 12848.00 ( -4.42%) When heavily loaded (e.g. 99.50th-qrtle-39 indicates 39 threads), there are small gains in many cases. Otherwise it depends on the quartile used where it can be bad -- e.g. 75.00th-qrtle-2. However, even these results are probably a co-incidence. For this workload, much depends on what node the threads get placed on and their relative locality and not wakeups from interrupt context. A larger component on how it behaves would be automatic NUMA balancing where a fault incurred to measure locality would be a much larger contributer to latency than the wakeup path. This is the results from an almost identical machine that happened to run the same test. They only differ in terms of storage which is irrelevant for this test. 4.15.0-rc3 4.15.0-rc3 vanilla noirq-v1r1 Lat 50.00th-qrtle-1 41.00 ( 0.00%) 41.00 ( 0.00%) Lat 75.00th-qrtle-1 42.00 ( 0.00%) 42.00 ( 0.00%) Lat 90.00th-qrtle-1 44.00 ( 0.00%) 43.00 ( 2.27%) Lat 95.00th-qrtle-1 53.00 ( 0.00%) 45.00 ( 15.09%) Lat 99.00th-qrtle-1 59.00 ( 0.00%) 58.00 ( 1.69%) Lat 99.50th-qrtle-1 60.00 ( 0.00%) 59.00 ( 1.67%) Lat 99.90th-qrtle-1 86.00 ( 0.00%) 61.00 ( 29.07%) Lat 50.00th-qrtle-2 52.00 ( 0.00%) 41.00 ( 21.15%) Lat 75.00th-qrtle-2 57.00 ( 0.00%) 46.00 ( 19.30%) Lat 90.00th-qrtle-2 60.00 ( 0.00%) 53.00 ( 11.67%) Lat 95.00th-qrtle-2 62.00 ( 0.00%) 57.00 ( 8.06%) Lat 99.00th-qrtle-2 73.00 ( 0.00%) 68.00 ( 6.85%) Lat 99.50th-qrtle-2 74.00 ( 0.00%) 71.00 ( 4.05%) Lat 99.90th-qrtle-2 90.00 ( 0.00%) 75.00 ( 16.67%) Lat 50.00th-qrtle-4 57.00 ( 0.00%) 52.00 ( 8.77%) Lat 75.00th-qrtle-4 60.00 ( 0.00%) 58.00 ( 3.33%) Lat 90.00th-qrtle-4 62.00 ( 0.00%) 62.00 ( 0.00%) Lat 95.00th-qrtle-4 65.00 ( 0.00%) 65.00 ( 0.00%) Lat 99.00th-qrtle-4 76.00 ( 0.00%) 75.00 ( 1.32%) Lat 99.50th-qrtle-4 77.00 ( 0.00%) 77.00 ( 0.00%) Lat 99.90th-qrtle-4 87.00 ( 0.00%) 81.00 ( 6.90%) Lat 50.00th-qrtle-8 59.00 ( 0.00%) 57.00 ( 3.39%) Lat 75.00th-qrtle-8 63.00 ( 0.00%) 62.00 ( 1.59%) Lat 90.00th-qrtle-8 66.00 ( 0.00%) 67.00 ( -1.52%) Lat 95.00th-qrtle-8 68.00 ( 0.00%) 70.00 ( -2.94%) Lat 99.00th-qrtle-8 79.00 ( 0.00%) 80.00 ( -1.27%) Lat 99.50th-qrtle-8 80.00 ( 0.00%) 84.00 ( -5.00%) Lat 99.90th-qrtle-8 84.00 ( 0.00%) 90.00 ( -7.14%) Lat 50.00th-qrtle-16 65.00 ( 0.00%) 65.00 ( 0.00%) Lat 75.00th-qrtle-16 77.00 ( 0.00%) 75.00 ( 2.60%) Lat 90.00th-qrtle-16 84.00 ( 0.00%) 83.00 ( 1.19%) Lat 95.00th-qrtle-16 88.00 ( 0.00%) 87.00 ( 1.14%) Lat 99.00th-qrtle-16 97.00 ( 0.00%) 96.00 ( 1.03%) Lat 99.50th-qrtle-16 100.00 ( 0.00%) 104.00 ( -4.00%) Lat 99.90th-qrtle-16 110.00 ( 0.00%) 126.00 ( -14.55%) Lat 50.00th-qrtle-32 70.00 ( 0.00%) 71.00 ( -1.43%) Lat 75.00th-qrtle-32 92.00 ( 0.00%) 94.00 ( -2.17%) Lat 90.00th-qrtle-32 110.00 ( 0.00%) 110.00 ( 0.00%) Lat 95.00th-qrtle-32 121.00 ( 0.00%) 118.00 ( 2.48%) Lat 99.00th-qrtle-32 135.00 ( 0.00%) 137.00 ( -1.48%) Lat 99.50th-qrtle-32 140.00 ( 0.00%) 146.00 ( -4.29%) Lat 99.90th-qrtle-32 150.00 ( 0.00%) 160.00 ( -6.67%) Lat 50.00th-qrtle-39 80.00 ( 0.00%) 71.00 ( 11.25%) Lat 75.00th-qrtle-39 102.00 ( 0.00%) 91.00 ( 10.78%) Lat 90.00th-qrtle-39 118.00 ( 0.00%) 108.00 ( 8.47%) Lat 95.00th-qrtle-39 128.00 ( 0.00%) 117.00 ( 8.59%) Lat 99.00th-qrtle-39 149.00 ( 0.00%) 133.00 ( 10.74%) Lat 99.50th-qrtle-39 160.00 ( 0.00%) 139.00 ( 13.12%) Lat 99.90th-qrtle-39 13808.00 ( 0.00%) 4920.00 ( 64.37%) Despite being nearly identical, it showed a variety of major gains so I'm not convinced that heavy emphasis should be placed on this particular workload in terms of evaluating this particular patch. Further evidence of this is the fact that testing on a UMA machine showed small gains/losses even though the patch should be a no-op on UMA. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Matt Fleming <matt@codeblueprint.co.uk> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/20171219085947.13136-2-mgorman@techsingularity.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-12-19 08:59:47 +00:00
*/
if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
sched/fair: Fix wake_affine() for !NUMA_BALANCING In commit: 3fed382b46ba ("sched/numa: Implement NUMA node level wake_affine()") Rik changed wake_affine to consider NUMA information when balancing between LLC domains. There are a number of problems here which this patch tries to address: - LLC < NODE; in this case we'd use the wrong information to balance - !NUMA_BALANCING: in this case, the new code doesn't do any balancing at all - re-computes the NUMA data for every wakeup, this can mean iterating up to 64 CPUs for every wakeup. - default affine wakeups inside a cache We address these by saving the load/capacity values for each sched_domain during regular load-balance and using these values in wake_affine_llc(). The obvious down-side to using cached values is that they can be too old and poorly reflect reality. But this way we can use LLC wide information and thus not rely on assuming LLC matches NODE. We also don't rely on NUMA_BALANCING nor do we have to aggegate two nodes (or even cache domains) worth of CPUs for each wakeup. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Fixes: 3fed382b46ba ("sched/numa: Implement NUMA node level wake_affine()") [ Minor readability improvements. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-07-31 15:50:05 +00:00
sched/core: Fix wake_affine() performance regression Eric reported a sysbench regression against commit: 3fed382b46ba ("sched/numa: Implement NUMA node level wake_affine()") Similarly, Rik was looking at the NAS-lu.C benchmark, which regressed against his v3.10 enterprise kernel. PRE (current tip/master): ivb-ep sysbench: 2: [30 secs] transactions: 64110 (2136.94 per sec.) 5: [30 secs] transactions: 143644 (4787.99 per sec.) 10: [30 secs] transactions: 274298 (9142.93 per sec.) 20: [30 secs] transactions: 418683 (13955.45 per sec.) 40: [30 secs] transactions: 320731 (10690.15 per sec.) 80: [30 secs] transactions: 355096 (11834.28 per sec.) hsw-ex NAS: OMP_PROC_BIND/lu.C.x_threads_144_run_1.log: Time in seconds = 18.01 OMP_PROC_BIND/lu.C.x_threads_144_run_2.log: Time in seconds = 17.89 OMP_PROC_BIND/lu.C.x_threads_144_run_3.log: Time in seconds = 17.93 lu.C.x_threads_144_run_1.log: Time in seconds = 434.68 lu.C.x_threads_144_run_2.log: Time in seconds = 405.36 lu.C.x_threads_144_run_3.log: Time in seconds = 433.83 POST (+patch): ivb-ep sysbench: 2: [30 secs] transactions: 64494 (2149.75 per sec.) 5: [30 secs] transactions: 145114 (4836.99 per sec.) 10: [30 secs] transactions: 278311 (9276.69 per sec.) 20: [30 secs] transactions: 437169 (14571.60 per sec.) 40: [30 secs] transactions: 669837 (22326.73 per sec.) 80: [30 secs] transactions: 631739 (21055.88 per sec.) hsw-ex NAS: lu.C.x_threads_144_run_1.log: Time in seconds = 23.36 lu.C.x_threads_144_run_2.log: Time in seconds = 22.96 lu.C.x_threads_144_run_3.log: Time in seconds = 22.52 This patch takes out all the shiny wake_affine() stuff and goes back to utter basics. Between the two CPUs involved with the wakeup (the CPU doing the wakeup and the CPU we ran on previously) pick the CPU we can run on _now_. This restores much of the regressions against the older kernels, but leaves some ground in the overloaded case. The default-enabled WA_WEIGHT (which will be introduced in the next patch) is an attempt to address the overloaded situation. Reported-by: Eric Farman <farman@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Matthew Rosato <mjrosato@linux.vnet.ibm.com> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: jinpuwang@gmail.com Cc: vcaputo@pengaru.com Fixes: 3fed382b46ba ("sched/numa: Implement NUMA node level wake_affine()") Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-09-27 09:35:30 +00:00
if (sync && cpu_rq(this_cpu)->nr_running == 1)
return this_cpu;
sched/fair: Fix wake_affine() for !NUMA_BALANCING In commit: 3fed382b46ba ("sched/numa: Implement NUMA node level wake_affine()") Rik changed wake_affine to consider NUMA information when balancing between LLC domains. There are a number of problems here which this patch tries to address: - LLC < NODE; in this case we'd use the wrong information to balance - !NUMA_BALANCING: in this case, the new code doesn't do any balancing at all - re-computes the NUMA data for every wakeup, this can mean iterating up to 64 CPUs for every wakeup. - default affine wakeups inside a cache We address these by saving the load/capacity values for each sched_domain during regular load-balance and using these values in wake_affine_llc(). The obvious down-side to using cached values is that they can be too old and poorly reflect reality. But this way we can use LLC wide information and thus not rely on assuming LLC matches NODE. We also don't rely on NUMA_BALANCING nor do we have to aggegate two nodes (or even cache domains) worth of CPUs for each wakeup. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Fixes: 3fed382b46ba ("sched/numa: Implement NUMA node level wake_affine()") [ Minor readability improvements. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-07-31 15:50:05 +00:00
return nr_cpumask_bits;
sched/fair: Fix wake_affine() for !NUMA_BALANCING In commit: 3fed382b46ba ("sched/numa: Implement NUMA node level wake_affine()") Rik changed wake_affine to consider NUMA information when balancing between LLC domains. There are a number of problems here which this patch tries to address: - LLC < NODE; in this case we'd use the wrong information to balance - !NUMA_BALANCING: in this case, the new code doesn't do any balancing at all - re-computes the NUMA data for every wakeup, this can mean iterating up to 64 CPUs for every wakeup. - default affine wakeups inside a cache We address these by saving the load/capacity values for each sched_domain during regular load-balance and using these values in wake_affine_llc(). The obvious down-side to using cached values is that they can be too old and poorly reflect reality. But this way we can use LLC wide information and thus not rely on assuming LLC matches NODE. We also don't rely on NUMA_BALANCING nor do we have to aggegate two nodes (or even cache domains) worth of CPUs for each wakeup. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Fixes: 3fed382b46ba ("sched/numa: Implement NUMA node level wake_affine()") [ Minor readability improvements. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-07-31 15:50:05 +00:00
}
static int
sched/core: Address more wake_affine() regressions The trivial wake_affine_idle() implementation is very good for a number of workloads, but it comes apart at the moment there are no idle CPUs left, IOW. the overloaded case. hackbench: NO_WA_WEIGHT WA_WEIGHT hackbench-20 : 7.362717561 seconds 6.450509391 seconds (win) netperf: NO_WA_WEIGHT WA_WEIGHT TCP_SENDFILE-1 : Avg: 54524.6 Avg: 52224.3 TCP_SENDFILE-10 : Avg: 48185.2 Avg: 46504.3 TCP_SENDFILE-20 : Avg: 29031.2 Avg: 28610.3 TCP_SENDFILE-40 : Avg: 9819.72 Avg: 9253.12 TCP_SENDFILE-80 : Avg: 5355.3 Avg: 4687.4 TCP_STREAM-1 : Avg: 41448.3 Avg: 42254 TCP_STREAM-10 : Avg: 24123.2 Avg: 25847.9 TCP_STREAM-20 : Avg: 15834.5 Avg: 18374.4 TCP_STREAM-40 : Avg: 5583.91 Avg: 5599.57 TCP_STREAM-80 : Avg: 2329.66 Avg: 2726.41 TCP_RR-1 : Avg: 80473.5 Avg: 82638.8 TCP_RR-10 : Avg: 72660.5 Avg: 73265.1 TCP_RR-20 : Avg: 52607.1 Avg: 52634.5 TCP_RR-40 : Avg: 57199.2 Avg: 56302.3 TCP_RR-80 : Avg: 25330.3 Avg: 26867.9 UDP_RR-1 : Avg: 108266 Avg: 107844 UDP_RR-10 : Avg: 95480 Avg: 95245.2 UDP_RR-20 : Avg: 68770.8 Avg: 68673.7 UDP_RR-40 : Avg: 76231 Avg: 75419.1 UDP_RR-80 : Avg: 34578.3 Avg: 35639.1 UDP_STREAM-1 : Avg: 64684.3 Avg: 66606 UDP_STREAM-10 : Avg: 52701.2 Avg: 52959.5 UDP_STREAM-20 : Avg: 30376.4 Avg: 29704 UDP_STREAM-40 : Avg: 15685.8 Avg: 15266.5 UDP_STREAM-80 : Avg: 8415.13 Avg: 7388.97 (wins and losses) sysbench: NO_WA_WEIGHT WA_WEIGHT sysbench-mysql-2 : 2135.17 per sec. 2142.51 per sec. sysbench-mysql-5 : 4809.68 per sec. 4800.19 per sec. sysbench-mysql-10 : 9158.59 per sec. 9157.05 per sec. sysbench-mysql-20 : 14570.70 per sec. 14543.55 per sec. sysbench-mysql-40 : 22130.56 per sec. 22184.82 per sec. sysbench-mysql-80 : 20995.56 per sec. 21904.18 per sec. sysbench-psql-2 : 1679.58 per sec. 1705.06 per sec. sysbench-psql-5 : 3797.69 per sec. 3879.93 per sec. sysbench-psql-10 : 7253.22 per sec. 7258.06 per sec. sysbench-psql-20 : 11166.75 per sec. 11220.00 per sec. sysbench-psql-40 : 17277.28 per sec. 17359.78 per sec. sysbench-psql-80 : 17112.44 per sec. 17221.16 per sec. (increase on the top end) tbench: NO_WA_WEIGHT Throughput 685.211 MB/sec 2 clients 2 procs max_latency=0.123 ms Throughput 1596.64 MB/sec 5 clients 5 procs max_latency=0.119 ms Throughput 2985.47 MB/sec 10 clients 10 procs max_latency=0.262 ms Throughput 4521.15 MB/sec 20 clients 20 procs max_latency=0.506 ms Throughput 9438.1 MB/sec 40 clients 40 procs max_latency=2.052 ms Throughput 8210.5 MB/sec 80 clients 80 procs max_latency=8.310 ms WA_WEIGHT Throughput 697.292 MB/sec 2 clients 2 procs max_latency=0.127 ms Throughput 1596.48 MB/sec 5 clients 5 procs max_latency=0.080 ms Throughput 2975.22 MB/sec 10 clients 10 procs max_latency=0.254 ms Throughput 4575.14 MB/sec 20 clients 20 procs max_latency=0.502 ms Throughput 9468.65 MB/sec 40 clients 40 procs max_latency=2.069 ms Throughput 8631.73 MB/sec 80 clients 80 procs max_latency=8.605 ms (increase on the top end) Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Rik van Riel <riel@redhat.com> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-10-06 07:23:24 +00:00
wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
int this_cpu, int prev_cpu, int sync)
sched/fair: Fix wake_affine() for !NUMA_BALANCING In commit: 3fed382b46ba ("sched/numa: Implement NUMA node level wake_affine()") Rik changed wake_affine to consider NUMA information when balancing between LLC domains. There are a number of problems here which this patch tries to address: - LLC < NODE; in this case we'd use the wrong information to balance - !NUMA_BALANCING: in this case, the new code doesn't do any balancing at all - re-computes the NUMA data for every wakeup, this can mean iterating up to 64 CPUs for every wakeup. - default affine wakeups inside a cache We address these by saving the load/capacity values for each sched_domain during regular load-balance and using these values in wake_affine_llc(). The obvious down-side to using cached values is that they can be too old and poorly reflect reality. But this way we can use LLC wide information and thus not rely on assuming LLC matches NODE. We also don't rely on NUMA_BALANCING nor do we have to aggegate two nodes (or even cache domains) worth of CPUs for each wakeup. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Fixes: 3fed382b46ba ("sched/numa: Implement NUMA node level wake_affine()") [ Minor readability improvements. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-07-31 15:50:05 +00:00
{
s64 this_eff_load, prev_eff_load;
unsigned long task_load;
this_eff_load = cpu_load(cpu_rq(this_cpu));
sched/fair: Fix wake_affine() for !NUMA_BALANCING In commit: 3fed382b46ba ("sched/numa: Implement NUMA node level wake_affine()") Rik changed wake_affine to consider NUMA information when balancing between LLC domains. There are a number of problems here which this patch tries to address: - LLC < NODE; in this case we'd use the wrong information to balance - !NUMA_BALANCING: in this case, the new code doesn't do any balancing at all - re-computes the NUMA data for every wakeup, this can mean iterating up to 64 CPUs for every wakeup. - default affine wakeups inside a cache We address these by saving the load/capacity values for each sched_domain during regular load-balance and using these values in wake_affine_llc(). The obvious down-side to using cached values is that they can be too old and poorly reflect reality. But this way we can use LLC wide information and thus not rely on assuming LLC matches NODE. We also don't rely on NUMA_BALANCING nor do we have to aggegate two nodes (or even cache domains) worth of CPUs for each wakeup. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Fixes: 3fed382b46ba ("sched/numa: Implement NUMA node level wake_affine()") [ Minor readability improvements. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-07-31 15:50:05 +00:00
if (sync) {
unsigned long current_load = task_h_load(current);
sched/core: Address more wake_affine() regressions The trivial wake_affine_idle() implementation is very good for a number of workloads, but it comes apart at the moment there are no idle CPUs left, IOW. the overloaded case. hackbench: NO_WA_WEIGHT WA_WEIGHT hackbench-20 : 7.362717561 seconds 6.450509391 seconds (win) netperf: NO_WA_WEIGHT WA_WEIGHT TCP_SENDFILE-1 : Avg: 54524.6 Avg: 52224.3 TCP_SENDFILE-10 : Avg: 48185.2 Avg: 46504.3 TCP_SENDFILE-20 : Avg: 29031.2 Avg: 28610.3 TCP_SENDFILE-40 : Avg: 9819.72 Avg: 9253.12 TCP_SENDFILE-80 : Avg: 5355.3 Avg: 4687.4 TCP_STREAM-1 : Avg: 41448.3 Avg: 42254 TCP_STREAM-10 : Avg: 24123.2 Avg: 25847.9 TCP_STREAM-20 : Avg: 15834.5 Avg: 18374.4 TCP_STREAM-40 : Avg: 5583.91 Avg: 5599.57 TCP_STREAM-80 : Avg: 2329.66 Avg: 2726.41 TCP_RR-1 : Avg: 80473.5 Avg: 82638.8 TCP_RR-10 : Avg: 72660.5 Avg: 73265.1 TCP_RR-20 : Avg: 52607.1 Avg: 52634.5 TCP_RR-40 : Avg: 57199.2 Avg: 56302.3 TCP_RR-80 : Avg: 25330.3 Avg: 26867.9 UDP_RR-1 : Avg: 108266 Avg: 107844 UDP_RR-10 : Avg: 95480 Avg: 95245.2 UDP_RR-20 : Avg: 68770.8 Avg: 68673.7 UDP_RR-40 : Avg: 76231 Avg: 75419.1 UDP_RR-80 : Avg: 34578.3 Avg: 35639.1 UDP_STREAM-1 : Avg: 64684.3 Avg: 66606 UDP_STREAM-10 : Avg: 52701.2 Avg: 52959.5 UDP_STREAM-20 : Avg: 30376.4 Avg: 29704 UDP_STREAM-40 : Avg: 15685.8 Avg: 15266.5 UDP_STREAM-80 : Avg: 8415.13 Avg: 7388.97 (wins and losses) sysbench: NO_WA_WEIGHT WA_WEIGHT sysbench-mysql-2 : 2135.17 per sec. 2142.51 per sec. sysbench-mysql-5 : 4809.68 per sec. 4800.19 per sec. sysbench-mysql-10 : 9158.59 per sec. 9157.05 per sec. sysbench-mysql-20 : 14570.70 per sec. 14543.55 per sec. sysbench-mysql-40 : 22130.56 per sec. 22184.82 per sec. sysbench-mysql-80 : 20995.56 per sec. 21904.18 per sec. sysbench-psql-2 : 1679.58 per sec. 1705.06 per sec. sysbench-psql-5 : 3797.69 per sec. 3879.93 per sec. sysbench-psql-10 : 7253.22 per sec. 7258.06 per sec. sysbench-psql-20 : 11166.75 per sec. 11220.00 per sec. sysbench-psql-40 : 17277.28 per sec. 17359.78 per sec. sysbench-psql-80 : 17112.44 per sec. 17221.16 per sec. (increase on the top end) tbench: NO_WA_WEIGHT Throughput 685.211 MB/sec 2 clients 2 procs max_latency=0.123 ms Throughput 1596.64 MB/sec 5 clients 5 procs max_latency=0.119 ms Throughput 2985.47 MB/sec 10 clients 10 procs max_latency=0.262 ms Throughput 4521.15 MB/sec 20 clients 20 procs max_latency=0.506 ms Throughput 9438.1 MB/sec 40 clients 40 procs max_latency=2.052 ms Throughput 8210.5 MB/sec 80 clients 80 procs max_latency=8.310 ms WA_WEIGHT Throughput 697.292 MB/sec 2 clients 2 procs max_latency=0.127 ms Throughput 1596.48 MB/sec 5 clients 5 procs max_latency=0.080 ms Throughput 2975.22 MB/sec 10 clients 10 procs max_latency=0.254 ms Throughput 4575.14 MB/sec 20 clients 20 procs max_latency=0.502 ms Throughput 9468.65 MB/sec 40 clients 40 procs max_latency=2.069 ms Throughput 8631.73 MB/sec 80 clients 80 procs max_latency=8.605 ms (increase on the top end) Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Rik van Riel <riel@redhat.com> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-10-06 07:23:24 +00:00
if (current_load > this_eff_load)
return this_cpu;
sched/fair: Fix wake_affine() for !NUMA_BALANCING In commit: 3fed382b46ba ("sched/numa: Implement NUMA node level wake_affine()") Rik changed wake_affine to consider NUMA information when balancing between LLC domains. There are a number of problems here which this patch tries to address: - LLC < NODE; in this case we'd use the wrong information to balance - !NUMA_BALANCING: in this case, the new code doesn't do any balancing at all - re-computes the NUMA data for every wakeup, this can mean iterating up to 64 CPUs for every wakeup. - default affine wakeups inside a cache We address these by saving the load/capacity values for each sched_domain during regular load-balance and using these values in wake_affine_llc(). The obvious down-side to using cached values is that they can be too old and poorly reflect reality. But this way we can use LLC wide information and thus not rely on assuming LLC matches NODE. We also don't rely on NUMA_BALANCING nor do we have to aggegate two nodes (or even cache domains) worth of CPUs for each wakeup. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Fixes: 3fed382b46ba ("sched/numa: Implement NUMA node level wake_affine()") [ Minor readability improvements. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-07-31 15:50:05 +00:00
sched/core: Address more wake_affine() regressions The trivial wake_affine_idle() implementation is very good for a number of workloads, but it comes apart at the moment there are no idle CPUs left, IOW. the overloaded case. hackbench: NO_WA_WEIGHT WA_WEIGHT hackbench-20 : 7.362717561 seconds 6.450509391 seconds (win) netperf: NO_WA_WEIGHT WA_WEIGHT TCP_SENDFILE-1 : Avg: 54524.6 Avg: 52224.3 TCP_SENDFILE-10 : Avg: 48185.2 Avg: 46504.3 TCP_SENDFILE-20 : Avg: 29031.2 Avg: 28610.3 TCP_SENDFILE-40 : Avg: 9819.72 Avg: 9253.12 TCP_SENDFILE-80 : Avg: 5355.3 Avg: 4687.4 TCP_STREAM-1 : Avg: 41448.3 Avg: 42254 TCP_STREAM-10 : Avg: 24123.2 Avg: 25847.9 TCP_STREAM-20 : Avg: 15834.5 Avg: 18374.4 TCP_STREAM-40 : Avg: 5583.91 Avg: 5599.57 TCP_STREAM-80 : Avg: 2329.66 Avg: 2726.41 TCP_RR-1 : Avg: 80473.5 Avg: 82638.8 TCP_RR-10 : Avg: 72660.5 Avg: 73265.1 TCP_RR-20 : Avg: 52607.1 Avg: 52634.5 TCP_RR-40 : Avg: 57199.2 Avg: 56302.3 TCP_RR-80 : Avg: 25330.3 Avg: 26867.9 UDP_RR-1 : Avg: 108266 Avg: 107844 UDP_RR-10 : Avg: 95480 Avg: 95245.2 UDP_RR-20 : Avg: 68770.8 Avg: 68673.7 UDP_RR-40 : Avg: 76231 Avg: 75419.1 UDP_RR-80 : Avg: 34578.3 Avg: 35639.1 UDP_STREAM-1 : Avg: 64684.3 Avg: 66606 UDP_STREAM-10 : Avg: 52701.2 Avg: 52959.5 UDP_STREAM-20 : Avg: 30376.4 Avg: 29704 UDP_STREAM-40 : Avg: 15685.8 Avg: 15266.5 UDP_STREAM-80 : Avg: 8415.13 Avg: 7388.97 (wins and losses) sysbench: NO_WA_WEIGHT WA_WEIGHT sysbench-mysql-2 : 2135.17 per sec. 2142.51 per sec. sysbench-mysql-5 : 4809.68 per sec. 4800.19 per sec. sysbench-mysql-10 : 9158.59 per sec. 9157.05 per sec. sysbench-mysql-20 : 14570.70 per sec. 14543.55 per sec. sysbench-mysql-40 : 22130.56 per sec. 22184.82 per sec. sysbench-mysql-80 : 20995.56 per sec. 21904.18 per sec. sysbench-psql-2 : 1679.58 per sec. 1705.06 per sec. sysbench-psql-5 : 3797.69 per sec. 3879.93 per sec. sysbench-psql-10 : 7253.22 per sec. 7258.06 per sec. sysbench-psql-20 : 11166.75 per sec. 11220.00 per sec. sysbench-psql-40 : 17277.28 per sec. 17359.78 per sec. sysbench-psql-80 : 17112.44 per sec. 17221.16 per sec. (increase on the top end) tbench: NO_WA_WEIGHT Throughput 685.211 MB/sec 2 clients 2 procs max_latency=0.123 ms Throughput 1596.64 MB/sec 5 clients 5 procs max_latency=0.119 ms Throughput 2985.47 MB/sec 10 clients 10 procs max_latency=0.262 ms Throughput 4521.15 MB/sec 20 clients 20 procs max_latency=0.506 ms Throughput 9438.1 MB/sec 40 clients 40 procs max_latency=2.052 ms Throughput 8210.5 MB/sec 80 clients 80 procs max_latency=8.310 ms WA_WEIGHT Throughput 697.292 MB/sec 2 clients 2 procs max_latency=0.127 ms Throughput 1596.48 MB/sec 5 clients 5 procs max_latency=0.080 ms Throughput 2975.22 MB/sec 10 clients 10 procs max_latency=0.254 ms Throughput 4575.14 MB/sec 20 clients 20 procs max_latency=0.502 ms Throughput 9468.65 MB/sec 40 clients 40 procs max_latency=2.069 ms Throughput 8631.73 MB/sec 80 clients 80 procs max_latency=8.605 ms (increase on the top end) Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Rik van Riel <riel@redhat.com> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-10-06 07:23:24 +00:00
this_eff_load -= current_load;
sched/fair: Fix wake_affine() for !NUMA_BALANCING In commit: 3fed382b46ba ("sched/numa: Implement NUMA node level wake_affine()") Rik changed wake_affine to consider NUMA information when balancing between LLC domains. There are a number of problems here which this patch tries to address: - LLC < NODE; in this case we'd use the wrong information to balance - !NUMA_BALANCING: in this case, the new code doesn't do any balancing at all - re-computes the NUMA data for every wakeup, this can mean iterating up to 64 CPUs for every wakeup. - default affine wakeups inside a cache We address these by saving the load/capacity values for each sched_domain during regular load-balance and using these values in wake_affine_llc(). The obvious down-side to using cached values is that they can be too old and poorly reflect reality. But this way we can use LLC wide information and thus not rely on assuming LLC matches NODE. We also don't rely on NUMA_BALANCING nor do we have to aggegate two nodes (or even cache domains) worth of CPUs for each wakeup. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Fixes: 3fed382b46ba ("sched/numa: Implement NUMA node level wake_affine()") [ Minor readability improvements. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-07-31 15:50:05 +00:00
}
task_load = task_h_load(p);
sched/core: Address more wake_affine() regressions The trivial wake_affine_idle() implementation is very good for a number of workloads, but it comes apart at the moment there are no idle CPUs left, IOW. the overloaded case. hackbench: NO_WA_WEIGHT WA_WEIGHT hackbench-20 : 7.362717561 seconds 6.450509391 seconds (win) netperf: NO_WA_WEIGHT WA_WEIGHT TCP_SENDFILE-1 : Avg: 54524.6 Avg: 52224.3 TCP_SENDFILE-10 : Avg: 48185.2 Avg: 46504.3 TCP_SENDFILE-20 : Avg: 29031.2 Avg: 28610.3 TCP_SENDFILE-40 : Avg: 9819.72 Avg: 9253.12 TCP_SENDFILE-80 : Avg: 5355.3 Avg: 4687.4 TCP_STREAM-1 : Avg: 41448.3 Avg: 42254 TCP_STREAM-10 : Avg: 24123.2 Avg: 25847.9 TCP_STREAM-20 : Avg: 15834.5 Avg: 18374.4 TCP_STREAM-40 : Avg: 5583.91 Avg: 5599.57 TCP_STREAM-80 : Avg: 2329.66 Avg: 2726.41 TCP_RR-1 : Avg: 80473.5 Avg: 82638.8 TCP_RR-10 : Avg: 72660.5 Avg: 73265.1 TCP_RR-20 : Avg: 52607.1 Avg: 52634.5 TCP_RR-40 : Avg: 57199.2 Avg: 56302.3 TCP_RR-80 : Avg: 25330.3 Avg: 26867.9 UDP_RR-1 : Avg: 108266 Avg: 107844 UDP_RR-10 : Avg: 95480 Avg: 95245.2 UDP_RR-20 : Avg: 68770.8 Avg: 68673.7 UDP_RR-40 : Avg: 76231 Avg: 75419.1 UDP_RR-80 : Avg: 34578.3 Avg: 35639.1 UDP_STREAM-1 : Avg: 64684.3 Avg: 66606 UDP_STREAM-10 : Avg: 52701.2 Avg: 52959.5 UDP_STREAM-20 : Avg: 30376.4 Avg: 29704 UDP_STREAM-40 : Avg: 15685.8 Avg: 15266.5 UDP_STREAM-80 : Avg: 8415.13 Avg: 7388.97 (wins and losses) sysbench: NO_WA_WEIGHT WA_WEIGHT sysbench-mysql-2 : 2135.17 per sec. 2142.51 per sec. sysbench-mysql-5 : 4809.68 per sec. 4800.19 per sec. sysbench-mysql-10 : 9158.59 per sec. 9157.05 per sec. sysbench-mysql-20 : 14570.70 per sec. 14543.55 per sec. sysbench-mysql-40 : 22130.56 per sec. 22184.82 per sec. sysbench-mysql-80 : 20995.56 per sec. 21904.18 per sec. sysbench-psql-2 : 1679.58 per sec. 1705.06 per sec. sysbench-psql-5 : 3797.69 per sec. 3879.93 per sec. sysbench-psql-10 : 7253.22 per sec. 7258.06 per sec. sysbench-psql-20 : 11166.75 per sec. 11220.00 per sec. sysbench-psql-40 : 17277.28 per sec. 17359.78 per sec. sysbench-psql-80 : 17112.44 per sec. 17221.16 per sec. (increase on the top end) tbench: NO_WA_WEIGHT Throughput 685.211 MB/sec 2 clients 2 procs max_latency=0.123 ms Throughput 1596.64 MB/sec 5 clients 5 procs max_latency=0.119 ms Throughput 2985.47 MB/sec 10 clients 10 procs max_latency=0.262 ms Throughput 4521.15 MB/sec 20 clients 20 procs max_latency=0.506 ms Throughput 9438.1 MB/sec 40 clients 40 procs max_latency=2.052 ms Throughput 8210.5 MB/sec 80 clients 80 procs max_latency=8.310 ms WA_WEIGHT Throughput 697.292 MB/sec 2 clients 2 procs max_latency=0.127 ms Throughput 1596.48 MB/sec 5 clients 5 procs max_latency=0.080 ms Throughput 2975.22 MB/sec 10 clients 10 procs max_latency=0.254 ms Throughput 4575.14 MB/sec 20 clients 20 procs max_latency=0.502 ms Throughput 9468.65 MB/sec 40 clients 40 procs max_latency=2.069 ms Throughput 8631.73 MB/sec 80 clients 80 procs max_latency=8.605 ms (increase on the top end) Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Rik van Riel <riel@redhat.com> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-10-06 07:23:24 +00:00
this_eff_load += task_load;
if (sched_feat(WA_BIAS))
this_eff_load *= 100;
this_eff_load *= capacity_of(prev_cpu);
sched/fair: Fix wake_affine() for !NUMA_BALANCING In commit: 3fed382b46ba ("sched/numa: Implement NUMA node level wake_affine()") Rik changed wake_affine to consider NUMA information when balancing between LLC domains. There are a number of problems here which this patch tries to address: - LLC < NODE; in this case we'd use the wrong information to balance - !NUMA_BALANCING: in this case, the new code doesn't do any balancing at all - re-computes the NUMA data for every wakeup, this can mean iterating up to 64 CPUs for every wakeup. - default affine wakeups inside a cache We address these by saving the load/capacity values for each sched_domain during regular load-balance and using these values in wake_affine_llc(). The obvious down-side to using cached values is that they can be too old and poorly reflect reality. But this way we can use LLC wide information and thus not rely on assuming LLC matches NODE. We also don't rely on NUMA_BALANCING nor do we have to aggegate two nodes (or even cache domains) worth of CPUs for each wakeup. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Fixes: 3fed382b46ba ("sched/numa: Implement NUMA node level wake_affine()") [ Minor readability improvements. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-07-31 15:50:05 +00:00
prev_eff_load = cpu_load(cpu_rq(prev_cpu));
sched/core: Address more wake_affine() regressions The trivial wake_affine_idle() implementation is very good for a number of workloads, but it comes apart at the moment there are no idle CPUs left, IOW. the overloaded case. hackbench: NO_WA_WEIGHT WA_WEIGHT hackbench-20 : 7.362717561 seconds 6.450509391 seconds (win) netperf: NO_WA_WEIGHT WA_WEIGHT TCP_SENDFILE-1 : Avg: 54524.6 Avg: 52224.3 TCP_SENDFILE-10 : Avg: 48185.2 Avg: 46504.3 TCP_SENDFILE-20 : Avg: 29031.2 Avg: 28610.3 TCP_SENDFILE-40 : Avg: 9819.72 Avg: 9253.12 TCP_SENDFILE-80 : Avg: 5355.3 Avg: 4687.4 TCP_STREAM-1 : Avg: 41448.3 Avg: 42254 TCP_STREAM-10 : Avg: 24123.2 Avg: 25847.9 TCP_STREAM-20 : Avg: 15834.5 Avg: 18374.4 TCP_STREAM-40 : Avg: 5583.91 Avg: 5599.57 TCP_STREAM-80 : Avg: 2329.66 Avg: 2726.41 TCP_RR-1 : Avg: 80473.5 Avg: 82638.8 TCP_RR-10 : Avg: 72660.5 Avg: 73265.1 TCP_RR-20 : Avg: 52607.1 Avg: 52634.5 TCP_RR-40 : Avg: 57199.2 Avg: 56302.3 TCP_RR-80 : Avg: 25330.3 Avg: 26867.9 UDP_RR-1 : Avg: 108266 Avg: 107844 UDP_RR-10 : Avg: 95480 Avg: 95245.2 UDP_RR-20 : Avg: 68770.8 Avg: 68673.7 UDP_RR-40 : Avg: 76231 Avg: 75419.1 UDP_RR-80 : Avg: 34578.3 Avg: 35639.1 UDP_STREAM-1 : Avg: 64684.3 Avg: 66606 UDP_STREAM-10 : Avg: 52701.2 Avg: 52959.5 UDP_STREAM-20 : Avg: 30376.4 Avg: 29704 UDP_STREAM-40 : Avg: 15685.8 Avg: 15266.5 UDP_STREAM-80 : Avg: 8415.13 Avg: 7388.97 (wins and losses) sysbench: NO_WA_WEIGHT WA_WEIGHT sysbench-mysql-2 : 2135.17 per sec. 2142.51 per sec. sysbench-mysql-5 : 4809.68 per sec. 4800.19 per sec. sysbench-mysql-10 : 9158.59 per sec. 9157.05 per sec. sysbench-mysql-20 : 14570.70 per sec. 14543.55 per sec. sysbench-mysql-40 : 22130.56 per sec. 22184.82 per sec. sysbench-mysql-80 : 20995.56 per sec. 21904.18 per sec. sysbench-psql-2 : 1679.58 per sec. 1705.06 per sec. sysbench-psql-5 : 3797.69 per sec. 3879.93 per sec. sysbench-psql-10 : 7253.22 per sec. 7258.06 per sec. sysbench-psql-20 : 11166.75 per sec. 11220.00 per sec. sysbench-psql-40 : 17277.28 per sec. 17359.78 per sec. sysbench-psql-80 : 17112.44 per sec. 17221.16 per sec. (increase on the top end) tbench: NO_WA_WEIGHT Throughput 685.211 MB/sec 2 clients 2 procs max_latency=0.123 ms Throughput 1596.64 MB/sec 5 clients 5 procs max_latency=0.119 ms Throughput 2985.47 MB/sec 10 clients 10 procs max_latency=0.262 ms Throughput 4521.15 MB/sec 20 clients 20 procs max_latency=0.506 ms Throughput 9438.1 MB/sec 40 clients 40 procs max_latency=2.052 ms Throughput 8210.5 MB/sec 80 clients 80 procs max_latency=8.310 ms WA_WEIGHT Throughput 697.292 MB/sec 2 clients 2 procs max_latency=0.127 ms Throughput 1596.48 MB/sec 5 clients 5 procs max_latency=0.080 ms Throughput 2975.22 MB/sec 10 clients 10 procs max_latency=0.254 ms Throughput 4575.14 MB/sec 20 clients 20 procs max_latency=0.502 ms Throughput 9468.65 MB/sec 40 clients 40 procs max_latency=2.069 ms Throughput 8631.73 MB/sec 80 clients 80 procs max_latency=8.605 ms (increase on the top end) Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Rik van Riel <riel@redhat.com> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-10-06 07:23:24 +00:00
prev_eff_load -= task_load;
if (sched_feat(WA_BIAS))
prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
prev_eff_load *= capacity_of(this_cpu);
sched/fair: Fix wake_affine() for !NUMA_BALANCING In commit: 3fed382b46ba ("sched/numa: Implement NUMA node level wake_affine()") Rik changed wake_affine to consider NUMA information when balancing between LLC domains. There are a number of problems here which this patch tries to address: - LLC < NODE; in this case we'd use the wrong information to balance - !NUMA_BALANCING: in this case, the new code doesn't do any balancing at all - re-computes the NUMA data for every wakeup, this can mean iterating up to 64 CPUs for every wakeup. - default affine wakeups inside a cache We address these by saving the load/capacity values for each sched_domain during regular load-balance and using these values in wake_affine_llc(). The obvious down-side to using cached values is that they can be too old and poorly reflect reality. But this way we can use LLC wide information and thus not rely on assuming LLC matches NODE. We also don't rely on NUMA_BALANCING nor do we have to aggegate two nodes (or even cache domains) worth of CPUs for each wakeup. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Fixes: 3fed382b46ba ("sched/numa: Implement NUMA node level wake_affine()") [ Minor readability improvements. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-07-31 15:50:05 +00:00
sched/fair: Do not migrate on wake_affine_weight() if weights are equal wake_affine_weight() will consider migrating a task to, or near, the current CPU if there is a load imbalance. If the CPUs share LLC then either CPU is valid as a search-for-idle-sibling target and equally appropriate for stacking two tasks on one CPU if an idle sibling is unavailable. If they do not share cache then a cross-node migration potentially impacts locality so while they are equal from a CPU capacity point of view, they are not equal in terms of memory locality. In either case, it's more appropriate to migrate only if there is a difference in their effective load. This patch modifies wake_affine_weight() to only consider migrating a task if there is a load imbalance for normal wakeups but will allow potential stacking if the loads are equal and it's a sync wakeup. For the most part, the different in performance is marginal. For example, on a 4-socket server running netperf UDP_STREAM on localhost the differences are as follows: 4.15.0 4.15.0 16rc0 noequal-v1r23 Hmean send-64 355.47 ( 0.00%) 349.50 ( -1.68%) Hmean send-128 697.98 ( 0.00%) 693.35 ( -0.66%) Hmean send-256 1328.02 ( 0.00%) 1318.77 ( -0.70%) Hmean send-1024 5051.83 ( 0.00%) 5051.11 ( -0.01%) Hmean send-2048 9637.02 ( 0.00%) 9601.34 ( -0.37%) Hmean send-3312 14355.37 ( 0.00%) 14414.51 ( 0.41%) Hmean send-4096 16464.97 ( 0.00%) 16301.37 ( -0.99%) Hmean send-8192 26722.42 ( 0.00%) 26428.95 ( -1.10%) Hmean send-16384 38137.81 ( 0.00%) 38046.11 ( -0.24%) Hmean recv-64 355.47 ( 0.00%) 349.50 ( -1.68%) Hmean recv-128 697.98 ( 0.00%) 693.35 ( -0.66%) Hmean recv-256 1328.02 ( 0.00%) 1318.77 ( -0.70%) Hmean recv-1024 5051.83 ( 0.00%) 5051.11 ( -0.01%) Hmean recv-2048 9636.95 ( 0.00%) 9601.30 ( -0.37%) Hmean recv-3312 14355.32 ( 0.00%) 14414.48 ( 0.41%) Hmean recv-4096 16464.74 ( 0.00%) 16301.16 ( -0.99%) Hmean recv-8192 26721.63 ( 0.00%) 26428.17 ( -1.10%) Hmean recv-16384 38136.00 ( 0.00%) 38044.88 ( -0.24%) Stddev send-64 7.30 ( 0.00%) 4.75 ( 34.96%) Stddev send-128 15.15 ( 0.00%) 22.38 ( -47.66%) Stddev send-256 13.99 ( 0.00%) 19.14 ( -36.81%) Stddev send-1024 105.73 ( 0.00%) 67.38 ( 36.27%) Stddev send-2048 294.57 ( 0.00%) 223.88 ( 24.00%) Stddev send-3312 302.28 ( 0.00%) 271.74 ( 10.10%) Stddev send-4096 195.92 ( 0.00%) 121.10 ( 38.19%) Stddev send-8192 399.71 ( 0.00%) 563.77 ( -41.04%) Stddev send-16384 1163.47 ( 0.00%) 1103.68 ( 5.14%) Stddev recv-64 7.30 ( 0.00%) 4.75 ( 34.96%) Stddev recv-128 15.15 ( 0.00%) 22.38 ( -47.66%) Stddev recv-256 13.99 ( 0.00%) 19.14 ( -36.81%) Stddev recv-1024 105.73 ( 0.00%) 67.38 ( 36.27%) Stddev recv-2048 294.59 ( 0.00%) 223.89 ( 24.00%) Stddev recv-3312 302.24 ( 0.00%) 271.75 ( 10.09%) Stddev recv-4096 196.03 ( 0.00%) 121.14 ( 38.20%) Stddev recv-8192 399.86 ( 0.00%) 563.65 ( -40.96%) Stddev recv-16384 1163.79 ( 0.00%) 1103.86 ( 5.15%) The difference in overall performance is marginal but note that most measurements are less variable. There were similar observations for other netperf comparisons. hackbench with sockets or threads with processes or threads showed minor difference with some reduction of migration. tbench showed only marginal differences that were within the noise. dbench, regardless of filesystem, showed minor differences all of which are within noise. Multiple machines, both UMA and NUMA were tested without any regressions showing up. The biggest risk with a patch like this is affecting wakeup latencies. However, the schbench load from Facebook which is very sensitive to wakeup latency showed a mixed result with mostly improvements in wakeup latency: 4.15.0 4.15.0 16rc0 noequal-v1r23 Lat 50.00th-qrtle-1 38.00 ( 0.00%) 38.00 ( 0.00%) Lat 75.00th-qrtle-1 49.00 ( 0.00%) 41.00 ( 16.33%) Lat 90.00th-qrtle-1 52.00 ( 0.00%) 50.00 ( 3.85%) Lat 95.00th-qrtle-1 54.00 ( 0.00%) 51.00 ( 5.56%) Lat 99.00th-qrtle-1 63.00 ( 0.00%) 60.00 ( 4.76%) Lat 99.50th-qrtle-1 66.00 ( 0.00%) 61.00 ( 7.58%) Lat 99.90th-qrtle-1 78.00 ( 0.00%) 65.00 ( 16.67%) Lat 50.00th-qrtle-2 38.00 ( 0.00%) 38.00 ( 0.00%) Lat 75.00th-qrtle-2 42.00 ( 0.00%) 43.00 ( -2.38%) Lat 90.00th-qrtle-2 46.00 ( 0.00%) 48.00 ( -4.35%) Lat 95.00th-qrtle-2 49.00 ( 0.00%) 50.00 ( -2.04%) Lat 99.00th-qrtle-2 55.00 ( 0.00%) 57.00 ( -3.64%) Lat 99.50th-qrtle-2 58.00 ( 0.00%) 60.00 ( -3.45%) Lat 99.90th-qrtle-2 65.00 ( 0.00%) 68.00 ( -4.62%) Lat 50.00th-qrtle-4 41.00 ( 0.00%) 41.00 ( 0.00%) Lat 75.00th-qrtle-4 45.00 ( 0.00%) 46.00 ( -2.22%) Lat 90.00th-qrtle-4 50.00 ( 0.00%) 50.00 ( 0.00%) Lat 95.00th-qrtle-4 54.00 ( 0.00%) 53.00 ( 1.85%) Lat 99.00th-qrtle-4 61.00 ( 0.00%) 61.00 ( 0.00%) Lat 99.50th-qrtle-4 65.00 ( 0.00%) 64.00 ( 1.54%) Lat 99.90th-qrtle-4 76.00 ( 0.00%) 82.00 ( -7.89%) Lat 50.00th-qrtle-8 48.00 ( 0.00%) 46.00 ( 4.17%) Lat 75.00th-qrtle-8 55.00 ( 0.00%) 54.00 ( 1.82%) Lat 90.00th-qrtle-8 60.00 ( 0.00%) 59.00 ( 1.67%) Lat 95.00th-qrtle-8 63.00 ( 0.00%) 63.00 ( 0.00%) Lat 99.00th-qrtle-8 71.00 ( 0.00%) 69.00 ( 2.82%) Lat 99.50th-qrtle-8 74.00 ( 0.00%) 73.00 ( 1.35%) Lat 99.90th-qrtle-8 98.00 ( 0.00%) 90.00 ( 8.16%) Lat 50.00th-qrtle-16 56.00 ( 0.00%) 55.00 ( 1.79%) Lat 75.00th-qrtle-16 68.00 ( 0.00%) 67.00 ( 1.47%) Lat 90.00th-qrtle-16 77.00 ( 0.00%) 78.00 ( -1.30%) Lat 95.00th-qrtle-16 82.00 ( 0.00%) 84.00 ( -2.44%) Lat 99.00th-qrtle-16 90.00 ( 0.00%) 93.00 ( -3.33%) Lat 99.50th-qrtle-16 93.00 ( 0.00%) 97.00 ( -4.30%) Lat 99.90th-qrtle-16 110.00 ( 0.00%) 110.00 ( 0.00%) Lat 50.00th-qrtle-32 68.00 ( 0.00%) 62.00 ( 8.82%) Lat 75.00th-qrtle-32 90.00 ( 0.00%) 83.00 ( 7.78%) Lat 90.00th-qrtle-32 110.00 ( 0.00%) 100.00 ( 9.09%) Lat 95.00th-qrtle-32 122.00 ( 0.00%) 111.00 ( 9.02%) Lat 99.00th-qrtle-32 145.00 ( 0.00%) 133.00 ( 8.28%) Lat 99.50th-qrtle-32 154.00 ( 0.00%) 143.00 ( 7.14%) Lat 99.90th-qrtle-32 2316.00 ( 0.00%) 515.00 ( 77.76%) Lat 50.00th-qrtle-35 69.00 ( 0.00%) 72.00 ( -4.35%) Lat 75.00th-qrtle-35 92.00 ( 0.00%) 95.00 ( -3.26%) Lat 90.00th-qrtle-35 111.00 ( 0.00%) 114.00 ( -2.70%) Lat 95.00th-qrtle-35 122.00 ( 0.00%) 124.00 ( -1.64%) Lat 99.00th-qrtle-35 142.00 ( 0.00%) 144.00 ( -1.41%) Lat 99.50th-qrtle-35 150.00 ( 0.00%) 154.00 ( -2.67%) Lat 99.90th-qrtle-35 6104.00 ( 0.00%) 5640.00 ( 7.60%) Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Giovanni Gherdovich <ggherdovich@suse.cz> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Matt Fleming <matt@codeblueprint.co.uk> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/20180213133730.24064-4-mgorman@techsingularity.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-02-13 13:37:27 +00:00
/*
* If sync, adjust the weight of prev_eff_load such that if
* prev_eff == this_eff that select_idle_sibling() will consider
* stacking the wakee on top of the waker if no other CPU is
* idle.
*/
if (sync)
prev_eff_load += 1;
return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
sched/fair: Fix wake_affine() for !NUMA_BALANCING In commit: 3fed382b46ba ("sched/numa: Implement NUMA node level wake_affine()") Rik changed wake_affine to consider NUMA information when balancing between LLC domains. There are a number of problems here which this patch tries to address: - LLC < NODE; in this case we'd use the wrong information to balance - !NUMA_BALANCING: in this case, the new code doesn't do any balancing at all - re-computes the NUMA data for every wakeup, this can mean iterating up to 64 CPUs for every wakeup. - default affine wakeups inside a cache We address these by saving the load/capacity values for each sched_domain during regular load-balance and using these values in wake_affine_llc(). The obvious down-side to using cached values is that they can be too old and poorly reflect reality. But this way we can use LLC wide information and thus not rely on assuming LLC matches NODE. We also don't rely on NUMA_BALANCING nor do we have to aggegate two nodes (or even cache domains) worth of CPUs for each wakeup. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Fixes: 3fed382b46ba ("sched/numa: Implement NUMA node level wake_affine()") [ Minor readability improvements. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-07-31 15:50:05 +00:00
}
static int wake_affine(struct sched_domain *sd, struct task_struct *p,
int this_cpu, int prev_cpu, int sync)
{
int target = nr_cpumask_bits;
if (sched_feat(WA_IDLE))
target = wake_affine_idle(this_cpu, prev_cpu, sync);
sched/fair: Fix wake_affine() for !NUMA_BALANCING In commit: 3fed382b46ba ("sched/numa: Implement NUMA node level wake_affine()") Rik changed wake_affine to consider NUMA information when balancing between LLC domains. There are a number of problems here which this patch tries to address: - LLC < NODE; in this case we'd use the wrong information to balance - !NUMA_BALANCING: in this case, the new code doesn't do any balancing at all - re-computes the NUMA data for every wakeup, this can mean iterating up to 64 CPUs for every wakeup. - default affine wakeups inside a cache We address these by saving the load/capacity values for each sched_domain during regular load-balance and using these values in wake_affine_llc(). The obvious down-side to using cached values is that they can be too old and poorly reflect reality. But this way we can use LLC wide information and thus not rely on assuming LLC matches NODE. We also don't rely on NUMA_BALANCING nor do we have to aggegate two nodes (or even cache domains) worth of CPUs for each wakeup. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Fixes: 3fed382b46ba ("sched/numa: Implement NUMA node level wake_affine()") [ Minor readability improvements. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-07-31 15:50:05 +00:00
if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
if (target == nr_cpumask_bits)
return prev_cpu;
schedstat_inc(sd->ttwu_move_affine);
schedstat_inc(p->se.statistics.nr_wakeups_affine);
return target;
}
static struct sched_group *
find_idlest_group(struct sched_domain *sd, struct task_struct *p,
int this_cpu, int sd_flag);
/*
* find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
*/
static int
find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
{
unsigned long load, min_load = ULONG_MAX;
sched/fair: Leverage the idle state info when choosing the "idlest" cpu The code in find_idlest_cpu() looks for the CPU with the smallest load. However, if multiple CPUs are idle, the first idle CPU is selected irrespective of the depth of its idle state. Among the idle CPUs we should pick the one with with the shallowest idle state, or the latest to have gone idle if all idle CPUs are in the same state. The later applies even when cpuidle is configured out. This patch doesn't cover the following issues: - The idle exit latency of a CPU might be larger than the time needed to migrate the waking task to an already running CPU with sufficient capacity, and therefore performance would benefit from task packing in such case (in most cases task packing is about power saving). - Some idle states have a non negligible and non abortable entry latency which needs to run to completion before the exit latency can start. A concurrent patch series is making this info available to the cpuidle core. Once available, the entry latency with the idle timestamp could determine when the exit latency may be effective. Those issues will be handled in due course. In the mean time, what is implemented here should improve things already compared to the current state of affairs. Based on an initial patch from Daniel Lezcano. Signed-off-by: Nicolas Pitre <nico@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Daniel Lezcano <daniel.lezcano@linaro.org> Cc: "Rafael J. Wysocki" <rjw@rjwysocki.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: linux-pm@vger.kernel.org Cc: linaro-kernel@lists.linaro.org Link: http://lkml.kernel.org/n/tip-@git.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-09-04 15:32:10 +00:00
unsigned int min_exit_latency = UINT_MAX;
u64 latest_idle_timestamp = 0;
int least_loaded_cpu = this_cpu;
sched/fair: Make sched-idle CPU selection consistent throughout There are instances where we keep searching for an idle CPU despite already having a sched-idle CPU (in find_idlest_group_cpu(), select_idle_smt() and select_idle_cpu() and then there are places where we don't necessarily do that and return a sched-idle CPU as soon as we find one (in select_idle_sibling()). This looks a bit inconsistent and it may be worth having the same policy everywhere. On the other hand, choosing a sched-idle CPU over a idle one shall be beneficial from performance and power point of view as well, as we don't need to get the CPU online from a deep idle state which wastes quite a lot of time and energy and delays the scheduling of the newly woken up task. This patch tries to simplify code around sched-idle CPU selection and make it consistent throughout. Testing is done with the help of rt-app on hikey board (ARM64 octa-core, 2 clusters, 0-3 and 4-7). The cpufreq governor was set to performance to avoid any side affects from CPU frequency. Following are the tests performed: Test 1: 1-cfs-task: A single SCHED_NORMAL task is pinned to CPU5 which runs for 2333 us out of 7777 us (so gives time for the cluster to go in deep idle state). Test 2: 1-cfs-1-idle-task: A single SCHED_NORMAL task is pinned on CPU5 and single SCHED_IDLE task is pinned on CPU6 (to make sure cluster 1 doesn't go in deep idle state). Test 3: 1-cfs-8-idle-task: A single SCHED_NORMAL task is pinned on CPU5 and eight SCHED_IDLE tasks are created which run forever (not pinned anywhere, so they run on all CPUs). Checked with kernelshark that as soon as NORMAL task sleeps, the SCHED_IDLE task starts running on CPU5. And here are the results on mean latency (in us), using the "st" tool. $ st 1-cfs-task/rt-app-cfs_thread-0.log N min max sum mean stddev 642 90 592 197180 307.134 109.906 $ st 1-cfs-1-idle-task/rt-app-cfs_thread-0.log N min max sum mean stddev 642 67 311 113850 177.336 41.4251 $ st 1-cfs-8-idle-task/rt-app-cfs_thread-0.log N min max sum mean stddev 643 29 173 41364 64.3297 13.2344 The mean latency when we need to: - wakeup from deep idle state is 307 us. - wakeup from shallow idle state is 177 us. - preempt a SCHED_IDLE task is 64 us. Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: https://lkml.kernel.org/r/b90cbcce608cef4e02a7bbfe178335f76d201bab.1573728344.git.viresh.kumar@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-11-14 10:49:27 +00:00
int shallowest_idle_cpu = -1;
int i;
/* Check if we have any choice: */
if (group->group_weight == 1)
return cpumask_first(sched_group_span(group));
/* Traverse only the allowed CPUs */
for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) {
sched/fair: Make sched-idle CPU selection consistent throughout There are instances where we keep searching for an idle CPU despite already having a sched-idle CPU (in find_idlest_group_cpu(), select_idle_smt() and select_idle_cpu() and then there are places where we don't necessarily do that and return a sched-idle CPU as soon as we find one (in select_idle_sibling()). This looks a bit inconsistent and it may be worth having the same policy everywhere. On the other hand, choosing a sched-idle CPU over a idle one shall be beneficial from performance and power point of view as well, as we don't need to get the CPU online from a deep idle state which wastes quite a lot of time and energy and delays the scheduling of the newly woken up task. This patch tries to simplify code around sched-idle CPU selection and make it consistent throughout. Testing is done with the help of rt-app on hikey board (ARM64 octa-core, 2 clusters, 0-3 and 4-7). The cpufreq governor was set to performance to avoid any side affects from CPU frequency. Following are the tests performed: Test 1: 1-cfs-task: A single SCHED_NORMAL task is pinned to CPU5 which runs for 2333 us out of 7777 us (so gives time for the cluster to go in deep idle state). Test 2: 1-cfs-1-idle-task: A single SCHED_NORMAL task is pinned on CPU5 and single SCHED_IDLE task is pinned on CPU6 (to make sure cluster 1 doesn't go in deep idle state). Test 3: 1-cfs-8-idle-task: A single SCHED_NORMAL task is pinned on CPU5 and eight SCHED_IDLE tasks are created which run forever (not pinned anywhere, so they run on all CPUs). Checked with kernelshark that as soon as NORMAL task sleeps, the SCHED_IDLE task starts running on CPU5. And here are the results on mean latency (in us), using the "st" tool. $ st 1-cfs-task/rt-app-cfs_thread-0.log N min max sum mean stddev 642 90 592 197180 307.134 109.906 $ st 1-cfs-1-idle-task/rt-app-cfs_thread-0.log N min max sum mean stddev 642 67 311 113850 177.336 41.4251 $ st 1-cfs-8-idle-task/rt-app-cfs_thread-0.log N min max sum mean stddev 643 29 173 41364 64.3297 13.2344 The mean latency when we need to: - wakeup from deep idle state is 307 us. - wakeup from shallow idle state is 177 us. - preempt a SCHED_IDLE task is 64 us. Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: https://lkml.kernel.org/r/b90cbcce608cef4e02a7bbfe178335f76d201bab.1573728344.git.viresh.kumar@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-11-14 10:49:27 +00:00
if (sched_idle_cpu(i))
return i;
if (available_idle_cpu(i)) {
sched/fair: Leverage the idle state info when choosing the "idlest" cpu The code in find_idlest_cpu() looks for the CPU with the smallest load. However, if multiple CPUs are idle, the first idle CPU is selected irrespective of the depth of its idle state. Among the idle CPUs we should pick the one with with the shallowest idle state, or the latest to have gone idle if all idle CPUs are in the same state. The later applies even when cpuidle is configured out. This patch doesn't cover the following issues: - The idle exit latency of a CPU might be larger than the time needed to migrate the waking task to an already running CPU with sufficient capacity, and therefore performance would benefit from task packing in such case (in most cases task packing is about power saving). - Some idle states have a non negligible and non abortable entry latency which needs to run to completion before the exit latency can start. A concurrent patch series is making this info available to the cpuidle core. Once available, the entry latency with the idle timestamp could determine when the exit latency may be effective. Those issues will be handled in due course. In the mean time, what is implemented here should improve things already compared to the current state of affairs. Based on an initial patch from Daniel Lezcano. Signed-off-by: Nicolas Pitre <nico@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Daniel Lezcano <daniel.lezcano@linaro.org> Cc: "Rafael J. Wysocki" <rjw@rjwysocki.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: linux-pm@vger.kernel.org Cc: linaro-kernel@lists.linaro.org Link: http://lkml.kernel.org/n/tip-@git.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-09-04 15:32:10 +00:00
struct rq *rq = cpu_rq(i);
struct cpuidle_state *idle = idle_get_state(rq);
if (idle && idle->exit_latency < min_exit_latency) {
/*
* We give priority to a CPU whose idle state
* has the smallest exit latency irrespective
* of any idle timestamp.
*/
min_exit_latency = idle->exit_latency;
latest_idle_timestamp = rq->idle_stamp;
shallowest_idle_cpu = i;
} else if ((!idle || idle->exit_latency == min_exit_latency) &&
rq->idle_stamp > latest_idle_timestamp) {
/*
* If equal or no active idle state, then
* the most recently idled CPU might have
* a warmer cache.
*/
latest_idle_timestamp = rq->idle_stamp;
shallowest_idle_cpu = i;
}
sched/fair: Make sched-idle CPU selection consistent throughout There are instances where we keep searching for an idle CPU despite already having a sched-idle CPU (in find_idlest_group_cpu(), select_idle_smt() and select_idle_cpu() and then there are places where we don't necessarily do that and return a sched-idle CPU as soon as we find one (in select_idle_sibling()). This looks a bit inconsistent and it may be worth having the same policy everywhere. On the other hand, choosing a sched-idle CPU over a idle one shall be beneficial from performance and power point of view as well, as we don't need to get the CPU online from a deep idle state which wastes quite a lot of time and energy and delays the scheduling of the newly woken up task. This patch tries to simplify code around sched-idle CPU selection and make it consistent throughout. Testing is done with the help of rt-app on hikey board (ARM64 octa-core, 2 clusters, 0-3 and 4-7). The cpufreq governor was set to performance to avoid any side affects from CPU frequency. Following are the tests performed: Test 1: 1-cfs-task: A single SCHED_NORMAL task is pinned to CPU5 which runs for 2333 us out of 7777 us (so gives time for the cluster to go in deep idle state). Test 2: 1-cfs-1-idle-task: A single SCHED_NORMAL task is pinned on CPU5 and single SCHED_IDLE task is pinned on CPU6 (to make sure cluster 1 doesn't go in deep idle state). Test 3: 1-cfs-8-idle-task: A single SCHED_NORMAL task is pinned on CPU5 and eight SCHED_IDLE tasks are created which run forever (not pinned anywhere, so they run on all CPUs). Checked with kernelshark that as soon as NORMAL task sleeps, the SCHED_IDLE task starts running on CPU5. And here are the results on mean latency (in us), using the "st" tool. $ st 1-cfs-task/rt-app-cfs_thread-0.log N min max sum mean stddev 642 90 592 197180 307.134 109.906 $ st 1-cfs-1-idle-task/rt-app-cfs_thread-0.log N min max sum mean stddev 642 67 311 113850 177.336 41.4251 $ st 1-cfs-8-idle-task/rt-app-cfs_thread-0.log N min max sum mean stddev 643 29 173 41364 64.3297 13.2344 The mean latency when we need to: - wakeup from deep idle state is 307 us. - wakeup from shallow idle state is 177 us. - preempt a SCHED_IDLE task is 64 us. Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: https://lkml.kernel.org/r/b90cbcce608cef4e02a7bbfe178335f76d201bab.1573728344.git.viresh.kumar@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-11-14 10:49:27 +00:00
} else if (shallowest_idle_cpu == -1) {
load = cpu_load(cpu_rq(i));
sched/fair: Remove impossible condition from find_idlest_group_cpu() find_idlest_group_cpu() goes through CPUs of a group previous selected by find_idlest_group(). find_idlest_group() returns NULL if the local group is the selected one and doesn't execute find_idlest_group_cpu if the group to which 'cpu' belongs to is chosen. So we're always guaranteed to call find_idlest_group_cpu() with a group to which 'cpu' is non-local. This makes one of the conditions in find_idlest_group_cpu() an impossible one, which we can get rid off. Signed-off-by: Joel Fernandes <joelaf@google.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Brendan Jackman <brendan.jackman@arm.com> Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Android Kernel <kernel-team@android.com> Cc: Atish Patra <atish.patra@oracle.com> Cc: Chris Redpath <Chris.Redpath@arm.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: EAS Dev <eas-dev@lists.linaro.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Cc: Josef Bacik <jbacik@fb.com> Cc: Juri Lelli <juri.lelli@arm.com> Cc: Len Brown <lenb@kernel.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten Ramussen <morten.rasmussen@arm.com> Cc: Patrick Bellasi <patrick.bellasi@arm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rafael J. Wysocki <rjw@rjwysocki.net> Cc: Rohit Jain <rohit.k.jain@oracle.com> Cc: Saravana Kannan <skannan@quicinc.com> Cc: Srinivas Pandruvada <srinivas.pandruvada@linux.intel.com> Cc: Steve Muckle <smuckle@google.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vikram Mulukutla <markivx@codeaurora.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Link: http://lkml.kernel.org/r/20171215153944.220146-3-joelaf@google.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-12-15 15:39:44 +00:00
if (load < min_load) {
sched/fair: Leverage the idle state info when choosing the "idlest" cpu The code in find_idlest_cpu() looks for the CPU with the smallest load. However, if multiple CPUs are idle, the first idle CPU is selected irrespective of the depth of its idle state. Among the idle CPUs we should pick the one with with the shallowest idle state, or the latest to have gone idle if all idle CPUs are in the same state. The later applies even when cpuidle is configured out. This patch doesn't cover the following issues: - The idle exit latency of a CPU might be larger than the time needed to migrate the waking task to an already running CPU with sufficient capacity, and therefore performance would benefit from task packing in such case (in most cases task packing is about power saving). - Some idle states have a non negligible and non abortable entry latency which needs to run to completion before the exit latency can start. A concurrent patch series is making this info available to the cpuidle core. Once available, the entry latency with the idle timestamp could determine when the exit latency may be effective. Those issues will be handled in due course. In the mean time, what is implemented here should improve things already compared to the current state of affairs. Based on an initial patch from Daniel Lezcano. Signed-off-by: Nicolas Pitre <nico@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Daniel Lezcano <daniel.lezcano@linaro.org> Cc: "Rafael J. Wysocki" <rjw@rjwysocki.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: linux-pm@vger.kernel.org Cc: linaro-kernel@lists.linaro.org Link: http://lkml.kernel.org/n/tip-@git.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-09-04 15:32:10 +00:00
min_load = load;
least_loaded_cpu = i;
}
}
}
sched/fair: Make sched-idle CPU selection consistent throughout There are instances where we keep searching for an idle CPU despite already having a sched-idle CPU (in find_idlest_group_cpu(), select_idle_smt() and select_idle_cpu() and then there are places where we don't necessarily do that and return a sched-idle CPU as soon as we find one (in select_idle_sibling()). This looks a bit inconsistent and it may be worth having the same policy everywhere. On the other hand, choosing a sched-idle CPU over a idle one shall be beneficial from performance and power point of view as well, as we don't need to get the CPU online from a deep idle state which wastes quite a lot of time and energy and delays the scheduling of the newly woken up task. This patch tries to simplify code around sched-idle CPU selection and make it consistent throughout. Testing is done with the help of rt-app on hikey board (ARM64 octa-core, 2 clusters, 0-3 and 4-7). The cpufreq governor was set to performance to avoid any side affects from CPU frequency. Following are the tests performed: Test 1: 1-cfs-task: A single SCHED_NORMAL task is pinned to CPU5 which runs for 2333 us out of 7777 us (so gives time for the cluster to go in deep idle state). Test 2: 1-cfs-1-idle-task: A single SCHED_NORMAL task is pinned on CPU5 and single SCHED_IDLE task is pinned on CPU6 (to make sure cluster 1 doesn't go in deep idle state). Test 3: 1-cfs-8-idle-task: A single SCHED_NORMAL task is pinned on CPU5 and eight SCHED_IDLE tasks are created which run forever (not pinned anywhere, so they run on all CPUs). Checked with kernelshark that as soon as NORMAL task sleeps, the SCHED_IDLE task starts running on CPU5. And here are the results on mean latency (in us), using the "st" tool. $ st 1-cfs-task/rt-app-cfs_thread-0.log N min max sum mean stddev 642 90 592 197180 307.134 109.906 $ st 1-cfs-1-idle-task/rt-app-cfs_thread-0.log N min max sum mean stddev 642 67 311 113850 177.336 41.4251 $ st 1-cfs-8-idle-task/rt-app-cfs_thread-0.log N min max sum mean stddev 643 29 173 41364 64.3297 13.2344 The mean latency when we need to: - wakeup from deep idle state is 307 us. - wakeup from shallow idle state is 177 us. - preempt a SCHED_IDLE task is 64 us. Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: https://lkml.kernel.org/r/b90cbcce608cef4e02a7bbfe178335f76d201bab.1573728344.git.viresh.kumar@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-11-14 10:49:27 +00:00
return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
}
static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
int cpu, int prev_cpu, int sd_flag)
{
int new_cpu = cpu;
if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr))
return prev_cpu;
/*
* We need task's util for cpu_util_without, sync it up to
sched/fair: Fix cpu_util_wake() for 'execl' type workloads A ~10% regression has been reported for UnixBench's execl throughput test by Aaron Lu and Ye Xiaolong: https://lkml.org/lkml/2018/10/30/765 That test is pretty simple, it does a "recursive" execve() syscall on the same binary. Starting from the syscall, this sequence is possible: do_execve() do_execveat_common() __do_execve_file() sched_exec() select_task_rq_fair() <==| Task already enqueued find_idlest_cpu() find_idlest_group() capacity_spare_wake() <==| Functions not called from cpu_util_wake() | the wakeup path which means we can end up calling cpu_util_wake() not only from the "wakeup path", as its name would suggest. Indeed, the task doing an execve() syscall is already enqueued on the CPU we want to get the cpu_util_wake() for. The estimated utilization for a CPU computed in cpu_util_wake() was written under the assumption that function can be called only from the wakeup path. If instead the task is already enqueued, we end up with a utilization which does not remove the current task's contribution from the estimated utilization of the CPU. This will wrongly assume a reduced spare capacity on the current CPU and increase the chances to migrate the task on execve. The regression is tracked down to: commit d519329f72a6 ("sched/fair: Update util_est only on util_avg updates") because in that patch we turn on by default the UTIL_EST sched feature. However, the real issue is introduced by: commit f9be3e5961c5 ("sched/fair: Use util_est in LB and WU paths") Let's fix this by ensuring to always discount the task estimated utilization from the CPU's estimated utilization when the task is also the current one. The same benchmark of the bug report, executed on a dual socket 40 CPUs Intel(R) Xeon(R) CPU E5-2690 v2 @ 3.00GHz machine, reports these "Execl Throughput" figures (higher the better): mainline : 48136.5 lps mainline+fix : 55376.5 lps which correspond to a 15% speedup. Moreover, since {cpu_util,capacity_spare}_wake() are not really only used from the wakeup path, let's remove this ambiguity by using a better matching name: {cpu_util,capacity_spare}_without(). Since we are at that, let's also improve the existing documentation. Reported-by: Aaron Lu <aaron.lu@intel.com> Reported-by: Ye Xiaolong <xiaolong.ye@intel.com> Tested-by: Aaron Lu <aaron.lu@intel.com> Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Quentin Perret <quentin.perret@arm.com> Cc: Steve Muckle <smuckle@google.com> Cc: Suren Baghdasaryan <surenb@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Todd Kjos <tkjos@google.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Fixes: f9be3e5961c5 (sched/fair: Use util_est in LB and WU paths) Link: https://lore.kernel.org/lkml/20181025093100.GB13236@e110439-lin/ Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-11-05 14:53:58 +00:00
* prev_cpu's last_update_time.
*/
if (!(sd_flag & SD_BALANCE_FORK))
sync_entity_load_avg(&p->se);
while (sd) {
struct sched_group *group;
struct sched_domain *tmp;
int weight;
if (!(sd->flags & sd_flag)) {
sd = sd->child;
continue;
}
group = find_idlest_group(sd, p, cpu, sd_flag);
if (!group) {
sd = sd->child;
continue;
}
new_cpu = find_idlest_group_cpu(group, p, cpu);
if (new_cpu == cpu) {
/* Now try balancing at a lower domain level of 'cpu': */
sd = sd->child;
continue;
}
/* Now try balancing at a lower domain level of 'new_cpu': */
cpu = new_cpu;
weight = sd->span_weight;
sd = NULL;
for_each_domain(cpu, tmp) {
if (weight <= tmp->span_weight)
break;
if (tmp->flags & sd_flag)
sd = tmp;
}
}
return new_cpu;
}
sched/core: Rewrite and improve select_idle_siblings() select_idle_siblings() is a known pain point for a number of workloads; it either does too much or not enough and sometimes just does plain wrong. This rewrite attempts to address a number of issues (but sadly not all). The current code does an unconditional sched_domain iteration; with the intent of finding an idle core (on SMT hardware). The problems which this patch tries to address are: - its pointless to look for idle cores if the machine is real busy; at which point you're just wasting cycles. - it's behaviour is inconsistent between SMT and !SMT hardware in that !SMT hardware ends up doing a scan for any idle CPU in the LLC domain, while SMT hardware does a scan for idle cores and if that fails, falls back to a scan for idle threads on the 'target' core. The new code replaces the sched_domain scan with 3 explicit scans: 1) search for an idle core in the LLC 2) search for an idle CPU in the LLC 3) search for an idle thread in the 'target' core where 1 and 3 are conditional on SMT support and 1 and 2 have runtime heuristics to skip the step. Step 1) is conditional on sd_llc_shared->has_idle_cores; when a cpu goes idle and sd_llc_shared->has_idle_cores is false, we scan all SMT siblings of the CPU going idle. Similarly, we clear sd_llc_shared->has_idle_cores when we fail to find an idle core. Step 2) tracks the average cost of the scan and compares this to the average idle time guestimate for the CPU doing the wakeup. There is a significant fudge factor involved to deal with the variability of the averages. Esp. hackbench was sensitive to this. Step 3) is unconditional; we assume (also per step 1) that scanning all SMT siblings in a core is 'cheap'. With this; SMT systems gain step 2, which cures a few benchmarks -- notably one from Facebook. One 'feature' of the sched_domain iteration, which we preserve in the new code, is that it would start scanning from the 'target' CPU, instead of scanning the cpumask in cpu id order. This avoids multiple CPUs in the LLC scanning for idle to gang up and find the same CPU quite as much. The down side is that tasks can end up hopping across the LLC for no apparent reason. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-05-09 08:38:05 +00:00
#ifdef CONFIG_SCHED_SMT
DEFINE_STATIC_KEY_FALSE(sched_smt_present);
cpu/hotplug: Fix "SMT disabled by BIOS" detection for KVM With the following commit: 73d5e2b47264 ("cpu/hotplug: detect SMT disabled by BIOS") ... the hotplug code attempted to detect when SMT was disabled by BIOS, in which case it reported SMT as permanently disabled. However, that code broke a virt hotplug scenario, where the guest is booted with only primary CPU threads, and a sibling is brought online later. The problem is that there doesn't seem to be a way to reliably distinguish between the HW "SMT disabled by BIOS" case and the virt "sibling not yet brought online" case. So the above-mentioned commit was a bit misguided, as it permanently disabled SMT for both cases, preventing future virt sibling hotplugs. Going back and reviewing the original problems which were attempted to be solved by that commit, when SMT was disabled in BIOS: 1) /sys/devices/system/cpu/smt/control showed "on" instead of "notsupported"; and 2) vmx_vm_init() was incorrectly showing the L1TF_MSG_SMT warning. I'd propose that we instead consider #1 above to not actually be a problem. Because, at least in the virt case, it's possible that SMT wasn't disabled by BIOS and a sibling thread could be brought online later. So it makes sense to just always default the smt control to "on" to allow for that possibility (assuming cpuid indicates that the CPU supports SMT). The real problem is #2, which has a simple fix: change vmx_vm_init() to query the actual current SMT state -- i.e., whether any siblings are currently online -- instead of looking at the SMT "control" sysfs value. So fix it by: a) reverting the original "fix" and its followup fix: 73d5e2b47264 ("cpu/hotplug: detect SMT disabled by BIOS") bc2d8d262cba ("cpu/hotplug: Fix SMT supported evaluation") and b) changing vmx_vm_init() to query the actual current SMT state -- instead of the sysfs control value -- to determine whether the L1TF warning is needed. This also requires the 'sched_smt_present' variable to exported, instead of 'cpu_smt_control'. Fixes: 73d5e2b47264 ("cpu/hotplug: detect SMT disabled by BIOS") Reported-by: Igor Mammedov <imammedo@redhat.com> Signed-off-by: Josh Poimboeuf <jpoimboe@redhat.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Cc: Joe Mario <jmario@redhat.com> Cc: Jiri Kosina <jikos@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: kvm@vger.kernel.org Cc: stable@vger.kernel.org Link: https://lkml.kernel.org/r/e3a85d585da28cc333ecbc1e78ee9216e6da9396.1548794349.git.jpoimboe@redhat.com
2019-01-30 13:13:58 +00:00
EXPORT_SYMBOL_GPL(sched_smt_present);
sched/core: Rewrite and improve select_idle_siblings() select_idle_siblings() is a known pain point for a number of workloads; it either does too much or not enough and sometimes just does plain wrong. This rewrite attempts to address a number of issues (but sadly not all). The current code does an unconditional sched_domain iteration; with the intent of finding an idle core (on SMT hardware). The problems which this patch tries to address are: - its pointless to look for idle cores if the machine is real busy; at which point you're just wasting cycles. - it's behaviour is inconsistent between SMT and !SMT hardware in that !SMT hardware ends up doing a scan for any idle CPU in the LLC domain, while SMT hardware does a scan for idle cores and if that fails, falls back to a scan for idle threads on the 'target' core. The new code replaces the sched_domain scan with 3 explicit scans: 1) search for an idle core in the LLC 2) search for an idle CPU in the LLC 3) search for an idle thread in the 'target' core where 1 and 3 are conditional on SMT support and 1 and 2 have runtime heuristics to skip the step. Step 1) is conditional on sd_llc_shared->has_idle_cores; when a cpu goes idle and sd_llc_shared->has_idle_cores is false, we scan all SMT siblings of the CPU going idle. Similarly, we clear sd_llc_shared->has_idle_cores when we fail to find an idle core. Step 2) tracks the average cost of the scan and compares this to the average idle time guestimate for the CPU doing the wakeup. There is a significant fudge factor involved to deal with the variability of the averages. Esp. hackbench was sensitive to this. Step 3) is unconditional; we assume (also per step 1) that scanning all SMT siblings in a core is 'cheap'. With this; SMT systems gain step 2, which cures a few benchmarks -- notably one from Facebook. One 'feature' of the sched_domain iteration, which we preserve in the new code, is that it would start scanning from the 'target' CPU, instead of scanning the cpumask in cpu id order. This avoids multiple CPUs in the LLC scanning for idle to gang up and find the same CPU quite as much. The down side is that tasks can end up hopping across the LLC for no apparent reason. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-05-09 08:38:05 +00:00
static inline void set_idle_cores(int cpu, int val)
{
struct sched_domain_shared *sds;
sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
if (sds)
WRITE_ONCE(sds->has_idle_cores, val);
}
static inline bool test_idle_cores(int cpu, bool def)
{
struct sched_domain_shared *sds;
sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
if (sds)
return READ_ONCE(sds->has_idle_cores);
return def;
}
/*
* Scans the local SMT mask to see if the entire core is idle, and records this
* information in sd_llc_shared->has_idle_cores.
*
* Since SMT siblings share all cache levels, inspecting this limited remote
* state should be fairly cheap.
*/
void __update_idle_core(struct rq *rq)
sched/core: Rewrite and improve select_idle_siblings() select_idle_siblings() is a known pain point for a number of workloads; it either does too much or not enough and sometimes just does plain wrong. This rewrite attempts to address a number of issues (but sadly not all). The current code does an unconditional sched_domain iteration; with the intent of finding an idle core (on SMT hardware). The problems which this patch tries to address are: - its pointless to look for idle cores if the machine is real busy; at which point you're just wasting cycles. - it's behaviour is inconsistent between SMT and !SMT hardware in that !SMT hardware ends up doing a scan for any idle CPU in the LLC domain, while SMT hardware does a scan for idle cores and if that fails, falls back to a scan for idle threads on the 'target' core. The new code replaces the sched_domain scan with 3 explicit scans: 1) search for an idle core in the LLC 2) search for an idle CPU in the LLC 3) search for an idle thread in the 'target' core where 1 and 3 are conditional on SMT support and 1 and 2 have runtime heuristics to skip the step. Step 1) is conditional on sd_llc_shared->has_idle_cores; when a cpu goes idle and sd_llc_shared->has_idle_cores is false, we scan all SMT siblings of the CPU going idle. Similarly, we clear sd_llc_shared->has_idle_cores when we fail to find an idle core. Step 2) tracks the average cost of the scan and compares this to the average idle time guestimate for the CPU doing the wakeup. There is a significant fudge factor involved to deal with the variability of the averages. Esp. hackbench was sensitive to this. Step 3) is unconditional; we assume (also per step 1) that scanning all SMT siblings in a core is 'cheap'. With this; SMT systems gain step 2, which cures a few benchmarks -- notably one from Facebook. One 'feature' of the sched_domain iteration, which we preserve in the new code, is that it would start scanning from the 'target' CPU, instead of scanning the cpumask in cpu id order. This avoids multiple CPUs in the LLC scanning for idle to gang up and find the same CPU quite as much. The down side is that tasks can end up hopping across the LLC for no apparent reason. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-05-09 08:38:05 +00:00
{
int core = cpu_of(rq);
int cpu;
rcu_read_lock();
if (test_idle_cores(core, true))
goto unlock;
for_each_cpu(cpu, cpu_smt_mask(core)) {
if (cpu == core)
continue;
if (!available_idle_cpu(cpu))
sched/core: Rewrite and improve select_idle_siblings() select_idle_siblings() is a known pain point for a number of workloads; it either does too much or not enough and sometimes just does plain wrong. This rewrite attempts to address a number of issues (but sadly not all). The current code does an unconditional sched_domain iteration; with the intent of finding an idle core (on SMT hardware). The problems which this patch tries to address are: - its pointless to look for idle cores if the machine is real busy; at which point you're just wasting cycles. - it's behaviour is inconsistent between SMT and !SMT hardware in that !SMT hardware ends up doing a scan for any idle CPU in the LLC domain, while SMT hardware does a scan for idle cores and if that fails, falls back to a scan for idle threads on the 'target' core. The new code replaces the sched_domain scan with 3 explicit scans: 1) search for an idle core in the LLC 2) search for an idle CPU in the LLC 3) search for an idle thread in the 'target' core where 1 and 3 are conditional on SMT support and 1 and 2 have runtime heuristics to skip the step. Step 1) is conditional on sd_llc_shared->has_idle_cores; when a cpu goes idle and sd_llc_shared->has_idle_cores is false, we scan all SMT siblings of the CPU going idle. Similarly, we clear sd_llc_shared->has_idle_cores when we fail to find an idle core. Step 2) tracks the average cost of the scan and compares this to the average idle time guestimate for the CPU doing the wakeup. There is a significant fudge factor involved to deal with the variability of the averages. Esp. hackbench was sensitive to this. Step 3) is unconditional; we assume (also per step 1) that scanning all SMT siblings in a core is 'cheap'. With this; SMT systems gain step 2, which cures a few benchmarks -- notably one from Facebook. One 'feature' of the sched_domain iteration, which we preserve in the new code, is that it would start scanning from the 'target' CPU, instead of scanning the cpumask in cpu id order. This avoids multiple CPUs in the LLC scanning for idle to gang up and find the same CPU quite as much. The down side is that tasks can end up hopping across the LLC for no apparent reason. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-05-09 08:38:05 +00:00
goto unlock;
}
set_idle_cores(core, 1);
unlock:
rcu_read_unlock();
}
/*
* Scan the entire LLC domain for idle cores; this dynamically switches off if
* there are no idle cores left in the system; tracked through
* sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
*/
static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
{
struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
int core, cpu;
sched/core: Rewrite and improve select_idle_siblings() select_idle_siblings() is a known pain point for a number of workloads; it either does too much or not enough and sometimes just does plain wrong. This rewrite attempts to address a number of issues (but sadly not all). The current code does an unconditional sched_domain iteration; with the intent of finding an idle core (on SMT hardware). The problems which this patch tries to address are: - its pointless to look for idle cores if the machine is real busy; at which point you're just wasting cycles. - it's behaviour is inconsistent between SMT and !SMT hardware in that !SMT hardware ends up doing a scan for any idle CPU in the LLC domain, while SMT hardware does a scan for idle cores and if that fails, falls back to a scan for idle threads on the 'target' core. The new code replaces the sched_domain scan with 3 explicit scans: 1) search for an idle core in the LLC 2) search for an idle CPU in the LLC 3) search for an idle thread in the 'target' core where 1 and 3 are conditional on SMT support and 1 and 2 have runtime heuristics to skip the step. Step 1) is conditional on sd_llc_shared->has_idle_cores; when a cpu goes idle and sd_llc_shared->has_idle_cores is false, we scan all SMT siblings of the CPU going idle. Similarly, we clear sd_llc_shared->has_idle_cores when we fail to find an idle core. Step 2) tracks the average cost of the scan and compares this to the average idle time guestimate for the CPU doing the wakeup. There is a significant fudge factor involved to deal with the variability of the averages. Esp. hackbench was sensitive to this. Step 3) is unconditional; we assume (also per step 1) that scanning all SMT siblings in a core is 'cheap'. With this; SMT systems gain step 2, which cures a few benchmarks -- notably one from Facebook. One 'feature' of the sched_domain iteration, which we preserve in the new code, is that it would start scanning from the 'target' CPU, instead of scanning the cpumask in cpu id order. This avoids multiple CPUs in the LLC scanning for idle to gang up and find the same CPU quite as much. The down side is that tasks can end up hopping across the LLC for no apparent reason. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-05-09 08:38:05 +00:00
if (!static_branch_likely(&sched_smt_present))
return -1;
sched/core: Rewrite and improve select_idle_siblings() select_idle_siblings() is a known pain point for a number of workloads; it either does too much or not enough and sometimes just does plain wrong. This rewrite attempts to address a number of issues (but sadly not all). The current code does an unconditional sched_domain iteration; with the intent of finding an idle core (on SMT hardware). The problems which this patch tries to address are: - its pointless to look for idle cores if the machine is real busy; at which point you're just wasting cycles. - it's behaviour is inconsistent between SMT and !SMT hardware in that !SMT hardware ends up doing a scan for any idle CPU in the LLC domain, while SMT hardware does a scan for idle cores and if that fails, falls back to a scan for idle threads on the 'target' core. The new code replaces the sched_domain scan with 3 explicit scans: 1) search for an idle core in the LLC 2) search for an idle CPU in the LLC 3) search for an idle thread in the 'target' core where 1 and 3 are conditional on SMT support and 1 and 2 have runtime heuristics to skip the step. Step 1) is conditional on sd_llc_shared->has_idle_cores; when a cpu goes idle and sd_llc_shared->has_idle_cores is false, we scan all SMT siblings of the CPU going idle. Similarly, we clear sd_llc_shared->has_idle_cores when we fail to find an idle core. Step 2) tracks the average cost of the scan and compares this to the average idle time guestimate for the CPU doing the wakeup. There is a significant fudge factor involved to deal with the variability of the averages. Esp. hackbench was sensitive to this. Step 3) is unconditional; we assume (also per step 1) that scanning all SMT siblings in a core is 'cheap'. With this; SMT systems gain step 2, which cures a few benchmarks -- notably one from Facebook. One 'feature' of the sched_domain iteration, which we preserve in the new code, is that it would start scanning from the 'target' CPU, instead of scanning the cpumask in cpu id order. This avoids multiple CPUs in the LLC scanning for idle to gang up and find the same CPU quite as much. The down side is that tasks can end up hopping across the LLC for no apparent reason. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-05-09 08:38:05 +00:00
if (!test_idle_cores(target, false))
return -1;
cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
sched/core: Rewrite and improve select_idle_siblings() select_idle_siblings() is a known pain point for a number of workloads; it either does too much or not enough and sometimes just does plain wrong. This rewrite attempts to address a number of issues (but sadly not all). The current code does an unconditional sched_domain iteration; with the intent of finding an idle core (on SMT hardware). The problems which this patch tries to address are: - its pointless to look for idle cores if the machine is real busy; at which point you're just wasting cycles. - it's behaviour is inconsistent between SMT and !SMT hardware in that !SMT hardware ends up doing a scan for any idle CPU in the LLC domain, while SMT hardware does a scan for idle cores and if that fails, falls back to a scan for idle threads on the 'target' core. The new code replaces the sched_domain scan with 3 explicit scans: 1) search for an idle core in the LLC 2) search for an idle CPU in the LLC 3) search for an idle thread in the 'target' core where 1 and 3 are conditional on SMT support and 1 and 2 have runtime heuristics to skip the step. Step 1) is conditional on sd_llc_shared->has_idle_cores; when a cpu goes idle and sd_llc_shared->has_idle_cores is false, we scan all SMT siblings of the CPU going idle. Similarly, we clear sd_llc_shared->has_idle_cores when we fail to find an idle core. Step 2) tracks the average cost of the scan and compares this to the average idle time guestimate for the CPU doing the wakeup. There is a significant fudge factor involved to deal with the variability of the averages. Esp. hackbench was sensitive to this. Step 3) is unconditional; we assume (also per step 1) that scanning all SMT siblings in a core is 'cheap'. With this; SMT systems gain step 2, which cures a few benchmarks -- notably one from Facebook. One 'feature' of the sched_domain iteration, which we preserve in the new code, is that it would start scanning from the 'target' CPU, instead of scanning the cpumask in cpu id order. This avoids multiple CPUs in the LLC scanning for idle to gang up and find the same CPU quite as much. The down side is that tasks can end up hopping across the LLC for no apparent reason. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-05-09 08:38:05 +00:00
for_each_cpu_wrap(core, cpus, target) {
sched/core: Rewrite and improve select_idle_siblings() select_idle_siblings() is a known pain point for a number of workloads; it either does too much or not enough and sometimes just does plain wrong. This rewrite attempts to address a number of issues (but sadly not all). The current code does an unconditional sched_domain iteration; with the intent of finding an idle core (on SMT hardware). The problems which this patch tries to address are: - its pointless to look for idle cores if the machine is real busy; at which point you're just wasting cycles. - it's behaviour is inconsistent between SMT and !SMT hardware in that !SMT hardware ends up doing a scan for any idle CPU in the LLC domain, while SMT hardware does a scan for idle cores and if that fails, falls back to a scan for idle threads on the 'target' core. The new code replaces the sched_domain scan with 3 explicit scans: 1) search for an idle core in the LLC 2) search for an idle CPU in the LLC 3) search for an idle thread in the 'target' core where 1 and 3 are conditional on SMT support and 1 and 2 have runtime heuristics to skip the step. Step 1) is conditional on sd_llc_shared->has_idle_cores; when a cpu goes idle and sd_llc_shared->has_idle_cores is false, we scan all SMT siblings of the CPU going idle. Similarly, we clear sd_llc_shared->has_idle_cores when we fail to find an idle core. Step 2) tracks the average cost of the scan and compares this to the average idle time guestimate for the CPU doing the wakeup. There is a significant fudge factor involved to deal with the variability of the averages. Esp. hackbench was sensitive to this. Step 3) is unconditional; we assume (also per step 1) that scanning all SMT siblings in a core is 'cheap'. With this; SMT systems gain step 2, which cures a few benchmarks -- notably one from Facebook. One 'feature' of the sched_domain iteration, which we preserve in the new code, is that it would start scanning from the 'target' CPU, instead of scanning the cpumask in cpu id order. This avoids multiple CPUs in the LLC scanning for idle to gang up and find the same CPU quite as much. The down side is that tasks can end up hopping across the LLC for no apparent reason. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-05-09 08:38:05 +00:00
bool idle = true;
for_each_cpu(cpu, cpu_smt_mask(core)) {
if (!available_idle_cpu(cpu)) {
sched/core: Rewrite and improve select_idle_siblings() select_idle_siblings() is a known pain point for a number of workloads; it either does too much or not enough and sometimes just does plain wrong. This rewrite attempts to address a number of issues (but sadly not all). The current code does an unconditional sched_domain iteration; with the intent of finding an idle core (on SMT hardware). The problems which this patch tries to address are: - its pointless to look for idle cores if the machine is real busy; at which point you're just wasting cycles. - it's behaviour is inconsistent between SMT and !SMT hardware in that !SMT hardware ends up doing a scan for any idle CPU in the LLC domain, while SMT hardware does a scan for idle cores and if that fails, falls back to a scan for idle threads on the 'target' core. The new code replaces the sched_domain scan with 3 explicit scans: 1) search for an idle core in the LLC 2) search for an idle CPU in the LLC 3) search for an idle thread in the 'target' core where 1 and 3 are conditional on SMT support and 1 and 2 have runtime heuristics to skip the step. Step 1) is conditional on sd_llc_shared->has_idle_cores; when a cpu goes idle and sd_llc_shared->has_idle_cores is false, we scan all SMT siblings of the CPU going idle. Similarly, we clear sd_llc_shared->has_idle_cores when we fail to find an idle core. Step 2) tracks the average cost of the scan and compares this to the average idle time guestimate for the CPU doing the wakeup. There is a significant fudge factor involved to deal with the variability of the averages. Esp. hackbench was sensitive to this. Step 3) is unconditional; we assume (also per step 1) that scanning all SMT siblings in a core is 'cheap'. With this; SMT systems gain step 2, which cures a few benchmarks -- notably one from Facebook. One 'feature' of the sched_domain iteration, which we preserve in the new code, is that it would start scanning from the 'target' CPU, instead of scanning the cpumask in cpu id order. This avoids multiple CPUs in the LLC scanning for idle to gang up and find the same CPU quite as much. The down side is that tasks can end up hopping across the LLC for no apparent reason. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-05-09 08:38:05 +00:00
idle = false;
break;
}
sched/core: Rewrite and improve select_idle_siblings() select_idle_siblings() is a known pain point for a number of workloads; it either does too much or not enough and sometimes just does plain wrong. This rewrite attempts to address a number of issues (but sadly not all). The current code does an unconditional sched_domain iteration; with the intent of finding an idle core (on SMT hardware). The problems which this patch tries to address are: - its pointless to look for idle cores if the machine is real busy; at which point you're just wasting cycles. - it's behaviour is inconsistent between SMT and !SMT hardware in that !SMT hardware ends up doing a scan for any idle CPU in the LLC domain, while SMT hardware does a scan for idle cores and if that fails, falls back to a scan for idle threads on the 'target' core. The new code replaces the sched_domain scan with 3 explicit scans: 1) search for an idle core in the LLC 2) search for an idle CPU in the LLC 3) search for an idle thread in the 'target' core where 1 and 3 are conditional on SMT support and 1 and 2 have runtime heuristics to skip the step. Step 1) is conditional on sd_llc_shared->has_idle_cores; when a cpu goes idle and sd_llc_shared->has_idle_cores is false, we scan all SMT siblings of the CPU going idle. Similarly, we clear sd_llc_shared->has_idle_cores when we fail to find an idle core. Step 2) tracks the average cost of the scan and compares this to the average idle time guestimate for the CPU doing the wakeup. There is a significant fudge factor involved to deal with the variability of the averages. Esp. hackbench was sensitive to this. Step 3) is unconditional; we assume (also per step 1) that scanning all SMT siblings in a core is 'cheap'. With this; SMT systems gain step 2, which cures a few benchmarks -- notably one from Facebook. One 'feature' of the sched_domain iteration, which we preserve in the new code, is that it would start scanning from the 'target' CPU, instead of scanning the cpumask in cpu id order. This avoids multiple CPUs in the LLC scanning for idle to gang up and find the same CPU quite as much. The down side is that tasks can end up hopping across the LLC for no apparent reason. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-05-09 08:38:05 +00:00
}
cpumask_andnot(cpus, cpus, cpu_smt_mask(core));
sched/core: Rewrite and improve select_idle_siblings() select_idle_siblings() is a known pain point for a number of workloads; it either does too much or not enough and sometimes just does plain wrong. This rewrite attempts to address a number of issues (but sadly not all). The current code does an unconditional sched_domain iteration; with the intent of finding an idle core (on SMT hardware). The problems which this patch tries to address are: - its pointless to look for idle cores if the machine is real busy; at which point you're just wasting cycles. - it's behaviour is inconsistent between SMT and !SMT hardware in that !SMT hardware ends up doing a scan for any idle CPU in the LLC domain, while SMT hardware does a scan for idle cores and if that fails, falls back to a scan for idle threads on the 'target' core. The new code replaces the sched_domain scan with 3 explicit scans: 1) search for an idle core in the LLC 2) search for an idle CPU in the LLC 3) search for an idle thread in the 'target' core where 1 and 3 are conditional on SMT support and 1 and 2 have runtime heuristics to skip the step. Step 1) is conditional on sd_llc_shared->has_idle_cores; when a cpu goes idle and sd_llc_shared->has_idle_cores is false, we scan all SMT siblings of the CPU going idle. Similarly, we clear sd_llc_shared->has_idle_cores when we fail to find an idle core. Step 2) tracks the average cost of the scan and compares this to the average idle time guestimate for the CPU doing the wakeup. There is a significant fudge factor involved to deal with the variability of the averages. Esp. hackbench was sensitive to this. Step 3) is unconditional; we assume (also per step 1) that scanning all SMT siblings in a core is 'cheap'. With this; SMT systems gain step 2, which cures a few benchmarks -- notably one from Facebook. One 'feature' of the sched_domain iteration, which we preserve in the new code, is that it would start scanning from the 'target' CPU, instead of scanning the cpumask in cpu id order. This avoids multiple CPUs in the LLC scanning for idle to gang up and find the same CPU quite as much. The down side is that tasks can end up hopping across the LLC for no apparent reason. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-05-09 08:38:05 +00:00
if (idle)
return core;
}
/*
* Failed to find an idle core; stop looking for one.
*/
set_idle_cores(target, 0);
return -1;
}
/*
* Scan the local SMT mask for idle CPUs.
*/
static int select_idle_smt(struct task_struct *p, int target)
sched/core: Rewrite and improve select_idle_siblings() select_idle_siblings() is a known pain point for a number of workloads; it either does too much or not enough and sometimes just does plain wrong. This rewrite attempts to address a number of issues (but sadly not all). The current code does an unconditional sched_domain iteration; with the intent of finding an idle core (on SMT hardware). The problems which this patch tries to address are: - its pointless to look for idle cores if the machine is real busy; at which point you're just wasting cycles. - it's behaviour is inconsistent between SMT and !SMT hardware in that !SMT hardware ends up doing a scan for any idle CPU in the LLC domain, while SMT hardware does a scan for idle cores and if that fails, falls back to a scan for idle threads on the 'target' core. The new code replaces the sched_domain scan with 3 explicit scans: 1) search for an idle core in the LLC 2) search for an idle CPU in the LLC 3) search for an idle thread in the 'target' core where 1 and 3 are conditional on SMT support and 1 and 2 have runtime heuristics to skip the step. Step 1) is conditional on sd_llc_shared->has_idle_cores; when a cpu goes idle and sd_llc_shared->has_idle_cores is false, we scan all SMT siblings of the CPU going idle. Similarly, we clear sd_llc_shared->has_idle_cores when we fail to find an idle core. Step 2) tracks the average cost of the scan and compares this to the average idle time guestimate for the CPU doing the wakeup. There is a significant fudge factor involved to deal with the variability of the averages. Esp. hackbench was sensitive to this. Step 3) is unconditional; we assume (also per step 1) that scanning all SMT siblings in a core is 'cheap'. With this; SMT systems gain step 2, which cures a few benchmarks -- notably one from Facebook. One 'feature' of the sched_domain iteration, which we preserve in the new code, is that it would start scanning from the 'target' CPU, instead of scanning the cpumask in cpu id order. This avoids multiple CPUs in the LLC scanning for idle to gang up and find the same CPU quite as much. The down side is that tasks can end up hopping across the LLC for no apparent reason. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-05-09 08:38:05 +00:00
{
sched/fair: Make sched-idle CPU selection consistent throughout There are instances where we keep searching for an idle CPU despite already having a sched-idle CPU (in find_idlest_group_cpu(), select_idle_smt() and select_idle_cpu() and then there are places where we don't necessarily do that and return a sched-idle CPU as soon as we find one (in select_idle_sibling()). This looks a bit inconsistent and it may be worth having the same policy everywhere. On the other hand, choosing a sched-idle CPU over a idle one shall be beneficial from performance and power point of view as well, as we don't need to get the CPU online from a deep idle state which wastes quite a lot of time and energy and delays the scheduling of the newly woken up task. This patch tries to simplify code around sched-idle CPU selection and make it consistent throughout. Testing is done with the help of rt-app on hikey board (ARM64 octa-core, 2 clusters, 0-3 and 4-7). The cpufreq governor was set to performance to avoid any side affects from CPU frequency. Following are the tests performed: Test 1: 1-cfs-task: A single SCHED_NORMAL task is pinned to CPU5 which runs for 2333 us out of 7777 us (so gives time for the cluster to go in deep idle state). Test 2: 1-cfs-1-idle-task: A single SCHED_NORMAL task is pinned on CPU5 and single SCHED_IDLE task is pinned on CPU6 (to make sure cluster 1 doesn't go in deep idle state). Test 3: 1-cfs-8-idle-task: A single SCHED_NORMAL task is pinned on CPU5 and eight SCHED_IDLE tasks are created which run forever (not pinned anywhere, so they run on all CPUs). Checked with kernelshark that as soon as NORMAL task sleeps, the SCHED_IDLE task starts running on CPU5. And here are the results on mean latency (in us), using the "st" tool. $ st 1-cfs-task/rt-app-cfs_thread-0.log N min max sum mean stddev 642 90 592 197180 307.134 109.906 $ st 1-cfs-1-idle-task/rt-app-cfs_thread-0.log N min max sum mean stddev 642 67 311 113850 177.336 41.4251 $ st 1-cfs-8-idle-task/rt-app-cfs_thread-0.log N min max sum mean stddev 643 29 173 41364 64.3297 13.2344 The mean latency when we need to: - wakeup from deep idle state is 307 us. - wakeup from shallow idle state is 177 us. - preempt a SCHED_IDLE task is 64 us. Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: https://lkml.kernel.org/r/b90cbcce608cef4e02a7bbfe178335f76d201bab.1573728344.git.viresh.kumar@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-11-14 10:49:27 +00:00
int cpu;
sched/core: Rewrite and improve select_idle_siblings() select_idle_siblings() is a known pain point for a number of workloads; it either does too much or not enough and sometimes just does plain wrong. This rewrite attempts to address a number of issues (but sadly not all). The current code does an unconditional sched_domain iteration; with the intent of finding an idle core (on SMT hardware). The problems which this patch tries to address are: - its pointless to look for idle cores if the machine is real busy; at which point you're just wasting cycles. - it's behaviour is inconsistent between SMT and !SMT hardware in that !SMT hardware ends up doing a scan for any idle CPU in the LLC domain, while SMT hardware does a scan for idle cores and if that fails, falls back to a scan for idle threads on the 'target' core. The new code replaces the sched_domain scan with 3 explicit scans: 1) search for an idle core in the LLC 2) search for an idle CPU in the LLC 3) search for an idle thread in the 'target' core where 1 and 3 are conditional on SMT support and 1 and 2 have runtime heuristics to skip the step. Step 1) is conditional on sd_llc_shared->has_idle_cores; when a cpu goes idle and sd_llc_shared->has_idle_cores is false, we scan all SMT siblings of the CPU going idle. Similarly, we clear sd_llc_shared->has_idle_cores when we fail to find an idle core. Step 2) tracks the average cost of the scan and compares this to the average idle time guestimate for the CPU doing the wakeup. There is a significant fudge factor involved to deal with the variability of the averages. Esp. hackbench was sensitive to this. Step 3) is unconditional; we assume (also per step 1) that scanning all SMT siblings in a core is 'cheap'. With this; SMT systems gain step 2, which cures a few benchmarks -- notably one from Facebook. One 'feature' of the sched_domain iteration, which we preserve in the new code, is that it would start scanning from the 'target' CPU, instead of scanning the cpumask in cpu id order. This avoids multiple CPUs in the LLC scanning for idle to gang up and find the same CPU quite as much. The down side is that tasks can end up hopping across the LLC for no apparent reason. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-05-09 08:38:05 +00:00
if (!static_branch_likely(&sched_smt_present))
return -1;
sched/core: Rewrite and improve select_idle_siblings() select_idle_siblings() is a known pain point for a number of workloads; it either does too much or not enough and sometimes just does plain wrong. This rewrite attempts to address a number of issues (but sadly not all). The current code does an unconditional sched_domain iteration; with the intent of finding an idle core (on SMT hardware). The problems which this patch tries to address are: - its pointless to look for idle cores if the machine is real busy; at which point you're just wasting cycles. - it's behaviour is inconsistent between SMT and !SMT hardware in that !SMT hardware ends up doing a scan for any idle CPU in the LLC domain, while SMT hardware does a scan for idle cores and if that fails, falls back to a scan for idle threads on the 'target' core. The new code replaces the sched_domain scan with 3 explicit scans: 1) search for an idle core in the LLC 2) search for an idle CPU in the LLC 3) search for an idle thread in the 'target' core where 1 and 3 are conditional on SMT support and 1 and 2 have runtime heuristics to skip the step. Step 1) is conditional on sd_llc_shared->has_idle_cores; when a cpu goes idle and sd_llc_shared->has_idle_cores is false, we scan all SMT siblings of the CPU going idle. Similarly, we clear sd_llc_shared->has_idle_cores when we fail to find an idle core. Step 2) tracks the average cost of the scan and compares this to the average idle time guestimate for the CPU doing the wakeup. There is a significant fudge factor involved to deal with the variability of the averages. Esp. hackbench was sensitive to this. Step 3) is unconditional; we assume (also per step 1) that scanning all SMT siblings in a core is 'cheap'. With this; SMT systems gain step 2, which cures a few benchmarks -- notably one from Facebook. One 'feature' of the sched_domain iteration, which we preserve in the new code, is that it would start scanning from the 'target' CPU, instead of scanning the cpumask in cpu id order. This avoids multiple CPUs in the LLC scanning for idle to gang up and find the same CPU quite as much. The down side is that tasks can end up hopping across the LLC for no apparent reason. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-05-09 08:38:05 +00:00
for_each_cpu(cpu, cpu_smt_mask(target)) {
if (!cpumask_test_cpu(cpu, p->cpus_ptr))
sched/core: Rewrite and improve select_idle_siblings() select_idle_siblings() is a known pain point for a number of workloads; it either does too much or not enough and sometimes just does plain wrong. This rewrite attempts to address a number of issues (but sadly not all). The current code does an unconditional sched_domain iteration; with the intent of finding an idle core (on SMT hardware). The problems which this patch tries to address are: - its pointless to look for idle cores if the machine is real busy; at which point you're just wasting cycles. - it's behaviour is inconsistent between SMT and !SMT hardware in that !SMT hardware ends up doing a scan for any idle CPU in the LLC domain, while SMT hardware does a scan for idle cores and if that fails, falls back to a scan for idle threads on the 'target' core. The new code replaces the sched_domain scan with 3 explicit scans: 1) search for an idle core in the LLC 2) search for an idle CPU in the LLC 3) search for an idle thread in the 'target' core where 1 and 3 are conditional on SMT support and 1 and 2 have runtime heuristics to skip the step. Step 1) is conditional on sd_llc_shared->has_idle_cores; when a cpu goes idle and sd_llc_shared->has_idle_cores is false, we scan all SMT siblings of the CPU going idle. Similarly, we clear sd_llc_shared->has_idle_cores when we fail to find an idle core. Step 2) tracks the average cost of the scan and compares this to the average idle time guestimate for the CPU doing the wakeup. There is a significant fudge factor involved to deal with the variability of the averages. Esp. hackbench was sensitive to this. Step 3) is unconditional; we assume (also per step 1) that scanning all SMT siblings in a core is 'cheap'. With this; SMT systems gain step 2, which cures a few benchmarks -- notably one from Facebook. One 'feature' of the sched_domain iteration, which we preserve in the new code, is that it would start scanning from the 'target' CPU, instead of scanning the cpumask in cpu id order. This avoids multiple CPUs in the LLC scanning for idle to gang up and find the same CPU quite as much. The down side is that tasks can end up hopping across the LLC for no apparent reason. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-05-09 08:38:05 +00:00
continue;
sched/fair: Make sched-idle CPU selection consistent throughout There are instances where we keep searching for an idle CPU despite already having a sched-idle CPU (in find_idlest_group_cpu(), select_idle_smt() and select_idle_cpu() and then there are places where we don't necessarily do that and return a sched-idle CPU as soon as we find one (in select_idle_sibling()). This looks a bit inconsistent and it may be worth having the same policy everywhere. On the other hand, choosing a sched-idle CPU over a idle one shall be beneficial from performance and power point of view as well, as we don't need to get the CPU online from a deep idle state which wastes quite a lot of time and energy and delays the scheduling of the newly woken up task. This patch tries to simplify code around sched-idle CPU selection and make it consistent throughout. Testing is done with the help of rt-app on hikey board (ARM64 octa-core, 2 clusters, 0-3 and 4-7). The cpufreq governor was set to performance to avoid any side affects from CPU frequency. Following are the tests performed: Test 1: 1-cfs-task: A single SCHED_NORMAL task is pinned to CPU5 which runs for 2333 us out of 7777 us (so gives time for the cluster to go in deep idle state). Test 2: 1-cfs-1-idle-task: A single SCHED_NORMAL task is pinned on CPU5 and single SCHED_IDLE task is pinned on CPU6 (to make sure cluster 1 doesn't go in deep idle state). Test 3: 1-cfs-8-idle-task: A single SCHED_NORMAL task is pinned on CPU5 and eight SCHED_IDLE tasks are created which run forever (not pinned anywhere, so they run on all CPUs). Checked with kernelshark that as soon as NORMAL task sleeps, the SCHED_IDLE task starts running on CPU5. And here are the results on mean latency (in us), using the "st" tool. $ st 1-cfs-task/rt-app-cfs_thread-0.log N min max sum mean stddev 642 90 592 197180 307.134 109.906 $ st 1-cfs-1-idle-task/rt-app-cfs_thread-0.log N min max sum mean stddev 642 67 311 113850 177.336 41.4251 $ st 1-cfs-8-idle-task/rt-app-cfs_thread-0.log N min max sum mean stddev 643 29 173 41364 64.3297 13.2344 The mean latency when we need to: - wakeup from deep idle state is 307 us. - wakeup from shallow idle state is 177 us. - preempt a SCHED_IDLE task is 64 us. Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: https://lkml.kernel.org/r/b90cbcce608cef4e02a7bbfe178335f76d201bab.1573728344.git.viresh.kumar@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-11-14 10:49:27 +00:00
if (available_idle_cpu(cpu) || sched_idle_cpu(cpu))
sched/core: Rewrite and improve select_idle_siblings() select_idle_siblings() is a known pain point for a number of workloads; it either does too much or not enough and sometimes just does plain wrong. This rewrite attempts to address a number of issues (but sadly not all). The current code does an unconditional sched_domain iteration; with the intent of finding an idle core (on SMT hardware). The problems which this patch tries to address are: - its pointless to look for idle cores if the machine is real busy; at which point you're just wasting cycles. - it's behaviour is inconsistent between SMT and !SMT hardware in that !SMT hardware ends up doing a scan for any idle CPU in the LLC domain, while SMT hardware does a scan for idle cores and if that fails, falls back to a scan for idle threads on the 'target' core. The new code replaces the sched_domain scan with 3 explicit scans: 1) search for an idle core in the LLC 2) search for an idle CPU in the LLC 3) search for an idle thread in the 'target' core where 1 and 3 are conditional on SMT support and 1 and 2 have runtime heuristics to skip the step. Step 1) is conditional on sd_llc_shared->has_idle_cores; when a cpu goes idle and sd_llc_shared->has_idle_cores is false, we scan all SMT siblings of the CPU going idle. Similarly, we clear sd_llc_shared->has_idle_cores when we fail to find an idle core. Step 2) tracks the average cost of the scan and compares this to the average idle time guestimate for the CPU doing the wakeup. There is a significant fudge factor involved to deal with the variability of the averages. Esp. hackbench was sensitive to this. Step 3) is unconditional; we assume (also per step 1) that scanning all SMT siblings in a core is 'cheap'. With this; SMT systems gain step 2, which cures a few benchmarks -- notably one from Facebook. One 'feature' of the sched_domain iteration, which we preserve in the new code, is that it would start scanning from the 'target' CPU, instead of scanning the cpumask in cpu id order. This avoids multiple CPUs in the LLC scanning for idle to gang up and find the same CPU quite as much. The down side is that tasks can end up hopping across the LLC for no apparent reason. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-05-09 08:38:05 +00:00
return cpu;
}
sched/fair: Make sched-idle CPU selection consistent throughout There are instances where we keep searching for an idle CPU despite already having a sched-idle CPU (in find_idlest_group_cpu(), select_idle_smt() and select_idle_cpu() and then there are places where we don't necessarily do that and return a sched-idle CPU as soon as we find one (in select_idle_sibling()). This looks a bit inconsistent and it may be worth having the same policy everywhere. On the other hand, choosing a sched-idle CPU over a idle one shall be beneficial from performance and power point of view as well, as we don't need to get the CPU online from a deep idle state which wastes quite a lot of time and energy and delays the scheduling of the newly woken up task. This patch tries to simplify code around sched-idle CPU selection and make it consistent throughout. Testing is done with the help of rt-app on hikey board (ARM64 octa-core, 2 clusters, 0-3 and 4-7). The cpufreq governor was set to performance to avoid any side affects from CPU frequency. Following are the tests performed: Test 1: 1-cfs-task: A single SCHED_NORMAL task is pinned to CPU5 which runs for 2333 us out of 7777 us (so gives time for the cluster to go in deep idle state). Test 2: 1-cfs-1-idle-task: A single SCHED_NORMAL task is pinned on CPU5 and single SCHED_IDLE task is pinned on CPU6 (to make sure cluster 1 doesn't go in deep idle state). Test 3: 1-cfs-8-idle-task: A single SCHED_NORMAL task is pinned on CPU5 and eight SCHED_IDLE tasks are created which run forever (not pinned anywhere, so they run on all CPUs). Checked with kernelshark that as soon as NORMAL task sleeps, the SCHED_IDLE task starts running on CPU5. And here are the results on mean latency (in us), using the "st" tool. $ st 1-cfs-task/rt-app-cfs_thread-0.log N min max sum mean stddev 642 90 592 197180 307.134 109.906 $ st 1-cfs-1-idle-task/rt-app-cfs_thread-0.log N min max sum mean stddev 642 67 311 113850 177.336 41.4251 $ st 1-cfs-8-idle-task/rt-app-cfs_thread-0.log N min max sum mean stddev 643 29 173 41364 64.3297 13.2344 The mean latency when we need to: - wakeup from deep idle state is 307 us. - wakeup from shallow idle state is 177 us. - preempt a SCHED_IDLE task is 64 us. Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: https://lkml.kernel.org/r/b90cbcce608cef4e02a7bbfe178335f76d201bab.1573728344.git.viresh.kumar@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-11-14 10:49:27 +00:00
return -1;
sched/core: Rewrite and improve select_idle_siblings() select_idle_siblings() is a known pain point for a number of workloads; it either does too much or not enough and sometimes just does plain wrong. This rewrite attempts to address a number of issues (but sadly not all). The current code does an unconditional sched_domain iteration; with the intent of finding an idle core (on SMT hardware). The problems which this patch tries to address are: - its pointless to look for idle cores if the machine is real busy; at which point you're just wasting cycles. - it's behaviour is inconsistent between SMT and !SMT hardware in that !SMT hardware ends up doing a scan for any idle CPU in the LLC domain, while SMT hardware does a scan for idle cores and if that fails, falls back to a scan for idle threads on the 'target' core. The new code replaces the sched_domain scan with 3 explicit scans: 1) search for an idle core in the LLC 2) search for an idle CPU in the LLC 3) search for an idle thread in the 'target' core where 1 and 3 are conditional on SMT support and 1 and 2 have runtime heuristics to skip the step. Step 1) is conditional on sd_llc_shared->has_idle_cores; when a cpu goes idle and sd_llc_shared->has_idle_cores is false, we scan all SMT siblings of the CPU going idle. Similarly, we clear sd_llc_shared->has_idle_cores when we fail to find an idle core. Step 2) tracks the average cost of the scan and compares this to the average idle time guestimate for the CPU doing the wakeup. There is a significant fudge factor involved to deal with the variability of the averages. Esp. hackbench was sensitive to this. Step 3) is unconditional; we assume (also per step 1) that scanning all SMT siblings in a core is 'cheap'. With this; SMT systems gain step 2, which cures a few benchmarks -- notably one from Facebook. One 'feature' of the sched_domain iteration, which we preserve in the new code, is that it would start scanning from the 'target' CPU, instead of scanning the cpumask in cpu id order. This avoids multiple CPUs in the LLC scanning for idle to gang up and find the same CPU quite as much. The down side is that tasks can end up hopping across the LLC for no apparent reason. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-05-09 08:38:05 +00:00
}
#else /* CONFIG_SCHED_SMT */
static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
{
return -1;
}
static inline int select_idle_smt(struct task_struct *p, int target)
sched/core: Rewrite and improve select_idle_siblings() select_idle_siblings() is a known pain point for a number of workloads; it either does too much or not enough and sometimes just does plain wrong. This rewrite attempts to address a number of issues (but sadly not all). The current code does an unconditional sched_domain iteration; with the intent of finding an idle core (on SMT hardware). The problems which this patch tries to address are: - its pointless to look for idle cores if the machine is real busy; at which point you're just wasting cycles. - it's behaviour is inconsistent between SMT and !SMT hardware in that !SMT hardware ends up doing a scan for any idle CPU in the LLC domain, while SMT hardware does a scan for idle cores and if that fails, falls back to a scan for idle threads on the 'target' core. The new code replaces the sched_domain scan with 3 explicit scans: 1) search for an idle core in the LLC 2) search for an idle CPU in the LLC 3) search for an idle thread in the 'target' core where 1 and 3 are conditional on SMT support and 1 and 2 have runtime heuristics to skip the step. Step 1) is conditional on sd_llc_shared->has_idle_cores; when a cpu goes idle and sd_llc_shared->has_idle_cores is false, we scan all SMT siblings of the CPU going idle. Similarly, we clear sd_llc_shared->has_idle_cores when we fail to find an idle core. Step 2) tracks the average cost of the scan and compares this to the average idle time guestimate for the CPU doing the wakeup. There is a significant fudge factor involved to deal with the variability of the averages. Esp. hackbench was sensitive to this. Step 3) is unconditional; we assume (also per step 1) that scanning all SMT siblings in a core is 'cheap'. With this; SMT systems gain step 2, which cures a few benchmarks -- notably one from Facebook. One 'feature' of the sched_domain iteration, which we preserve in the new code, is that it would start scanning from the 'target' CPU, instead of scanning the cpumask in cpu id order. This avoids multiple CPUs in the LLC scanning for idle to gang up and find the same CPU quite as much. The down side is that tasks can end up hopping across the LLC for no apparent reason. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-05-09 08:38:05 +00:00
{
return -1;
}
#endif /* CONFIG_SCHED_SMT */
/*
* Scan the LLC domain for idle CPUs; this is dynamically regulated by
* comparing the average scan cost (tracked in sd->avg_scan_cost) against the
* average idle time for this rq (as found in rq->avg_idle).
*/
sched/core: Rewrite and improve select_idle_siblings() select_idle_siblings() is a known pain point for a number of workloads; it either does too much or not enough and sometimes just does plain wrong. This rewrite attempts to address a number of issues (but sadly not all). The current code does an unconditional sched_domain iteration; with the intent of finding an idle core (on SMT hardware). The problems which this patch tries to address are: - its pointless to look for idle cores if the machine is real busy; at which point you're just wasting cycles. - it's behaviour is inconsistent between SMT and !SMT hardware in that !SMT hardware ends up doing a scan for any idle CPU in the LLC domain, while SMT hardware does a scan for idle cores and if that fails, falls back to a scan for idle threads on the 'target' core. The new code replaces the sched_domain scan with 3 explicit scans: 1) search for an idle core in the LLC 2) search for an idle CPU in the LLC 3) search for an idle thread in the 'target' core where 1 and 3 are conditional on SMT support and 1 and 2 have runtime heuristics to skip the step. Step 1) is conditional on sd_llc_shared->has_idle_cores; when a cpu goes idle and sd_llc_shared->has_idle_cores is false, we scan all SMT siblings of the CPU going idle. Similarly, we clear sd_llc_shared->has_idle_cores when we fail to find an idle core. Step 2) tracks the average cost of the scan and compares this to the average idle time guestimate for the CPU doing the wakeup. There is a significant fudge factor involved to deal with the variability of the averages. Esp. hackbench was sensitive to this. Step 3) is unconditional; we assume (also per step 1) that scanning all SMT siblings in a core is 'cheap'. With this; SMT systems gain step 2, which cures a few benchmarks -- notably one from Facebook. One 'feature' of the sched_domain iteration, which we preserve in the new code, is that it would start scanning from the 'target' CPU, instead of scanning the cpumask in cpu id order. This avoids multiple CPUs in the LLC scanning for idle to gang up and find the same CPU quite as much. The down side is that tasks can end up hopping across the LLC for no apparent reason. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-05-09 08:38:05 +00:00
static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
{
struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
sched/fair: Fix sched domains NULL dereference in select_idle_sibling() Commit: 10e2f1acd01 ("sched/core: Rewrite and improve select_idle_siblings()") ... improved select_idle_sibling(), but also triggered a regression (crash) during CPU-hotplug: BUG: unable to handle kernel NULL pointer dereference at 0000000000000078 IP: [<ffffffffb10cd332>] select_idle_sibling+0x1c2/0x4f0 Call Trace: <IRQ> select_task_rq_fair+0x749/0x930 ? select_task_rq_fair+0xb4/0x930 ? __lock_is_held+0x54/0x70 try_to_wake_up+0x19a/0x5b0 default_wake_function+0x12/0x20 autoremove_wake_function+0x12/0x40 __wake_up_common+0x55/0x90 __wake_up+0x39/0x50 wake_up_klogd_work_func+0x40/0x60 irq_work_run_list+0x57/0x80 irq_work_run+0x2c/0x30 smp_irq_work_interrupt+0x2e/0x40 irq_work_interrupt+0x96/0xa0 <EOI> ? _raw_spin_unlock_irqrestore+0x45/0x80 try_to_wake_up+0x4a/0x5b0 wake_up_state+0x10/0x20 __kthread_unpark+0x67/0x70 kthread_unpark+0x22/0x30 cpuhp_online_idle+0x3e/0x70 cpu_startup_entry+0x6a/0x450 start_secondary+0x154/0x180 This can be reproduced by running the ftrace test case of kselftest, the test case will hot-unplug the CPU and the CPU will attach to the NULL sched-domain during scheduler teardown. The step 2 for the rewrite select_idle_siblings(): | Step 2) tracks the average cost of the scan and compares this to the | average idle time guestimate for the CPU doing the wakeup. If the CPU which doing the wakeup is the going hot-unplug CPU, then NULL sched domain will be dereferenced to acquire the average cost of the scan. This patch fix it by failing the search of an idle CPU in the LLC process if this sched domain is NULL. Tested-by: Catalin Marinas <catalin.marinas@arm.com> Signed-off-by: Wanpeng Li <wanpeng.li@hotmail.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1475971443-3187-1-git-send-email-wanpeng.li@hotmail.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-10-09 00:04:03 +00:00
struct sched_domain *this_sd;
sched/core: Implement new approach to scale select_idle_cpu() Hackbench recently suffered a bunch of pain, first by commit: 4c77b18cf8b7 ("sched/fair: Make select_idle_cpu() more aggressive") and then by commit: c743f0a5c50f ("sched/fair, cpumask: Export for_each_cpu_wrap()") which fixed a bug in the initial for_each_cpu_wrap() implementation that made select_idle_cpu() even more expensive. The bug was that it would skip over CPUs when bits were consequtive in the bitmask. This however gave me an idea to fix select_idle_cpu(); where the old scheme was a cliff-edge throttle on idle scanning, this introduces a more gradual approach. Instead of stopping to scan entirely, we limit how many CPUs we scan. Initial benchmarks show that it mostly recovers hackbench while not hurting anything else, except Mason's schbench, but not as bad as the old thing. It also appears to recover the tbench high-end, which also suffered like hackbench. Tested-by: Matt Fleming <matt@codeblueprint.co.uk> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Chris Mason <clm@fb.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hpa@zytor.com Cc: kitsunyan <kitsunyan@inbox.ru> Cc: linux-kernel@vger.kernel.org Cc: lvenanci@redhat.com Cc: riel@redhat.com Cc: xiaolong.ye@intel.com Link: http://lkml.kernel.org/r/20170517105350.hk5m4h4jb6dfr65a@hirez.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-05-17 10:53:50 +00:00
u64 avg_cost, avg_idle;
sched/fair: Align rq->avg_idle and rq->avg_scan_cost sched/core.c uses update_avg() for rq->avg_idle and sched/fair.c uses an open-coded version (with the exact same decay factor) for rq->avg_scan_cost. On top of that, select_idle_cpu() expects to be able to compare these two fields. The only difference between the two is that rq->avg_scan_cost is computed using a pure division rather than a shift. Turns out it actually matters, first of all because the shifted value can be negative, and the standard has this to say about it: """ The result of E1 >> E2 is E1 right-shifted E2 bit positions. [...] If E1 has a signed type and a negative value, the resulting value is implementation-defined. """ Not only this, but (arithmetic) right shifting a negative value (using 2's complement) is *not* equivalent to dividing it by the corresponding power of 2. Let's look at a few examples: -4 -> 0xF..FC -4 >> 3 -> 0xF..FF == -1 != -4 / 8 -8 -> 0xF..F8 -8 >> 3 -> 0xF..FF == -1 == -8 / 8 -9 -> 0xF..F7 -9 >> 3 -> 0xF..FE == -2 != -9 / 8 Make update_avg() use a division, and export it to the private scheduler header to reuse it where relevant. Note that this still lets compilers use a shift here, but should prevent any unwanted surprise. The disassembly of select_idle_cpu() remains unchanged on arm64, and ttwu_do_wakeup() gains 2 instructions; the diff sort of looks like this: - sub x1, x1, x0 + subs x1, x1, x0 // set condition codes + add x0, x1, #0x7 + csel x0, x0, x1, mi // x0 = x1 < 0 ? x0 : x1 add x0, x3, x0, asr #3 which does the right thing (i.e. gives us the expected result while still using an arithmetic shift) Signed-off-by: Valentin Schneider <valentin.schneider@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Link: https://lkml.kernel.org/r/20200330090127.16294-1-valentin.schneider@arm.com
2020-03-30 09:01:27 +00:00
u64 time;
int this = smp_processor_id();
sched/fair: Make sched-idle CPU selection consistent throughout There are instances where we keep searching for an idle CPU despite already having a sched-idle CPU (in find_idlest_group_cpu(), select_idle_smt() and select_idle_cpu() and then there are places where we don't necessarily do that and return a sched-idle CPU as soon as we find one (in select_idle_sibling()). This looks a bit inconsistent and it may be worth having the same policy everywhere. On the other hand, choosing a sched-idle CPU over a idle one shall be beneficial from performance and power point of view as well, as we don't need to get the CPU online from a deep idle state which wastes quite a lot of time and energy and delays the scheduling of the newly woken up task. This patch tries to simplify code around sched-idle CPU selection and make it consistent throughout. Testing is done with the help of rt-app on hikey board (ARM64 octa-core, 2 clusters, 0-3 and 4-7). The cpufreq governor was set to performance to avoid any side affects from CPU frequency. Following are the tests performed: Test 1: 1-cfs-task: A single SCHED_NORMAL task is pinned to CPU5 which runs for 2333 us out of 7777 us (so gives time for the cluster to go in deep idle state). Test 2: 1-cfs-1-idle-task: A single SCHED_NORMAL task is pinned on CPU5 and single SCHED_IDLE task is pinned on CPU6 (to make sure cluster 1 doesn't go in deep idle state). Test 3: 1-cfs-8-idle-task: A single SCHED_NORMAL task is pinned on CPU5 and eight SCHED_IDLE tasks are created which run forever (not pinned anywhere, so they run on all CPUs). Checked with kernelshark that as soon as NORMAL task sleeps, the SCHED_IDLE task starts running on CPU5. And here are the results on mean latency (in us), using the "st" tool. $ st 1-cfs-task/rt-app-cfs_thread-0.log N min max sum mean stddev 642 90 592 197180 307.134 109.906 $ st 1-cfs-1-idle-task/rt-app-cfs_thread-0.log N min max sum mean stddev 642 67 311 113850 177.336 41.4251 $ st 1-cfs-8-idle-task/rt-app-cfs_thread-0.log N min max sum mean stddev 643 29 173 41364 64.3297 13.2344 The mean latency when we need to: - wakeup from deep idle state is 307 us. - wakeup from shallow idle state is 177 us. - preempt a SCHED_IDLE task is 64 us. Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: https://lkml.kernel.org/r/b90cbcce608cef4e02a7bbfe178335f76d201bab.1573728344.git.viresh.kumar@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-11-14 10:49:27 +00:00
int cpu, nr = INT_MAX;
sched/core: Rewrite and improve select_idle_siblings() select_idle_siblings() is a known pain point for a number of workloads; it either does too much or not enough and sometimes just does plain wrong. This rewrite attempts to address a number of issues (but sadly not all). The current code does an unconditional sched_domain iteration; with the intent of finding an idle core (on SMT hardware). The problems which this patch tries to address are: - its pointless to look for idle cores if the machine is real busy; at which point you're just wasting cycles. - it's behaviour is inconsistent between SMT and !SMT hardware in that !SMT hardware ends up doing a scan for any idle CPU in the LLC domain, while SMT hardware does a scan for idle cores and if that fails, falls back to a scan for idle threads on the 'target' core. The new code replaces the sched_domain scan with 3 explicit scans: 1) search for an idle core in the LLC 2) search for an idle CPU in the LLC 3) search for an idle thread in the 'target' core where 1 and 3 are conditional on SMT support and 1 and 2 have runtime heuristics to skip the step. Step 1) is conditional on sd_llc_shared->has_idle_cores; when a cpu goes idle and sd_llc_shared->has_idle_cores is false, we scan all SMT siblings of the CPU going idle. Similarly, we clear sd_llc_shared->has_idle_cores when we fail to find an idle core. Step 2) tracks the average cost of the scan and compares this to the average idle time guestimate for the CPU doing the wakeup. There is a significant fudge factor involved to deal with the variability of the averages. Esp. hackbench was sensitive to this. Step 3) is unconditional; we assume (also per step 1) that scanning all SMT siblings in a core is 'cheap'. With this; SMT systems gain step 2, which cures a few benchmarks -- notably one from Facebook. One 'feature' of the sched_domain iteration, which we preserve in the new code, is that it would start scanning from the 'target' CPU, instead of scanning the cpumask in cpu id order. This avoids multiple CPUs in the LLC scanning for idle to gang up and find the same CPU quite as much. The down side is that tasks can end up hopping across the LLC for no apparent reason. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-05-09 08:38:05 +00:00
sched/fair: Fix sched domains NULL dereference in select_idle_sibling() Commit: 10e2f1acd01 ("sched/core: Rewrite and improve select_idle_siblings()") ... improved select_idle_sibling(), but also triggered a regression (crash) during CPU-hotplug: BUG: unable to handle kernel NULL pointer dereference at 0000000000000078 IP: [<ffffffffb10cd332>] select_idle_sibling+0x1c2/0x4f0 Call Trace: <IRQ> select_task_rq_fair+0x749/0x930 ? select_task_rq_fair+0xb4/0x930 ? __lock_is_held+0x54/0x70 try_to_wake_up+0x19a/0x5b0 default_wake_function+0x12/0x20 autoremove_wake_function+0x12/0x40 __wake_up_common+0x55/0x90 __wake_up+0x39/0x50 wake_up_klogd_work_func+0x40/0x60 irq_work_run_list+0x57/0x80 irq_work_run+0x2c/0x30 smp_irq_work_interrupt+0x2e/0x40 irq_work_interrupt+0x96/0xa0 <EOI> ? _raw_spin_unlock_irqrestore+0x45/0x80 try_to_wake_up+0x4a/0x5b0 wake_up_state+0x10/0x20 __kthread_unpark+0x67/0x70 kthread_unpark+0x22/0x30 cpuhp_online_idle+0x3e/0x70 cpu_startup_entry+0x6a/0x450 start_secondary+0x154/0x180 This can be reproduced by running the ftrace test case of kselftest, the test case will hot-unplug the CPU and the CPU will attach to the NULL sched-domain during scheduler teardown. The step 2 for the rewrite select_idle_siblings(): | Step 2) tracks the average cost of the scan and compares this to the | average idle time guestimate for the CPU doing the wakeup. If the CPU which doing the wakeup is the going hot-unplug CPU, then NULL sched domain will be dereferenced to acquire the average cost of the scan. This patch fix it by failing the search of an idle CPU in the LLC process if this sched domain is NULL. Tested-by: Catalin Marinas <catalin.marinas@arm.com> Signed-off-by: Wanpeng Li <wanpeng.li@hotmail.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1475971443-3187-1-git-send-email-wanpeng.li@hotmail.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-10-09 00:04:03 +00:00
this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
if (!this_sd)
return -1;
sched/core: Rewrite and improve select_idle_siblings() select_idle_siblings() is a known pain point for a number of workloads; it either does too much or not enough and sometimes just does plain wrong. This rewrite attempts to address a number of issues (but sadly not all). The current code does an unconditional sched_domain iteration; with the intent of finding an idle core (on SMT hardware). The problems which this patch tries to address are: - its pointless to look for idle cores if the machine is real busy; at which point you're just wasting cycles. - it's behaviour is inconsistent between SMT and !SMT hardware in that !SMT hardware ends up doing a scan for any idle CPU in the LLC domain, while SMT hardware does a scan for idle cores and if that fails, falls back to a scan for idle threads on the 'target' core. The new code replaces the sched_domain scan with 3 explicit scans: 1) search for an idle core in the LLC 2) search for an idle CPU in the LLC 3) search for an idle thread in the 'target' core where 1 and 3 are conditional on SMT support and 1 and 2 have runtime heuristics to skip the step. Step 1) is conditional on sd_llc_shared->has_idle_cores; when a cpu goes idle and sd_llc_shared->has_idle_cores is false, we scan all SMT siblings of the CPU going idle. Similarly, we clear sd_llc_shared->has_idle_cores when we fail to find an idle core. Step 2) tracks the average cost of the scan and compares this to the average idle time guestimate for the CPU doing the wakeup. There is a significant fudge factor involved to deal with the variability of the averages. Esp. hackbench was sensitive to this. Step 3) is unconditional; we assume (also per step 1) that scanning all SMT siblings in a core is 'cheap'. With this; SMT systems gain step 2, which cures a few benchmarks -- notably one from Facebook. One 'feature' of the sched_domain iteration, which we preserve in the new code, is that it would start scanning from the 'target' CPU, instead of scanning the cpumask in cpu id order. This avoids multiple CPUs in the LLC scanning for idle to gang up and find the same CPU quite as much. The down side is that tasks can end up hopping across the LLC for no apparent reason. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-05-09 08:38:05 +00:00
/*
* Due to large variance we need a large fuzz factor; hackbench in
* particularly is sensitive here.
*/
sched/core: Implement new approach to scale select_idle_cpu() Hackbench recently suffered a bunch of pain, first by commit: 4c77b18cf8b7 ("sched/fair: Make select_idle_cpu() more aggressive") and then by commit: c743f0a5c50f ("sched/fair, cpumask: Export for_each_cpu_wrap()") which fixed a bug in the initial for_each_cpu_wrap() implementation that made select_idle_cpu() even more expensive. The bug was that it would skip over CPUs when bits were consequtive in the bitmask. This however gave me an idea to fix select_idle_cpu(); where the old scheme was a cliff-edge throttle on idle scanning, this introduces a more gradual approach. Instead of stopping to scan entirely, we limit how many CPUs we scan. Initial benchmarks show that it mostly recovers hackbench while not hurting anything else, except Mason's schbench, but not as bad as the old thing. It also appears to recover the tbench high-end, which also suffered like hackbench. Tested-by: Matt Fleming <matt@codeblueprint.co.uk> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Chris Mason <clm@fb.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hpa@zytor.com Cc: kitsunyan <kitsunyan@inbox.ru> Cc: linux-kernel@vger.kernel.org Cc: lvenanci@redhat.com Cc: riel@redhat.com Cc: xiaolong.ye@intel.com Link: http://lkml.kernel.org/r/20170517105350.hk5m4h4jb6dfr65a@hirez.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-05-17 10:53:50 +00:00
avg_idle = this_rq()->avg_idle / 512;
avg_cost = this_sd->avg_scan_cost + 1;
if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost)
sched/core: Rewrite and improve select_idle_siblings() select_idle_siblings() is a known pain point for a number of workloads; it either does too much or not enough and sometimes just does plain wrong. This rewrite attempts to address a number of issues (but sadly not all). The current code does an unconditional sched_domain iteration; with the intent of finding an idle core (on SMT hardware). The problems which this patch tries to address are: - its pointless to look for idle cores if the machine is real busy; at which point you're just wasting cycles. - it's behaviour is inconsistent between SMT and !SMT hardware in that !SMT hardware ends up doing a scan for any idle CPU in the LLC domain, while SMT hardware does a scan for idle cores and if that fails, falls back to a scan for idle threads on the 'target' core. The new code replaces the sched_domain scan with 3 explicit scans: 1) search for an idle core in the LLC 2) search for an idle CPU in the LLC 3) search for an idle thread in the 'target' core where 1 and 3 are conditional on SMT support and 1 and 2 have runtime heuristics to skip the step. Step 1) is conditional on sd_llc_shared->has_idle_cores; when a cpu goes idle and sd_llc_shared->has_idle_cores is false, we scan all SMT siblings of the CPU going idle. Similarly, we clear sd_llc_shared->has_idle_cores when we fail to find an idle core. Step 2) tracks the average cost of the scan and compares this to the average idle time guestimate for the CPU doing the wakeup. There is a significant fudge factor involved to deal with the variability of the averages. Esp. hackbench was sensitive to this. Step 3) is unconditional; we assume (also per step 1) that scanning all SMT siblings in a core is 'cheap'. With this; SMT systems gain step 2, which cures a few benchmarks -- notably one from Facebook. One 'feature' of the sched_domain iteration, which we preserve in the new code, is that it would start scanning from the 'target' CPU, instead of scanning the cpumask in cpu id order. This avoids multiple CPUs in the LLC scanning for idle to gang up and find the same CPU quite as much. The down side is that tasks can end up hopping across the LLC for no apparent reason. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-05-09 08:38:05 +00:00
return -1;
sched/core: Implement new approach to scale select_idle_cpu() Hackbench recently suffered a bunch of pain, first by commit: 4c77b18cf8b7 ("sched/fair: Make select_idle_cpu() more aggressive") and then by commit: c743f0a5c50f ("sched/fair, cpumask: Export for_each_cpu_wrap()") which fixed a bug in the initial for_each_cpu_wrap() implementation that made select_idle_cpu() even more expensive. The bug was that it would skip over CPUs when bits were consequtive in the bitmask. This however gave me an idea to fix select_idle_cpu(); where the old scheme was a cliff-edge throttle on idle scanning, this introduces a more gradual approach. Instead of stopping to scan entirely, we limit how many CPUs we scan. Initial benchmarks show that it mostly recovers hackbench while not hurting anything else, except Mason's schbench, but not as bad as the old thing. It also appears to recover the tbench high-end, which also suffered like hackbench. Tested-by: Matt Fleming <matt@codeblueprint.co.uk> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Chris Mason <clm@fb.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hpa@zytor.com Cc: kitsunyan <kitsunyan@inbox.ru> Cc: linux-kernel@vger.kernel.org Cc: lvenanci@redhat.com Cc: riel@redhat.com Cc: xiaolong.ye@intel.com Link: http://lkml.kernel.org/r/20170517105350.hk5m4h4jb6dfr65a@hirez.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-05-17 10:53:50 +00:00
if (sched_feat(SIS_PROP)) {
u64 span_avg = sd->span_weight * avg_idle;
if (span_avg > 4*avg_cost)
nr = div_u64(span_avg, avg_cost);
else
nr = 4;
}
time = cpu_clock(this);
sched/core: Rewrite and improve select_idle_siblings() select_idle_siblings() is a known pain point for a number of workloads; it either does too much or not enough and sometimes just does plain wrong. This rewrite attempts to address a number of issues (but sadly not all). The current code does an unconditional sched_domain iteration; with the intent of finding an idle core (on SMT hardware). The problems which this patch tries to address are: - its pointless to look for idle cores if the machine is real busy; at which point you're just wasting cycles. - it's behaviour is inconsistent between SMT and !SMT hardware in that !SMT hardware ends up doing a scan for any idle CPU in the LLC domain, while SMT hardware does a scan for idle cores and if that fails, falls back to a scan for idle threads on the 'target' core. The new code replaces the sched_domain scan with 3 explicit scans: 1) search for an idle core in the LLC 2) search for an idle CPU in the LLC 3) search for an idle thread in the 'target' core where 1 and 3 are conditional on SMT support and 1 and 2 have runtime heuristics to skip the step. Step 1) is conditional on sd_llc_shared->has_idle_cores; when a cpu goes idle and sd_llc_shared->has_idle_cores is false, we scan all SMT siblings of the CPU going idle. Similarly, we clear sd_llc_shared->has_idle_cores when we fail to find an idle core. Step 2) tracks the average cost of the scan and compares this to the average idle time guestimate for the CPU doing the wakeup. There is a significant fudge factor involved to deal with the variability of the averages. Esp. hackbench was sensitive to this. Step 3) is unconditional; we assume (also per step 1) that scanning all SMT siblings in a core is 'cheap'. With this; SMT systems gain step 2, which cures a few benchmarks -- notably one from Facebook. One 'feature' of the sched_domain iteration, which we preserve in the new code, is that it would start scanning from the 'target' CPU, instead of scanning the cpumask in cpu id order. This avoids multiple CPUs in the LLC scanning for idle to gang up and find the same CPU quite as much. The down side is that tasks can end up hopping across the LLC for no apparent reason. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-05-09 08:38:05 +00:00
cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
for_each_cpu_wrap(cpu, cpus, target) {
sched/core: Implement new approach to scale select_idle_cpu() Hackbench recently suffered a bunch of pain, first by commit: 4c77b18cf8b7 ("sched/fair: Make select_idle_cpu() more aggressive") and then by commit: c743f0a5c50f ("sched/fair, cpumask: Export for_each_cpu_wrap()") which fixed a bug in the initial for_each_cpu_wrap() implementation that made select_idle_cpu() even more expensive. The bug was that it would skip over CPUs when bits were consequtive in the bitmask. This however gave me an idea to fix select_idle_cpu(); where the old scheme was a cliff-edge throttle on idle scanning, this introduces a more gradual approach. Instead of stopping to scan entirely, we limit how many CPUs we scan. Initial benchmarks show that it mostly recovers hackbench while not hurting anything else, except Mason's schbench, but not as bad as the old thing. It also appears to recover the tbench high-end, which also suffered like hackbench. Tested-by: Matt Fleming <matt@codeblueprint.co.uk> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Chris Mason <clm@fb.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hpa@zytor.com Cc: kitsunyan <kitsunyan@inbox.ru> Cc: linux-kernel@vger.kernel.org Cc: lvenanci@redhat.com Cc: riel@redhat.com Cc: xiaolong.ye@intel.com Link: http://lkml.kernel.org/r/20170517105350.hk5m4h4jb6dfr65a@hirez.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-05-17 10:53:50 +00:00
if (!--nr)
sched/fair: Make sched-idle CPU selection consistent throughout There are instances where we keep searching for an idle CPU despite already having a sched-idle CPU (in find_idlest_group_cpu(), select_idle_smt() and select_idle_cpu() and then there are places where we don't necessarily do that and return a sched-idle CPU as soon as we find one (in select_idle_sibling()). This looks a bit inconsistent and it may be worth having the same policy everywhere. On the other hand, choosing a sched-idle CPU over a idle one shall be beneficial from performance and power point of view as well, as we don't need to get the CPU online from a deep idle state which wastes quite a lot of time and energy and delays the scheduling of the newly woken up task. This patch tries to simplify code around sched-idle CPU selection and make it consistent throughout. Testing is done with the help of rt-app on hikey board (ARM64 octa-core, 2 clusters, 0-3 and 4-7). The cpufreq governor was set to performance to avoid any side affects from CPU frequency. Following are the tests performed: Test 1: 1-cfs-task: A single SCHED_NORMAL task is pinned to CPU5 which runs for 2333 us out of 7777 us (so gives time for the cluster to go in deep idle state). Test 2: 1-cfs-1-idle-task: A single SCHED_NORMAL task is pinned on CPU5 and single SCHED_IDLE task is pinned on CPU6 (to make sure cluster 1 doesn't go in deep idle state). Test 3: 1-cfs-8-idle-task: A single SCHED_NORMAL task is pinned on CPU5 and eight SCHED_IDLE tasks are created which run forever (not pinned anywhere, so they run on all CPUs). Checked with kernelshark that as soon as NORMAL task sleeps, the SCHED_IDLE task starts running on CPU5. And here are the results on mean latency (in us), using the "st" tool. $ st 1-cfs-task/rt-app-cfs_thread-0.log N min max sum mean stddev 642 90 592 197180 307.134 109.906 $ st 1-cfs-1-idle-task/rt-app-cfs_thread-0.log N min max sum mean stddev 642 67 311 113850 177.336 41.4251 $ st 1-cfs-8-idle-task/rt-app-cfs_thread-0.log N min max sum mean stddev 643 29 173 41364 64.3297 13.2344 The mean latency when we need to: - wakeup from deep idle state is 307 us. - wakeup from shallow idle state is 177 us. - preempt a SCHED_IDLE task is 64 us. Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: https://lkml.kernel.org/r/b90cbcce608cef4e02a7bbfe178335f76d201bab.1573728344.git.viresh.kumar@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-11-14 10:49:27 +00:00
return -1;
if (available_idle_cpu(cpu) || sched_idle_cpu(cpu))
sched/core: Rewrite and improve select_idle_siblings() select_idle_siblings() is a known pain point for a number of workloads; it either does too much or not enough and sometimes just does plain wrong. This rewrite attempts to address a number of issues (but sadly not all). The current code does an unconditional sched_domain iteration; with the intent of finding an idle core (on SMT hardware). The problems which this patch tries to address are: - its pointless to look for idle cores if the machine is real busy; at which point you're just wasting cycles. - it's behaviour is inconsistent between SMT and !SMT hardware in that !SMT hardware ends up doing a scan for any idle CPU in the LLC domain, while SMT hardware does a scan for idle cores and if that fails, falls back to a scan for idle threads on the 'target' core. The new code replaces the sched_domain scan with 3 explicit scans: 1) search for an idle core in the LLC 2) search for an idle CPU in the LLC 3) search for an idle thread in the 'target' core where 1 and 3 are conditional on SMT support and 1 and 2 have runtime heuristics to skip the step. Step 1) is conditional on sd_llc_shared->has_idle_cores; when a cpu goes idle and sd_llc_shared->has_idle_cores is false, we scan all SMT siblings of the CPU going idle. Similarly, we clear sd_llc_shared->has_idle_cores when we fail to find an idle core. Step 2) tracks the average cost of the scan and compares this to the average idle time guestimate for the CPU doing the wakeup. There is a significant fudge factor involved to deal with the variability of the averages. Esp. hackbench was sensitive to this. Step 3) is unconditional; we assume (also per step 1) that scanning all SMT siblings in a core is 'cheap'. With this; SMT systems gain step 2, which cures a few benchmarks -- notably one from Facebook. One 'feature' of the sched_domain iteration, which we preserve in the new code, is that it would start scanning from the 'target' CPU, instead of scanning the cpumask in cpu id order. This avoids multiple CPUs in the LLC scanning for idle to gang up and find the same CPU quite as much. The down side is that tasks can end up hopping across the LLC for no apparent reason. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-05-09 08:38:05 +00:00
break;
}
time = cpu_clock(this) - time;
sched/fair: Align rq->avg_idle and rq->avg_scan_cost sched/core.c uses update_avg() for rq->avg_idle and sched/fair.c uses an open-coded version (with the exact same decay factor) for rq->avg_scan_cost. On top of that, select_idle_cpu() expects to be able to compare these two fields. The only difference between the two is that rq->avg_scan_cost is computed using a pure division rather than a shift. Turns out it actually matters, first of all because the shifted value can be negative, and the standard has this to say about it: """ The result of E1 >> E2 is E1 right-shifted E2 bit positions. [...] If E1 has a signed type and a negative value, the resulting value is implementation-defined. """ Not only this, but (arithmetic) right shifting a negative value (using 2's complement) is *not* equivalent to dividing it by the corresponding power of 2. Let's look at a few examples: -4 -> 0xF..FC -4 >> 3 -> 0xF..FF == -1 != -4 / 8 -8 -> 0xF..F8 -8 >> 3 -> 0xF..FF == -1 == -8 / 8 -9 -> 0xF..F7 -9 >> 3 -> 0xF..FE == -2 != -9 / 8 Make update_avg() use a division, and export it to the private scheduler header to reuse it where relevant. Note that this still lets compilers use a shift here, but should prevent any unwanted surprise. The disassembly of select_idle_cpu() remains unchanged on arm64, and ttwu_do_wakeup() gains 2 instructions; the diff sort of looks like this: - sub x1, x1, x0 + subs x1, x1, x0 // set condition codes + add x0, x1, #0x7 + csel x0, x0, x1, mi // x0 = x1 < 0 ? x0 : x1 add x0, x3, x0, asr #3 which does the right thing (i.e. gives us the expected result while still using an arithmetic shift) Signed-off-by: Valentin Schneider <valentin.schneider@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Link: https://lkml.kernel.org/r/20200330090127.16294-1-valentin.schneider@arm.com
2020-03-30 09:01:27 +00:00
update_avg(&this_sd->avg_scan_cost, time);
sched/core: Rewrite and improve select_idle_siblings() select_idle_siblings() is a known pain point for a number of workloads; it either does too much or not enough and sometimes just does plain wrong. This rewrite attempts to address a number of issues (but sadly not all). The current code does an unconditional sched_domain iteration; with the intent of finding an idle core (on SMT hardware). The problems which this patch tries to address are: - its pointless to look for idle cores if the machine is real busy; at which point you're just wasting cycles. - it's behaviour is inconsistent between SMT and !SMT hardware in that !SMT hardware ends up doing a scan for any idle CPU in the LLC domain, while SMT hardware does a scan for idle cores and if that fails, falls back to a scan for idle threads on the 'target' core. The new code replaces the sched_domain scan with 3 explicit scans: 1) search for an idle core in the LLC 2) search for an idle CPU in the LLC 3) search for an idle thread in the 'target' core where 1 and 3 are conditional on SMT support and 1 and 2 have runtime heuristics to skip the step. Step 1) is conditional on sd_llc_shared->has_idle_cores; when a cpu goes idle and sd_llc_shared->has_idle_cores is false, we scan all SMT siblings of the CPU going idle. Similarly, we clear sd_llc_shared->has_idle_cores when we fail to find an idle core. Step 2) tracks the average cost of the scan and compares this to the average idle time guestimate for the CPU doing the wakeup. There is a significant fudge factor involved to deal with the variability of the averages. Esp. hackbench was sensitive to this. Step 3) is unconditional; we assume (also per step 1) that scanning all SMT siblings in a core is 'cheap'. With this; SMT systems gain step 2, which cures a few benchmarks -- notably one from Facebook. One 'feature' of the sched_domain iteration, which we preserve in the new code, is that it would start scanning from the 'target' CPU, instead of scanning the cpumask in cpu id order. This avoids multiple CPUs in the LLC scanning for idle to gang up and find the same CPU quite as much. The down side is that tasks can end up hopping across the LLC for no apparent reason. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-05-09 08:38:05 +00:00
return cpu;
}
sched/fair: Add asymmetric CPU capacity wakeup scan Issue ===== On asymmetric CPU capacity topologies, we currently rely on wake_cap() to drive select_task_rq_fair() towards either: - its slow-path (find_idlest_cpu()) if either the previous or current (waking) CPU has too little capacity for the waking task - its fast-path (select_idle_sibling()) otherwise Commit: 3273163c6775 ("sched/fair: Let asymmetric CPU configurations balance at wake-up") points out that this relies on the assumption that "[...]the CPU capacities within an SD_SHARE_PKG_RESOURCES domain (sd_llc) are homogeneous". This assumption no longer holds on newer generations of big.LITTLE systems (DynamIQ), which can accommodate CPUs of different compute capacity within a single LLC domain. To hopefully paint a better picture, a regular big.LITTLE topology would look like this: +---------+ +---------+ | L2 | | L2 | +----+----+ +----+----+ |CPU0|CPU1| |CPU2|CPU3| +----+----+ +----+----+ ^^^ ^^^ LITTLEs bigs which would result in the following scheduler topology: DIE [ ] <- sd_asym_cpucapacity MC [ ] [ ] <- sd_llc 0 1 2 3 Conversely, a DynamIQ topology could look like: +-------------------+ | L3 | +----+----+----+----+ | L2 | L2 | L2 | L2 | +----+----+----+----+ |CPU0|CPU1|CPU2|CPU3| +----+----+----+----+ ^^^^^ ^^^^^ LITTLEs bigs which would result in the following scheduler topology: MC [ ] <- sd_llc, sd_asym_cpucapacity 0 1 2 3 What this means is that, on DynamIQ systems, we could pass the wake_cap() test (IOW presume the waking task fits on the CPU capacities of some LLC domain), thus go through select_idle_sibling(). This function operates on an LLC domain, which here spans both bigs and LITTLEs, so it could very well pick a CPU of too small capacity for the task, despite there being fitting idle CPUs - it very much depends on the CPU iteration order, on which we have absolutely no guarantees capacity-wise. Implementation ============== Introduce yet another select_idle_sibling() helper function that takes CPU capacity into account. The policy is to pick the first idle CPU which is big enough for the task (task_util * margin < cpu_capacity). If no idle CPU is big enough, we pick the idle one with the highest capacity. Unlike other select_idle_sibling() helpers, this one operates on the sd_asym_cpucapacity sched_domain pointer, which is guaranteed to span all known CPU capacities in the system. As such, this will work for both "legacy" big.LITTLE (LITTLEs & bigs split at MC, joined at DIE) and for newer DynamIQ systems (e.g. LITTLEs and bigs in the same MC domain). Note that this limits the scope of select_idle_sibling() to select_idle_capacity() for asymmetric CPU capacity systems - the LLC domain will not be scanned, and no further heuristic will be applied. Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Valentin Schneider <valentin.schneider@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Reviewed-by: Quentin Perret <qperret@google.com> Link: https://lkml.kernel.org/r/20200206191957.12325-2-valentin.schneider@arm.com
2020-02-06 19:19:54 +00:00
/*
* Scan the asym_capacity domain for idle CPUs; pick the first idle one on which
* the task fits. If no CPU is big enough, but there are idle ones, try to
* maximize capacity.
*/
static int
select_idle_capacity(struct task_struct *p, struct sched_domain *sd, int target)
{
unsigned long best_cap = 0;
int cpu, best_cpu = -1;
struct cpumask *cpus;
sync_entity_load_avg(&p->se);
cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
for_each_cpu_wrap(cpu, cpus, target) {
unsigned long cpu_cap = capacity_of(cpu);
if (!available_idle_cpu(cpu) && !sched_idle_cpu(cpu))
continue;
if (task_fits_capacity(p, cpu_cap))
return cpu;
if (cpu_cap > best_cap) {
best_cap = cpu_cap;
best_cpu = cpu;
}
}
return best_cpu;
}
sched/core: Rewrite and improve select_idle_siblings() select_idle_siblings() is a known pain point for a number of workloads; it either does too much or not enough and sometimes just does plain wrong. This rewrite attempts to address a number of issues (but sadly not all). The current code does an unconditional sched_domain iteration; with the intent of finding an idle core (on SMT hardware). The problems which this patch tries to address are: - its pointless to look for idle cores if the machine is real busy; at which point you're just wasting cycles. - it's behaviour is inconsistent between SMT and !SMT hardware in that !SMT hardware ends up doing a scan for any idle CPU in the LLC domain, while SMT hardware does a scan for idle cores and if that fails, falls back to a scan for idle threads on the 'target' core. The new code replaces the sched_domain scan with 3 explicit scans: 1) search for an idle core in the LLC 2) search for an idle CPU in the LLC 3) search for an idle thread in the 'target' core where 1 and 3 are conditional on SMT support and 1 and 2 have runtime heuristics to skip the step. Step 1) is conditional on sd_llc_shared->has_idle_cores; when a cpu goes idle and sd_llc_shared->has_idle_cores is false, we scan all SMT siblings of the CPU going idle. Similarly, we clear sd_llc_shared->has_idle_cores when we fail to find an idle core. Step 2) tracks the average cost of the scan and compares this to the average idle time guestimate for the CPU doing the wakeup. There is a significant fudge factor involved to deal with the variability of the averages. Esp. hackbench was sensitive to this. Step 3) is unconditional; we assume (also per step 1) that scanning all SMT siblings in a core is 'cheap'. With this; SMT systems gain step 2, which cures a few benchmarks -- notably one from Facebook. One 'feature' of the sched_domain iteration, which we preserve in the new code, is that it would start scanning from the 'target' CPU, instead of scanning the cpumask in cpu id order. This avoids multiple CPUs in the LLC scanning for idle to gang up and find the same CPU quite as much. The down side is that tasks can end up hopping across the LLC for no apparent reason. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-05-09 08:38:05 +00:00
/*
* Try and locate an idle core/thread in the LLC cache domain.
*/
static int select_idle_sibling(struct task_struct *p, int prev, int target)
{
sched: Fix select_idle_sibling() logic in select_task_rq_fair() Issues in the current select_idle_sibling() logic in select_task_rq_fair() in the context of a task wake-up: a) Once we select the idle sibling, we use that domain (spanning the cpu that the task is currently woken-up and the idle sibling that we found) in our wake_affine() decisions. This domain is completely different from the domain(we are supposed to use) that spans the cpu that the task currently woken-up and the cpu where the task previously ran. b) We do select_idle_sibling() check only for the cpu that the task is currently woken-up on. If select_task_rq_fair() selects the previously run cpu for waking the task, doing a select_idle_sibling() check for that cpu also helps and we don't do this currently. c) In the scenarios where the cpu that the task is woken-up is busy but with its HT siblings are idle, we are selecting the task be woken-up on the idle HT sibling instead of a core that it previously ran and currently completely idle. i.e., we are not taking decisions based on wake_affine() but directly selecting an idle sibling that can cause an imbalance at the SMT/MC level which will be later corrected by the periodic load balancer. Fix this by first going through the load imbalance calculations using wake_affine() and once we make a decision of woken-up cpu vs previously-ran cpu, then choose a possible idle sibling for waking up the task on. Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> LKML-Reference: <1270079265.7835.8.camel@sbs-t61.sc.intel.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-03-31 23:47:45 +00:00
struct sched_domain *sd;
sched/fair: Use a recently used CPU as an idle candidate and the basis for SIS The select_idle_sibling() (SIS) rewrite in commit: 10e2f1acd010 ("sched/core: Rewrite and improve select_idle_siblings()") ... replaced a domain iteration with a search that broadly speaking does a wrapped walk of the scheduler domain sharing a last-level-cache. While this had a number of improvements, one consequence is that two tasks that share a waker/wakee relationship push each other around a socket. Even though two tasks may be active, all cores are evenly used. This is great from a search perspective and spreads a load across individual cores, but it has adverse consequences for cpufreq. As each CPU has relatively low utilisation, cpufreq may decide the utilisation is too low to used a higher P-state and overall computation throughput suffers. While individual cpufreq and cpuidle drivers may compensate by artifically boosting P-state (at c0) or avoiding lower C-states (during idle), it does not help if hardware-based cpufreq (e.g. HWP) is used. This patch tracks a recently used CPU based on what CPU a task was running on when it last was a waker a CPU it was recently using when a task is a wakee. During SIS, the recently used CPU is used as a target if it's still allowed by the task and is idle. The benefit may be non-obvious so consider an example of two tasks communicating back and forth. Task A may be an application doing IO where task B is a kworker or kthread like journald. Task A may issue IO, wake B and B wakes up A on completion. With the existing scheme this may look like the following (potentially different IDs if SMT is in use but similar principal applies). A (cpu 0) wake B (wakes on cpu 1) B (cpu 1) wake A (wakes on cpu 2) A (cpu 2) wake B (wakes on cpu 3) etc. A careful reader may wonder why CPU 0 was not idle when B wakes A the first time and it's simply due to the fact that A can be rescheduled to another CPU and the pattern is that prev == target when B tries to wakeup A and the information about CPU 0 has been lost. With this patch, the pattern is more likely to be: A (cpu 0) wake B (wakes on cpu 1) B (cpu 1) wake A (wakes on cpu 0) A (cpu 0) wake B (wakes on cpu 1) etc i.e. two communicating casts are more likely to use just two cores instead of all available cores sharing a LLC. The most dramatic speedup was noticed on dbench using the XFS filesystem on UMA as clients interact heavily with workqueues in that configuration. Note that a similar speedup is not observed on ext4 as the wakeup pattern is different: 4.15.0-rc9 4.15.0-rc9 waprev-v1 biasancestor-v1 Hmean 1 287.54 ( 0.00%) 817.01 ( 184.14%) Hmean 2 1268.12 ( 0.00%) 1781.24 ( 40.46%) Hmean 4 1739.68 ( 0.00%) 1594.47 ( -8.35%) Hmean 8 2464.12 ( 0.00%) 2479.56 ( 0.63%) Hmean 64 1455.57 ( 0.00%) 1434.68 ( -1.44%) The results can be less dramatic on NUMA where automatic balancing interferes with the test. It's also known that network benchmarks running on localhost also benefit quite a bit from this patch (roughly 10% on netperf RR for UDP and TCP depending on the machine). Hackbench also seens small improvements (6-11% depending on machine and thread count). The facebook schbench was also tested but in most cases showed little or no different to wakeup latencies. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Matt Fleming <matt@codeblueprint.co.uk> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/20180130104555.4125-5-mgorman@techsingularity.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-01-30 10:45:55 +00:00
int i, recent_used_cpu;
sched/fair: Add asymmetric CPU capacity wakeup scan Issue ===== On asymmetric CPU capacity topologies, we currently rely on wake_cap() to drive select_task_rq_fair() towards either: - its slow-path (find_idlest_cpu()) if either the previous or current (waking) CPU has too little capacity for the waking task - its fast-path (select_idle_sibling()) otherwise Commit: 3273163c6775 ("sched/fair: Let asymmetric CPU configurations balance at wake-up") points out that this relies on the assumption that "[...]the CPU capacities within an SD_SHARE_PKG_RESOURCES domain (sd_llc) are homogeneous". This assumption no longer holds on newer generations of big.LITTLE systems (DynamIQ), which can accommodate CPUs of different compute capacity within a single LLC domain. To hopefully paint a better picture, a regular big.LITTLE topology would look like this: +---------+ +---------+ | L2 | | L2 | +----+----+ +----+----+ |CPU0|CPU1| |CPU2|CPU3| +----+----+ +----+----+ ^^^ ^^^ LITTLEs bigs which would result in the following scheduler topology: DIE [ ] <- sd_asym_cpucapacity MC [ ] [ ] <- sd_llc 0 1 2 3 Conversely, a DynamIQ topology could look like: +-------------------+ | L3 | +----+----+----+----+ | L2 | L2 | L2 | L2 | +----+----+----+----+ |CPU0|CPU1|CPU2|CPU3| +----+----+----+----+ ^^^^^ ^^^^^ LITTLEs bigs which would result in the following scheduler topology: MC [ ] <- sd_llc, sd_asym_cpucapacity 0 1 2 3 What this means is that, on DynamIQ systems, we could pass the wake_cap() test (IOW presume the waking task fits on the CPU capacities of some LLC domain), thus go through select_idle_sibling(). This function operates on an LLC domain, which here spans both bigs and LITTLEs, so it could very well pick a CPU of too small capacity for the task, despite there being fitting idle CPUs - it very much depends on the CPU iteration order, on which we have absolutely no guarantees capacity-wise. Implementation ============== Introduce yet another select_idle_sibling() helper function that takes CPU capacity into account. The policy is to pick the first idle CPU which is big enough for the task (task_util * margin < cpu_capacity). If no idle CPU is big enough, we pick the idle one with the highest capacity. Unlike other select_idle_sibling() helpers, this one operates on the sd_asym_cpucapacity sched_domain pointer, which is guaranteed to span all known CPU capacities in the system. As such, this will work for both "legacy" big.LITTLE (LITTLEs & bigs split at MC, joined at DIE) and for newer DynamIQ systems (e.g. LITTLEs and bigs in the same MC domain). Note that this limits the scope of select_idle_sibling() to select_idle_capacity() for asymmetric CPU capacity systems - the LLC domain will not be scanned, and no further heuristic will be applied. Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Valentin Schneider <valentin.schneider@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Reviewed-by: Quentin Perret <qperret@google.com> Link: https://lkml.kernel.org/r/20200206191957.12325-2-valentin.schneider@arm.com
2020-02-06 19:19:54 +00:00
/*
* For asymmetric CPU capacity systems, our domain of interest is
* sd_asym_cpucapacity rather than sd_llc.
*/
if (static_branch_unlikely(&sched_asym_cpucapacity)) {
sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, target));
/*
* On an asymmetric CPU capacity system where an exclusive
* cpuset defines a symmetric island (i.e. one unique
* capacity_orig value through the cpuset), the key will be set
* but the CPUs within that cpuset will not have a domain with
* SD_ASYM_CPUCAPACITY. These should follow the usual symmetric
* capacity path.
*/
if (!sd)
goto symmetric;
i = select_idle_capacity(p, sd, target);
return ((unsigned)i < nr_cpumask_bits) ? i : target;
}
symmetric:
if (available_idle_cpu(target) || sched_idle_cpu(target))
return target;
sched: Fix select_idle_sibling() logic in select_task_rq_fair() Issues in the current select_idle_sibling() logic in select_task_rq_fair() in the context of a task wake-up: a) Once we select the idle sibling, we use that domain (spanning the cpu that the task is currently woken-up and the idle sibling that we found) in our wake_affine() decisions. This domain is completely different from the domain(we are supposed to use) that spans the cpu that the task currently woken-up and the cpu where the task previously ran. b) We do select_idle_sibling() check only for the cpu that the task is currently woken-up on. If select_task_rq_fair() selects the previously run cpu for waking the task, doing a select_idle_sibling() check for that cpu also helps and we don't do this currently. c) In the scenarios where the cpu that the task is woken-up is busy but with its HT siblings are idle, we are selecting the task be woken-up on the idle HT sibling instead of a core that it previously ran and currently completely idle. i.e., we are not taking decisions based on wake_affine() but directly selecting an idle sibling that can cause an imbalance at the SMT/MC level which will be later corrected by the periodic load balancer. Fix this by first going through the load imbalance calculations using wake_affine() and once we make a decision of woken-up cpu vs previously-ran cpu, then choose a possible idle sibling for waking up the task on. Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> LKML-Reference: <1270079265.7835.8.camel@sbs-t61.sc.intel.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-03-31 23:47:45 +00:00
/*
* If the previous CPU is cache affine and idle, don't be stupid:
sched: Fix select_idle_sibling() logic in select_task_rq_fair() Issues in the current select_idle_sibling() logic in select_task_rq_fair() in the context of a task wake-up: a) Once we select the idle sibling, we use that domain (spanning the cpu that the task is currently woken-up and the idle sibling that we found) in our wake_affine() decisions. This domain is completely different from the domain(we are supposed to use) that spans the cpu that the task currently woken-up and the cpu where the task previously ran. b) We do select_idle_sibling() check only for the cpu that the task is currently woken-up on. If select_task_rq_fair() selects the previously run cpu for waking the task, doing a select_idle_sibling() check for that cpu also helps and we don't do this currently. c) In the scenarios where the cpu that the task is woken-up is busy but with its HT siblings are idle, we are selecting the task be woken-up on the idle HT sibling instead of a core that it previously ran and currently completely idle. i.e., we are not taking decisions based on wake_affine() but directly selecting an idle sibling that can cause an imbalance at the SMT/MC level which will be later corrected by the periodic load balancer. Fix this by first going through the load imbalance calculations using wake_affine() and once we make a decision of woken-up cpu vs previously-ran cpu, then choose a possible idle sibling for waking up the task on. Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> LKML-Reference: <1270079265.7835.8.camel@sbs-t61.sc.intel.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-03-31 23:47:45 +00:00
*/
if (prev != target && cpus_share_cache(prev, target) &&
(available_idle_cpu(prev) || sched_idle_cpu(prev)))
return prev;
sched/fair: Allow a per-CPU kthread waking a task to stack on the same CPU, to fix XFS performance regression The following XFS commit: 8ab39f11d974 ("xfs: prevent CIL push holdoff in log recovery") changed the logic from using bound workqueues to using unbound workqueues. Functionally this makes sense but it was observed at the time that the dbench performance dropped quite a lot and CPU migrations were increased. The current pattern of the task migration is straight-forward. With XFS, an IO issuer delegates work to xlog_cil_push_work ()on an unbound kworker. This runs on a nearby CPU and on completion, dbench wakes up on its old CPU as it is still idle and no migration occurs. dbench then queues the real IO on the blk_mq_requeue_work() work item which runs on a bound kworker which is forced to run on the same CPU as dbench. When IO completes, the bound kworker wakes dbench but as the kworker is a bound but, real task, the CPU is not considered idle and dbench gets migrated by select_idle_sibling() to a new CPU. dbench may ping-pong between two CPUs for a while but ultimately it starts a round-robin of all CPUs sharing the same LLC. High-frequency migration on each IO completion has poor performance overall. It has negative implications both in commication costs and power management. mpstat confirmed that at low thread counts that all CPUs sharing an LLC has low level of activity. Note that even if the CIL patch was reverted, there still would be migrations but the impact is less noticeable. It turns out that individually the scheduler, XFS, blk-mq and workqueues all made sensible decisions but in combination, the overall effect was sub-optimal. This patch special cases the IO issue/completion pattern and allows a bound kworker waker and a task wakee to stack on the same CPU if there is a strong chance they are directly related. The expectation is that the kworker is likely going back to sleep shortly. This is not guaranteed as the IO could be queued asynchronously but there is a very strong relationship between the task and kworker in this case that would justify stacking on the same CPU instead of migrating. There should be few concerns about kworker starvation given that the special casing is only when the kworker is the waker. DBench on XFS MMTests config: io-dbench4-async modified to run on a fresh XFS filesystem UMA machine with 8 cores sharing LLC 5.5.0-rc7 5.5.0-rc7 tipsched-20200124 kworkerstack Amean 1 22.63 ( 0.00%) 20.54 * 9.23%* Amean 2 25.56 ( 0.00%) 23.40 * 8.44%* Amean 4 28.63 ( 0.00%) 27.85 * 2.70%* Amean 8 37.66 ( 0.00%) 37.68 ( -0.05%) Amean 64 469.47 ( 0.00%) 468.26 ( 0.26%) Stddev 1 1.00 ( 0.00%) 0.72 ( 28.12%) Stddev 2 1.62 ( 0.00%) 1.97 ( -21.54%) Stddev 4 2.53 ( 0.00%) 3.58 ( -41.19%) Stddev 8 5.30 ( 0.00%) 5.20 ( 1.92%) Stddev 64 86.36 ( 0.00%) 94.53 ( -9.46%) NUMA machine, 48 CPUs total, 24 CPUs share cache 5.5.0-rc7 5.5.0-rc7 tipsched-20200124 kworkerstack-v1r2 Amean 1 58.69 ( 0.00%) 30.21 * 48.53%* Amean 2 60.90 ( 0.00%) 35.29 * 42.05%* Amean 4 66.77 ( 0.00%) 46.55 * 30.28%* Amean 8 81.41 ( 0.00%) 68.46 * 15.91%* Amean 16 113.29 ( 0.00%) 107.79 * 4.85%* Amean 32 199.10 ( 0.00%) 198.22 * 0.44%* Amean 64 478.99 ( 0.00%) 477.06 * 0.40%* Amean 128 1345.26 ( 0.00%) 1372.64 * -2.04%* Stddev 1 2.64 ( 0.00%) 4.17 ( -58.08%) Stddev 2 4.35 ( 0.00%) 5.38 ( -23.73%) Stddev 4 6.77 ( 0.00%) 6.56 ( 3.00%) Stddev 8 11.61 ( 0.00%) 10.91 ( 6.04%) Stddev 16 18.63 ( 0.00%) 19.19 ( -3.01%) Stddev 32 38.71 ( 0.00%) 38.30 ( 1.06%) Stddev 64 100.28 ( 0.00%) 91.24 ( 9.02%) Stddev 128 186.87 ( 0.00%) 160.34 ( 14.20%) Dbench has been modified to report the time to complete a single "load file". This is a more meaningful metric for dbench that a throughput metric as the benchmark makes many different system calls that are not throughput-related Patch shows a 9.23% and 48.53% reduction in the time to process a load file with the difference partially explained by the number of CPUs sharing a LLC. In a separate run, task migrations were almost eliminated by the patch for low client counts. In case people have issue with the metric used for the benchmark, this is a comparison of the throughputs as reported by dbench on the NUMA machine. dbench4 Throughput (misleading but traditional) 5.5.0-rc7 5.5.0-rc7 tipsched-20200124 kworkerstack-v1r2 Hmean 1 321.41 ( 0.00%) 617.82 * 92.22%* Hmean 2 622.87 ( 0.00%) 1066.80 * 71.27%* Hmean 4 1134.56 ( 0.00%) 1623.74 * 43.12%* Hmean 8 1869.96 ( 0.00%) 2212.67 * 18.33%* Hmean 16 2673.11 ( 0.00%) 2806.13 * 4.98%* Hmean 32 3032.74 ( 0.00%) 3039.54 ( 0.22%) Hmean 64 2514.25 ( 0.00%) 2498.96 * -0.61%* Hmean 128 1778.49 ( 0.00%) 1746.05 * -1.82%* Note that this is somewhat specific to XFS and ext4 shows no performance difference as it does not rely on kworkers in the same way. No major problem was observed running other workloads on different machines although not all tests have completed yet. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/20200128154006.GD3466@techsingularity.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2020-01-28 15:40:06 +00:00
/*
* Allow a per-cpu kthread to stack with the wakee if the
* kworker thread and the tasks previous CPUs are the same.
* The assumption is that the wakee queued work for the
* per-cpu kthread that is now complete and the wakeup is
* essentially a sync wakeup. An obvious example of this
* pattern is IO completions.
*/
if (is_per_cpu_kthread(current) &&
prev == smp_processor_id() &&
this_rq()->nr_running <= 1) {
return prev;
}
/* Check a recently used CPU as a potential idle candidate: */
sched/fair: Use a recently used CPU as an idle candidate and the basis for SIS The select_idle_sibling() (SIS) rewrite in commit: 10e2f1acd010 ("sched/core: Rewrite and improve select_idle_siblings()") ... replaced a domain iteration with a search that broadly speaking does a wrapped walk of the scheduler domain sharing a last-level-cache. While this had a number of improvements, one consequence is that two tasks that share a waker/wakee relationship push each other around a socket. Even though two tasks may be active, all cores are evenly used. This is great from a search perspective and spreads a load across individual cores, but it has adverse consequences for cpufreq. As each CPU has relatively low utilisation, cpufreq may decide the utilisation is too low to used a higher P-state and overall computation throughput suffers. While individual cpufreq and cpuidle drivers may compensate by artifically boosting P-state (at c0) or avoiding lower C-states (during idle), it does not help if hardware-based cpufreq (e.g. HWP) is used. This patch tracks a recently used CPU based on what CPU a task was running on when it last was a waker a CPU it was recently using when a task is a wakee. During SIS, the recently used CPU is used as a target if it's still allowed by the task and is idle. The benefit may be non-obvious so consider an example of two tasks communicating back and forth. Task A may be an application doing IO where task B is a kworker or kthread like journald. Task A may issue IO, wake B and B wakes up A on completion. With the existing scheme this may look like the following (potentially different IDs if SMT is in use but similar principal applies). A (cpu 0) wake B (wakes on cpu 1) B (cpu 1) wake A (wakes on cpu 2) A (cpu 2) wake B (wakes on cpu 3) etc. A careful reader may wonder why CPU 0 was not idle when B wakes A the first time and it's simply due to the fact that A can be rescheduled to another CPU and the pattern is that prev == target when B tries to wakeup A and the information about CPU 0 has been lost. With this patch, the pattern is more likely to be: A (cpu 0) wake B (wakes on cpu 1) B (cpu 1) wake A (wakes on cpu 0) A (cpu 0) wake B (wakes on cpu 1) etc i.e. two communicating casts are more likely to use just two cores instead of all available cores sharing a LLC. The most dramatic speedup was noticed on dbench using the XFS filesystem on UMA as clients interact heavily with workqueues in that configuration. Note that a similar speedup is not observed on ext4 as the wakeup pattern is different: 4.15.0-rc9 4.15.0-rc9 waprev-v1 biasancestor-v1 Hmean 1 287.54 ( 0.00%) 817.01 ( 184.14%) Hmean 2 1268.12 ( 0.00%) 1781.24 ( 40.46%) Hmean 4 1739.68 ( 0.00%) 1594.47 ( -8.35%) Hmean 8 2464.12 ( 0.00%) 2479.56 ( 0.63%) Hmean 64 1455.57 ( 0.00%) 1434.68 ( -1.44%) The results can be less dramatic on NUMA where automatic balancing interferes with the test. It's also known that network benchmarks running on localhost also benefit quite a bit from this patch (roughly 10% on netperf RR for UDP and TCP depending on the machine). Hackbench also seens small improvements (6-11% depending on machine and thread count). The facebook schbench was also tested but in most cases showed little or no different to wakeup latencies. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Matt Fleming <matt@codeblueprint.co.uk> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/20180130104555.4125-5-mgorman@techsingularity.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-01-30 10:45:55 +00:00
recent_used_cpu = p->recent_used_cpu;
if (recent_used_cpu != prev &&
recent_used_cpu != target &&
cpus_share_cache(recent_used_cpu, target) &&
(available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) &&
cpumask_test_cpu(p->recent_used_cpu, p->cpus_ptr)) {
sched/fair: Use a recently used CPU as an idle candidate and the basis for SIS The select_idle_sibling() (SIS) rewrite in commit: 10e2f1acd010 ("sched/core: Rewrite and improve select_idle_siblings()") ... replaced a domain iteration with a search that broadly speaking does a wrapped walk of the scheduler domain sharing a last-level-cache. While this had a number of improvements, one consequence is that two tasks that share a waker/wakee relationship push each other around a socket. Even though two tasks may be active, all cores are evenly used. This is great from a search perspective and spreads a load across individual cores, but it has adverse consequences for cpufreq. As each CPU has relatively low utilisation, cpufreq may decide the utilisation is too low to used a higher P-state and overall computation throughput suffers. While individual cpufreq and cpuidle drivers may compensate by artifically boosting P-state (at c0) or avoiding lower C-states (during idle), it does not help if hardware-based cpufreq (e.g. HWP) is used. This patch tracks a recently used CPU based on what CPU a task was running on when it last was a waker a CPU it was recently using when a task is a wakee. During SIS, the recently used CPU is used as a target if it's still allowed by the task and is idle. The benefit may be non-obvious so consider an example of two tasks communicating back and forth. Task A may be an application doing IO where task B is a kworker or kthread like journald. Task A may issue IO, wake B and B wakes up A on completion. With the existing scheme this may look like the following (potentially different IDs if SMT is in use but similar principal applies). A (cpu 0) wake B (wakes on cpu 1) B (cpu 1) wake A (wakes on cpu 2) A (cpu 2) wake B (wakes on cpu 3) etc. A careful reader may wonder why CPU 0 was not idle when B wakes A the first time and it's simply due to the fact that A can be rescheduled to another CPU and the pattern is that prev == target when B tries to wakeup A and the information about CPU 0 has been lost. With this patch, the pattern is more likely to be: A (cpu 0) wake B (wakes on cpu 1) B (cpu 1) wake A (wakes on cpu 0) A (cpu 0) wake B (wakes on cpu 1) etc i.e. two communicating casts are more likely to use just two cores instead of all available cores sharing a LLC. The most dramatic speedup was noticed on dbench using the XFS filesystem on UMA as clients interact heavily with workqueues in that configuration. Note that a similar speedup is not observed on ext4 as the wakeup pattern is different: 4.15.0-rc9 4.15.0-rc9 waprev-v1 biasancestor-v1 Hmean 1 287.54 ( 0.00%) 817.01 ( 184.14%) Hmean 2 1268.12 ( 0.00%) 1781.24 ( 40.46%) Hmean 4 1739.68 ( 0.00%) 1594.47 ( -8.35%) Hmean 8 2464.12 ( 0.00%) 2479.56 ( 0.63%) Hmean 64 1455.57 ( 0.00%) 1434.68 ( -1.44%) The results can be less dramatic on NUMA where automatic balancing interferes with the test. It's also known that network benchmarks running on localhost also benefit quite a bit from this patch (roughly 10% on netperf RR for UDP and TCP depending on the machine). Hackbench also seens small improvements (6-11% depending on machine and thread count). The facebook schbench was also tested but in most cases showed little or no different to wakeup latencies. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Matt Fleming <matt@codeblueprint.co.uk> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/20180130104555.4125-5-mgorman@techsingularity.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-01-30 10:45:55 +00:00
/*
* Replace recent_used_cpu with prev as it is a potential
* candidate for the next wake:
sched/fair: Use a recently used CPU as an idle candidate and the basis for SIS The select_idle_sibling() (SIS) rewrite in commit: 10e2f1acd010 ("sched/core: Rewrite and improve select_idle_siblings()") ... replaced a domain iteration with a search that broadly speaking does a wrapped walk of the scheduler domain sharing a last-level-cache. While this had a number of improvements, one consequence is that two tasks that share a waker/wakee relationship push each other around a socket. Even though two tasks may be active, all cores are evenly used. This is great from a search perspective and spreads a load across individual cores, but it has adverse consequences for cpufreq. As each CPU has relatively low utilisation, cpufreq may decide the utilisation is too low to used a higher P-state and overall computation throughput suffers. While individual cpufreq and cpuidle drivers may compensate by artifically boosting P-state (at c0) or avoiding lower C-states (during idle), it does not help if hardware-based cpufreq (e.g. HWP) is used. This patch tracks a recently used CPU based on what CPU a task was running on when it last was a waker a CPU it was recently using when a task is a wakee. During SIS, the recently used CPU is used as a target if it's still allowed by the task and is idle. The benefit may be non-obvious so consider an example of two tasks communicating back and forth. Task A may be an application doing IO where task B is a kworker or kthread like journald. Task A may issue IO, wake B and B wakes up A on completion. With the existing scheme this may look like the following (potentially different IDs if SMT is in use but similar principal applies). A (cpu 0) wake B (wakes on cpu 1) B (cpu 1) wake A (wakes on cpu 2) A (cpu 2) wake B (wakes on cpu 3) etc. A careful reader may wonder why CPU 0 was not idle when B wakes A the first time and it's simply due to the fact that A can be rescheduled to another CPU and the pattern is that prev == target when B tries to wakeup A and the information about CPU 0 has been lost. With this patch, the pattern is more likely to be: A (cpu 0) wake B (wakes on cpu 1) B (cpu 1) wake A (wakes on cpu 0) A (cpu 0) wake B (wakes on cpu 1) etc i.e. two communicating casts are more likely to use just two cores instead of all available cores sharing a LLC. The most dramatic speedup was noticed on dbench using the XFS filesystem on UMA as clients interact heavily with workqueues in that configuration. Note that a similar speedup is not observed on ext4 as the wakeup pattern is different: 4.15.0-rc9 4.15.0-rc9 waprev-v1 biasancestor-v1 Hmean 1 287.54 ( 0.00%) 817.01 ( 184.14%) Hmean 2 1268.12 ( 0.00%) 1781.24 ( 40.46%) Hmean 4 1739.68 ( 0.00%) 1594.47 ( -8.35%) Hmean 8 2464.12 ( 0.00%) 2479.56 ( 0.63%) Hmean 64 1455.57 ( 0.00%) 1434.68 ( -1.44%) The results can be less dramatic on NUMA where automatic balancing interferes with the test. It's also known that network benchmarks running on localhost also benefit quite a bit from this patch (roughly 10% on netperf RR for UDP and TCP depending on the machine). Hackbench also seens small improvements (6-11% depending on machine and thread count). The facebook schbench was also tested but in most cases showed little or no different to wakeup latencies. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Matt Fleming <matt@codeblueprint.co.uk> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/20180130104555.4125-5-mgorman@techsingularity.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-01-30 10:45:55 +00:00
*/
p->recent_used_cpu = prev;
return recent_used_cpu;
}
sd = rcu_dereference(per_cpu(sd_llc, target));
sched/core: Rewrite and improve select_idle_siblings() select_idle_siblings() is a known pain point for a number of workloads; it either does too much or not enough and sometimes just does plain wrong. This rewrite attempts to address a number of issues (but sadly not all). The current code does an unconditional sched_domain iteration; with the intent of finding an idle core (on SMT hardware). The problems which this patch tries to address are: - its pointless to look for idle cores if the machine is real busy; at which point you're just wasting cycles. - it's behaviour is inconsistent between SMT and !SMT hardware in that !SMT hardware ends up doing a scan for any idle CPU in the LLC domain, while SMT hardware does a scan for idle cores and if that fails, falls back to a scan for idle threads on the 'target' core. The new code replaces the sched_domain scan with 3 explicit scans: 1) search for an idle core in the LLC 2) search for an idle CPU in the LLC 3) search for an idle thread in the 'target' core where 1 and 3 are conditional on SMT support and 1 and 2 have runtime heuristics to skip the step. Step 1) is conditional on sd_llc_shared->has_idle_cores; when a cpu goes idle and sd_llc_shared->has_idle_cores is false, we scan all SMT siblings of the CPU going idle. Similarly, we clear sd_llc_shared->has_idle_cores when we fail to find an idle core. Step 2) tracks the average cost of the scan and compares this to the average idle time guestimate for the CPU doing the wakeup. There is a significant fudge factor involved to deal with the variability of the averages. Esp. hackbench was sensitive to this. Step 3) is unconditional; we assume (also per step 1) that scanning all SMT siblings in a core is 'cheap'. With this; SMT systems gain step 2, which cures a few benchmarks -- notably one from Facebook. One 'feature' of the sched_domain iteration, which we preserve in the new code, is that it would start scanning from the 'target' CPU, instead of scanning the cpumask in cpu id order. This avoids multiple CPUs in the LLC scanning for idle to gang up and find the same CPU quite as much. The down side is that tasks can end up hopping across the LLC for no apparent reason. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-05-09 08:38:05 +00:00
if (!sd)
return target;
sched/core: Rewrite and improve select_idle_siblings() select_idle_siblings() is a known pain point for a number of workloads; it either does too much or not enough and sometimes just does plain wrong. This rewrite attempts to address a number of issues (but sadly not all). The current code does an unconditional sched_domain iteration; with the intent of finding an idle core (on SMT hardware). The problems which this patch tries to address are: - its pointless to look for idle cores if the machine is real busy; at which point you're just wasting cycles. - it's behaviour is inconsistent between SMT and !SMT hardware in that !SMT hardware ends up doing a scan for any idle CPU in the LLC domain, while SMT hardware does a scan for idle cores and if that fails, falls back to a scan for idle threads on the 'target' core. The new code replaces the sched_domain scan with 3 explicit scans: 1) search for an idle core in the LLC 2) search for an idle CPU in the LLC 3) search for an idle thread in the 'target' core where 1 and 3 are conditional on SMT support and 1 and 2 have runtime heuristics to skip the step. Step 1) is conditional on sd_llc_shared->has_idle_cores; when a cpu goes idle and sd_llc_shared->has_idle_cores is false, we scan all SMT siblings of the CPU going idle. Similarly, we clear sd_llc_shared->has_idle_cores when we fail to find an idle core. Step 2) tracks the average cost of the scan and compares this to the average idle time guestimate for the CPU doing the wakeup. There is a significant fudge factor involved to deal with the variability of the averages. Esp. hackbench was sensitive to this. Step 3) is unconditional; we assume (also per step 1) that scanning all SMT siblings in a core is 'cheap'. With this; SMT systems gain step 2, which cures a few benchmarks -- notably one from Facebook. One 'feature' of the sched_domain iteration, which we preserve in the new code, is that it would start scanning from the 'target' CPU, instead of scanning the cpumask in cpu id order. This avoids multiple CPUs in the LLC scanning for idle to gang up and find the same CPU quite as much. The down side is that tasks can end up hopping across the LLC for no apparent reason. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-05-09 08:38:05 +00:00
i = select_idle_core(p, sd, target);
if ((unsigned)i < nr_cpumask_bits)
return i;
sched/core: Rewrite and improve select_idle_siblings() select_idle_siblings() is a known pain point for a number of workloads; it either does too much or not enough and sometimes just does plain wrong. This rewrite attempts to address a number of issues (but sadly not all). The current code does an unconditional sched_domain iteration; with the intent of finding an idle core (on SMT hardware). The problems which this patch tries to address are: - its pointless to look for idle cores if the machine is real busy; at which point you're just wasting cycles. - it's behaviour is inconsistent between SMT and !SMT hardware in that !SMT hardware ends up doing a scan for any idle CPU in the LLC domain, while SMT hardware does a scan for idle cores and if that fails, falls back to a scan for idle threads on the 'target' core. The new code replaces the sched_domain scan with 3 explicit scans: 1) search for an idle core in the LLC 2) search for an idle CPU in the LLC 3) search for an idle thread in the 'target' core where 1 and 3 are conditional on SMT support and 1 and 2 have runtime heuristics to skip the step. Step 1) is conditional on sd_llc_shared->has_idle_cores; when a cpu goes idle and sd_llc_shared->has_idle_cores is false, we scan all SMT siblings of the CPU going idle. Similarly, we clear sd_llc_shared->has_idle_cores when we fail to find an idle core. Step 2) tracks the average cost of the scan and compares this to the average idle time guestimate for the CPU doing the wakeup. There is a significant fudge factor involved to deal with the variability of the averages. Esp. hackbench was sensitive to this. Step 3) is unconditional; we assume (also per step 1) that scanning all SMT siblings in a core is 'cheap'. With this; SMT systems gain step 2, which cures a few benchmarks -- notably one from Facebook. One 'feature' of the sched_domain iteration, which we preserve in the new code, is that it would start scanning from the 'target' CPU, instead of scanning the cpumask in cpu id order. This avoids multiple CPUs in the LLC scanning for idle to gang up and find the same CPU quite as much. The down side is that tasks can end up hopping across the LLC for no apparent reason. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-05-09 08:38:05 +00:00
i = select_idle_cpu(p, sd, target);
if ((unsigned)i < nr_cpumask_bits)
return i;
i = select_idle_smt(p, target);
sched/core: Rewrite and improve select_idle_siblings() select_idle_siblings() is a known pain point for a number of workloads; it either does too much or not enough and sometimes just does plain wrong. This rewrite attempts to address a number of issues (but sadly not all). The current code does an unconditional sched_domain iteration; with the intent of finding an idle core (on SMT hardware). The problems which this patch tries to address are: - its pointless to look for idle cores if the machine is real busy; at which point you're just wasting cycles. - it's behaviour is inconsistent between SMT and !SMT hardware in that !SMT hardware ends up doing a scan for any idle CPU in the LLC domain, while SMT hardware does a scan for idle cores and if that fails, falls back to a scan for idle threads on the 'target' core. The new code replaces the sched_domain scan with 3 explicit scans: 1) search for an idle core in the LLC 2) search for an idle CPU in the LLC 3) search for an idle thread in the 'target' core where 1 and 3 are conditional on SMT support and 1 and 2 have runtime heuristics to skip the step. Step 1) is conditional on sd_llc_shared->has_idle_cores; when a cpu goes idle and sd_llc_shared->has_idle_cores is false, we scan all SMT siblings of the CPU going idle. Similarly, we clear sd_llc_shared->has_idle_cores when we fail to find an idle core. Step 2) tracks the average cost of the scan and compares this to the average idle time guestimate for the CPU doing the wakeup. There is a significant fudge factor involved to deal with the variability of the averages. Esp. hackbench was sensitive to this. Step 3) is unconditional; we assume (also per step 1) that scanning all SMT siblings in a core is 'cheap'. With this; SMT systems gain step 2, which cures a few benchmarks -- notably one from Facebook. One 'feature' of the sched_domain iteration, which we preserve in the new code, is that it would start scanning from the 'target' CPU, instead of scanning the cpumask in cpu id order. This avoids multiple CPUs in the LLC scanning for idle to gang up and find the same CPU quite as much. The down side is that tasks can end up hopping across the LLC for no apparent reason. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-05-09 08:38:05 +00:00
if ((unsigned)i < nr_cpumask_bits)
return i;
return target;
}
sched/fair: Use util_est in LB and WU paths When the scheduler looks at the CPU utilization, the current PELT value for a CPU is returned straight away. In certain scenarios this can have undesired side effects on task placement. For example, since the task utilization is decayed at wakeup time, when a long sleeping big task is enqueued it does not add immediately a significant contribution to the target CPU. As a result we generate a race condition where other tasks can be placed on the same CPU while it is still considered relatively empty. In order to reduce this kind of race conditions, this patch introduces the required support to integrate the usage of the CPU's estimated utilization in the wakeup path, via cpu_util_wake(), as well as in the load-balance path, via cpu_util() which is used by update_sg_lb_stats(). The estimated utilization of a CPU is defined to be the maximum between its PELT's utilization and the sum of the estimated utilization (at previous dequeue time) of all the tasks currently RUNNABLE on that CPU. This allows to properly represent the spare capacity of a CPU which, for example, has just got a big task running since a long sleep period. Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Joel Fernandes <joelaf@google.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rafael J . Wysocki <rafael.j.wysocki@intel.com> Cc: Steve Muckle <smuckle@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Todd Kjos <tkjos@android.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Link: http://lkml.kernel.org/r/20180309095245.11071-3-patrick.bellasi@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-03-09 09:52:43 +00:00
/**
* Amount of capacity of a CPU that is (estimated to be) used by CFS tasks
* @cpu: the CPU to get the utilization of
*
* The unit of the return value must be the one of capacity so we can compare
* the utilization with the capacity of the CPU that is available for CFS task
* (ie cpu_capacity).
*
* cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
* recent utilization of currently non-runnable tasks on a CPU. It represents
* the amount of utilization of a CPU in the range [0..capacity_orig] where
* capacity_orig is the cpu_capacity available at the highest frequency
* (arch_scale_freq_capacity()).
* The utilization of a CPU converges towards a sum equal to or less than the
* current capacity (capacity_curr <= capacity_orig) of the CPU because it is
* the running time on this CPU scaled by capacity_curr.
*
sched/fair: Use util_est in LB and WU paths When the scheduler looks at the CPU utilization, the current PELT value for a CPU is returned straight away. In certain scenarios this can have undesired side effects on task placement. For example, since the task utilization is decayed at wakeup time, when a long sleeping big task is enqueued it does not add immediately a significant contribution to the target CPU. As a result we generate a race condition where other tasks can be placed on the same CPU while it is still considered relatively empty. In order to reduce this kind of race conditions, this patch introduces the required support to integrate the usage of the CPU's estimated utilization in the wakeup path, via cpu_util_wake(), as well as in the load-balance path, via cpu_util() which is used by update_sg_lb_stats(). The estimated utilization of a CPU is defined to be the maximum between its PELT's utilization and the sum of the estimated utilization (at previous dequeue time) of all the tasks currently RUNNABLE on that CPU. This allows to properly represent the spare capacity of a CPU which, for example, has just got a big task running since a long sleep period. Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Joel Fernandes <joelaf@google.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rafael J . Wysocki <rafael.j.wysocki@intel.com> Cc: Steve Muckle <smuckle@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Todd Kjos <tkjos@android.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Link: http://lkml.kernel.org/r/20180309095245.11071-3-patrick.bellasi@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-03-09 09:52:43 +00:00
* The estimated utilization of a CPU is defined to be the maximum between its
* cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks
* currently RUNNABLE on that CPU.
* This allows to properly represent the expected utilization of a CPU which
* has just got a big task running since a long sleep period. At the same time
* however it preserves the benefits of the "blocked utilization" in
* describing the potential for other tasks waking up on the same CPU.
*
* Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
* higher than capacity_orig because of unfortunate rounding in
* cfs.avg.util_avg or just after migrating tasks and new task wakeups until
* the average stabilizes with the new running time. We need to check that the
* utilization stays within the range of [0..capacity_orig] and cap it if
* necessary. Without utilization capping, a group could be seen as overloaded
* (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
* available capacity. We allow utilization to overshoot capacity_curr (but not
* capacity_orig) as it useful for predicting the capacity required after task
* migrations (scheduler-driven DVFS).
sched/fair: Use util_est in LB and WU paths When the scheduler looks at the CPU utilization, the current PELT value for a CPU is returned straight away. In certain scenarios this can have undesired side effects on task placement. For example, since the task utilization is decayed at wakeup time, when a long sleeping big task is enqueued it does not add immediately a significant contribution to the target CPU. As a result we generate a race condition where other tasks can be placed on the same CPU while it is still considered relatively empty. In order to reduce this kind of race conditions, this patch introduces the required support to integrate the usage of the CPU's estimated utilization in the wakeup path, via cpu_util_wake(), as well as in the load-balance path, via cpu_util() which is used by update_sg_lb_stats(). The estimated utilization of a CPU is defined to be the maximum between its PELT's utilization and the sum of the estimated utilization (at previous dequeue time) of all the tasks currently RUNNABLE on that CPU. This allows to properly represent the spare capacity of a CPU which, for example, has just got a big task running since a long sleep period. Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Joel Fernandes <joelaf@google.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rafael J . Wysocki <rafael.j.wysocki@intel.com> Cc: Steve Muckle <smuckle@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Todd Kjos <tkjos@android.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Link: http://lkml.kernel.org/r/20180309095245.11071-3-patrick.bellasi@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-03-09 09:52:43 +00:00
*
* Return: the (estimated) utilization for the specified CPU
sched: Calculate CPU's usage statistic and put it into struct sg_lb_stats::group_usage Monitor the usage level of each group of each sched_domain level. The usage is the portion of cpu_capacity_orig that is currently used on a CPU or group of CPUs. We use the utilization_load_avg to evaluate the usage level of each group. The utilization_load_avg only takes into account the running time of the CFS tasks on a CPU with a maximum value of SCHED_LOAD_SCALE when the CPU is fully utilized. Nevertheless, we must cap utilization_load_avg which can be temporally greater than SCHED_LOAD_SCALE after the migration of a task on this CPU and until the metrics are stabilized. The utilization_load_avg is in the range [0..SCHED_LOAD_SCALE] to reflect the running load on the CPU whereas the available capacity for the CFS task is in the range [0..cpu_capacity_orig]. In order to test if a CPU is fully utilized by CFS tasks, we have to scale the utilization in the cpu_capacity_orig range of the CPU to get the usage of the latter. The usage can then be compared with the available capacity (ie cpu_capacity) to deduct the usage level of a CPU. The frequency scaling invariance of the usage is not taken into account in this patch, it will be solved in another patch which will deal with frequency scaling invariance on the utilization_load_avg. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Morten.Rasmussen@arm.com Cc: dietmar.eggemann@arm.com Cc: efault@gmx.de Cc: kamalesh@linux.vnet.ibm.com Cc: linaro-kernel@lists.linaro.org Cc: nicolas.pitre@linaro.org Cc: preeti@linux.vnet.ibm.com Cc: riel@redhat.com Link: http://lkml.kernel.org/r/1425455327-13508-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-03-04 07:48:47 +00:00
*/
sched/fair: Use util_est in LB and WU paths When the scheduler looks at the CPU utilization, the current PELT value for a CPU is returned straight away. In certain scenarios this can have undesired side effects on task placement. For example, since the task utilization is decayed at wakeup time, when a long sleeping big task is enqueued it does not add immediately a significant contribution to the target CPU. As a result we generate a race condition where other tasks can be placed on the same CPU while it is still considered relatively empty. In order to reduce this kind of race conditions, this patch introduces the required support to integrate the usage of the CPU's estimated utilization in the wakeup path, via cpu_util_wake(), as well as in the load-balance path, via cpu_util() which is used by update_sg_lb_stats(). The estimated utilization of a CPU is defined to be the maximum between its PELT's utilization and the sum of the estimated utilization (at previous dequeue time) of all the tasks currently RUNNABLE on that CPU. This allows to properly represent the spare capacity of a CPU which, for example, has just got a big task running since a long sleep period. Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Joel Fernandes <joelaf@google.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rafael J . Wysocki <rafael.j.wysocki@intel.com> Cc: Steve Muckle <smuckle@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Todd Kjos <tkjos@android.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Link: http://lkml.kernel.org/r/20180309095245.11071-3-patrick.bellasi@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-03-09 09:52:43 +00:00
static inline unsigned long cpu_util(int cpu)
sched: Calculate CPU's usage statistic and put it into struct sg_lb_stats::group_usage Monitor the usage level of each group of each sched_domain level. The usage is the portion of cpu_capacity_orig that is currently used on a CPU or group of CPUs. We use the utilization_load_avg to evaluate the usage level of each group. The utilization_load_avg only takes into account the running time of the CFS tasks on a CPU with a maximum value of SCHED_LOAD_SCALE when the CPU is fully utilized. Nevertheless, we must cap utilization_load_avg which can be temporally greater than SCHED_LOAD_SCALE after the migration of a task on this CPU and until the metrics are stabilized. The utilization_load_avg is in the range [0..SCHED_LOAD_SCALE] to reflect the running load on the CPU whereas the available capacity for the CFS task is in the range [0..cpu_capacity_orig]. In order to test if a CPU is fully utilized by CFS tasks, we have to scale the utilization in the cpu_capacity_orig range of the CPU to get the usage of the latter. The usage can then be compared with the available capacity (ie cpu_capacity) to deduct the usage level of a CPU. The frequency scaling invariance of the usage is not taken into account in this patch, it will be solved in another patch which will deal with frequency scaling invariance on the utilization_load_avg. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Morten.Rasmussen@arm.com Cc: dietmar.eggemann@arm.com Cc: efault@gmx.de Cc: kamalesh@linux.vnet.ibm.com Cc: linaro-kernel@lists.linaro.org Cc: nicolas.pitre@linaro.org Cc: preeti@linux.vnet.ibm.com Cc: riel@redhat.com Link: http://lkml.kernel.org/r/1425455327-13508-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-03-04 07:48:47 +00:00
{
sched/fair: Use util_est in LB and WU paths When the scheduler looks at the CPU utilization, the current PELT value for a CPU is returned straight away. In certain scenarios this can have undesired side effects on task placement. For example, since the task utilization is decayed at wakeup time, when a long sleeping big task is enqueued it does not add immediately a significant contribution to the target CPU. As a result we generate a race condition where other tasks can be placed on the same CPU while it is still considered relatively empty. In order to reduce this kind of race conditions, this patch introduces the required support to integrate the usage of the CPU's estimated utilization in the wakeup path, via cpu_util_wake(), as well as in the load-balance path, via cpu_util() which is used by update_sg_lb_stats(). The estimated utilization of a CPU is defined to be the maximum between its PELT's utilization and the sum of the estimated utilization (at previous dequeue time) of all the tasks currently RUNNABLE on that CPU. This allows to properly represent the spare capacity of a CPU which, for example, has just got a big task running since a long sleep period. Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Joel Fernandes <joelaf@google.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rafael J . Wysocki <rafael.j.wysocki@intel.com> Cc: Steve Muckle <smuckle@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Todd Kjos <tkjos@android.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Link: http://lkml.kernel.org/r/20180309095245.11071-3-patrick.bellasi@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-03-09 09:52:43 +00:00
struct cfs_rq *cfs_rq;
unsigned int util;
cfs_rq = &cpu_rq(cpu)->cfs;
util = READ_ONCE(cfs_rq->avg.util_avg);
if (sched_feat(UTIL_EST))
util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued));
sched: Calculate CPU's usage statistic and put it into struct sg_lb_stats::group_usage Monitor the usage level of each group of each sched_domain level. The usage is the portion of cpu_capacity_orig that is currently used on a CPU or group of CPUs. We use the utilization_load_avg to evaluate the usage level of each group. The utilization_load_avg only takes into account the running time of the CFS tasks on a CPU with a maximum value of SCHED_LOAD_SCALE when the CPU is fully utilized. Nevertheless, we must cap utilization_load_avg which can be temporally greater than SCHED_LOAD_SCALE after the migration of a task on this CPU and until the metrics are stabilized. The utilization_load_avg is in the range [0..SCHED_LOAD_SCALE] to reflect the running load on the CPU whereas the available capacity for the CFS task is in the range [0..cpu_capacity_orig]. In order to test if a CPU is fully utilized by CFS tasks, we have to scale the utilization in the cpu_capacity_orig range of the CPU to get the usage of the latter. The usage can then be compared with the available capacity (ie cpu_capacity) to deduct the usage level of a CPU. The frequency scaling invariance of the usage is not taken into account in this patch, it will be solved in another patch which will deal with frequency scaling invariance on the utilization_load_avg. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Morten.Rasmussen@arm.com Cc: dietmar.eggemann@arm.com Cc: efault@gmx.de Cc: kamalesh@linux.vnet.ibm.com Cc: linaro-kernel@lists.linaro.org Cc: nicolas.pitre@linaro.org Cc: preeti@linux.vnet.ibm.com Cc: riel@redhat.com Link: http://lkml.kernel.org/r/1425455327-13508-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-03-04 07:48:47 +00:00
sched/fair: Use util_est in LB and WU paths When the scheduler looks at the CPU utilization, the current PELT value for a CPU is returned straight away. In certain scenarios this can have undesired side effects on task placement. For example, since the task utilization is decayed at wakeup time, when a long sleeping big task is enqueued it does not add immediately a significant contribution to the target CPU. As a result we generate a race condition where other tasks can be placed on the same CPU while it is still considered relatively empty. In order to reduce this kind of race conditions, this patch introduces the required support to integrate the usage of the CPU's estimated utilization in the wakeup path, via cpu_util_wake(), as well as in the load-balance path, via cpu_util() which is used by update_sg_lb_stats(). The estimated utilization of a CPU is defined to be the maximum between its PELT's utilization and the sum of the estimated utilization (at previous dequeue time) of all the tasks currently RUNNABLE on that CPU. This allows to properly represent the spare capacity of a CPU which, for example, has just got a big task running since a long sleep period. Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Joel Fernandes <joelaf@google.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rafael J . Wysocki <rafael.j.wysocki@intel.com> Cc: Steve Muckle <smuckle@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Todd Kjos <tkjos@android.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Link: http://lkml.kernel.org/r/20180309095245.11071-3-patrick.bellasi@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-03-09 09:52:43 +00:00
return min_t(unsigned long, util, capacity_orig_of(cpu));
sched: Calculate CPU's usage statistic and put it into struct sg_lb_stats::group_usage Monitor the usage level of each group of each sched_domain level. The usage is the portion of cpu_capacity_orig that is currently used on a CPU or group of CPUs. We use the utilization_load_avg to evaluate the usage level of each group. The utilization_load_avg only takes into account the running time of the CFS tasks on a CPU with a maximum value of SCHED_LOAD_SCALE when the CPU is fully utilized. Nevertheless, we must cap utilization_load_avg which can be temporally greater than SCHED_LOAD_SCALE after the migration of a task on this CPU and until the metrics are stabilized. The utilization_load_avg is in the range [0..SCHED_LOAD_SCALE] to reflect the running load on the CPU whereas the available capacity for the CFS task is in the range [0..cpu_capacity_orig]. In order to test if a CPU is fully utilized by CFS tasks, we have to scale the utilization in the cpu_capacity_orig range of the CPU to get the usage of the latter. The usage can then be compared with the available capacity (ie cpu_capacity) to deduct the usage level of a CPU. The frequency scaling invariance of the usage is not taken into account in this patch, it will be solved in another patch which will deal with frequency scaling invariance on the utilization_load_avg. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Morten.Rasmussen@arm.com Cc: dietmar.eggemann@arm.com Cc: efault@gmx.de Cc: kamalesh@linux.vnet.ibm.com Cc: linaro-kernel@lists.linaro.org Cc: nicolas.pitre@linaro.org Cc: preeti@linux.vnet.ibm.com Cc: riel@redhat.com Link: http://lkml.kernel.org/r/1425455327-13508-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-03-04 07:48:47 +00:00
}
sched/fair: Compute task/cpu utilization at wake-up correctly At task wake-up load-tracking isn't updated until the task is enqueued. The task's own view of its utilization contribution may therefore not be aligned with its contribution to the cfs_rq load-tracking which may have been updated in the meantime. Basically, the task's own utilization hasn't yet accounted for the sleep decay, while the cfs_rq may have (partially). Estimating the cfs_rq utilization in case the task is migrated at wake-up as task_rq(p)->cfs.avg.util_avg - p->se.avg.util_avg is therefore incorrect as the two load-tracking signals aren't time synchronized (different last update). To solve this problem, this patch synchronizes the task utilization with its previous rq before the task utilization is used in the wake-up path. Currently the update/synchronization is done _after_ the task has been placed by select_task_rq_fair(). The synchronization is done without having to take the rq lock using the existing mechanism used in remove_entity_load_avg(). Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: freedom.tan@mediatek.com Cc: keita.kobayashi.ym@renesas.com Cc: mgalbraith@suse.de Cc: sgurrappadi@nvidia.com Cc: vincent.guittot@linaro.org Cc: yuyang.du@intel.com Link: http://lkml.kernel.org/r/1476452472-24740-2-git-send-email-morten.rasmussen@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-10-14 13:41:07 +00:00
/*
sched/fair: Fix cpu_util_wake() for 'execl' type workloads A ~10% regression has been reported for UnixBench's execl throughput test by Aaron Lu and Ye Xiaolong: https://lkml.org/lkml/2018/10/30/765 That test is pretty simple, it does a "recursive" execve() syscall on the same binary. Starting from the syscall, this sequence is possible: do_execve() do_execveat_common() __do_execve_file() sched_exec() select_task_rq_fair() <==| Task already enqueued find_idlest_cpu() find_idlest_group() capacity_spare_wake() <==| Functions not called from cpu_util_wake() | the wakeup path which means we can end up calling cpu_util_wake() not only from the "wakeup path", as its name would suggest. Indeed, the task doing an execve() syscall is already enqueued on the CPU we want to get the cpu_util_wake() for. The estimated utilization for a CPU computed in cpu_util_wake() was written under the assumption that function can be called only from the wakeup path. If instead the task is already enqueued, we end up with a utilization which does not remove the current task's contribution from the estimated utilization of the CPU. This will wrongly assume a reduced spare capacity on the current CPU and increase the chances to migrate the task on execve. The regression is tracked down to: commit d519329f72a6 ("sched/fair: Update util_est only on util_avg updates") because in that patch we turn on by default the UTIL_EST sched feature. However, the real issue is introduced by: commit f9be3e5961c5 ("sched/fair: Use util_est in LB and WU paths") Let's fix this by ensuring to always discount the task estimated utilization from the CPU's estimated utilization when the task is also the current one. The same benchmark of the bug report, executed on a dual socket 40 CPUs Intel(R) Xeon(R) CPU E5-2690 v2 @ 3.00GHz machine, reports these "Execl Throughput" figures (higher the better): mainline : 48136.5 lps mainline+fix : 55376.5 lps which correspond to a 15% speedup. Moreover, since {cpu_util,capacity_spare}_wake() are not really only used from the wakeup path, let's remove this ambiguity by using a better matching name: {cpu_util,capacity_spare}_without(). Since we are at that, let's also improve the existing documentation. Reported-by: Aaron Lu <aaron.lu@intel.com> Reported-by: Ye Xiaolong <xiaolong.ye@intel.com> Tested-by: Aaron Lu <aaron.lu@intel.com> Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Quentin Perret <quentin.perret@arm.com> Cc: Steve Muckle <smuckle@google.com> Cc: Suren Baghdasaryan <surenb@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Todd Kjos <tkjos@google.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Fixes: f9be3e5961c5 (sched/fair: Use util_est in LB and WU paths) Link: https://lore.kernel.org/lkml/20181025093100.GB13236@e110439-lin/ Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-11-05 14:53:58 +00:00
* cpu_util_without: compute cpu utilization without any contributions from *p
* @cpu: the CPU which utilization is requested
* @p: the task which utilization should be discounted
*
* The utilization of a CPU is defined by the utilization of tasks currently
* enqueued on that CPU as well as tasks which are currently sleeping after an
* execution on that CPU.
*
* This method returns the utilization of the specified CPU by discounting the
* utilization of the specified task, whenever the task is currently
* contributing to the CPU utilization.
sched/fair: Compute task/cpu utilization at wake-up correctly At task wake-up load-tracking isn't updated until the task is enqueued. The task's own view of its utilization contribution may therefore not be aligned with its contribution to the cfs_rq load-tracking which may have been updated in the meantime. Basically, the task's own utilization hasn't yet accounted for the sleep decay, while the cfs_rq may have (partially). Estimating the cfs_rq utilization in case the task is migrated at wake-up as task_rq(p)->cfs.avg.util_avg - p->se.avg.util_avg is therefore incorrect as the two load-tracking signals aren't time synchronized (different last update). To solve this problem, this patch synchronizes the task utilization with its previous rq before the task utilization is used in the wake-up path. Currently the update/synchronization is done _after_ the task has been placed by select_task_rq_fair(). The synchronization is done without having to take the rq lock using the existing mechanism used in remove_entity_load_avg(). Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: freedom.tan@mediatek.com Cc: keita.kobayashi.ym@renesas.com Cc: mgalbraith@suse.de Cc: sgurrappadi@nvidia.com Cc: vincent.guittot@linaro.org Cc: yuyang.du@intel.com Link: http://lkml.kernel.org/r/1476452472-24740-2-git-send-email-morten.rasmussen@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-10-14 13:41:07 +00:00
*/
sched/fair: Fix cpu_util_wake() for 'execl' type workloads A ~10% regression has been reported for UnixBench's execl throughput test by Aaron Lu and Ye Xiaolong: https://lkml.org/lkml/2018/10/30/765 That test is pretty simple, it does a "recursive" execve() syscall on the same binary. Starting from the syscall, this sequence is possible: do_execve() do_execveat_common() __do_execve_file() sched_exec() select_task_rq_fair() <==| Task already enqueued find_idlest_cpu() find_idlest_group() capacity_spare_wake() <==| Functions not called from cpu_util_wake() | the wakeup path which means we can end up calling cpu_util_wake() not only from the "wakeup path", as its name would suggest. Indeed, the task doing an execve() syscall is already enqueued on the CPU we want to get the cpu_util_wake() for. The estimated utilization for a CPU computed in cpu_util_wake() was written under the assumption that function can be called only from the wakeup path. If instead the task is already enqueued, we end up with a utilization which does not remove the current task's contribution from the estimated utilization of the CPU. This will wrongly assume a reduced spare capacity on the current CPU and increase the chances to migrate the task on execve. The regression is tracked down to: commit d519329f72a6 ("sched/fair: Update util_est only on util_avg updates") because in that patch we turn on by default the UTIL_EST sched feature. However, the real issue is introduced by: commit f9be3e5961c5 ("sched/fair: Use util_est in LB and WU paths") Let's fix this by ensuring to always discount the task estimated utilization from the CPU's estimated utilization when the task is also the current one. The same benchmark of the bug report, executed on a dual socket 40 CPUs Intel(R) Xeon(R) CPU E5-2690 v2 @ 3.00GHz machine, reports these "Execl Throughput" figures (higher the better): mainline : 48136.5 lps mainline+fix : 55376.5 lps which correspond to a 15% speedup. Moreover, since {cpu_util,capacity_spare}_wake() are not really only used from the wakeup path, let's remove this ambiguity by using a better matching name: {cpu_util,capacity_spare}_without(). Since we are at that, let's also improve the existing documentation. Reported-by: Aaron Lu <aaron.lu@intel.com> Reported-by: Ye Xiaolong <xiaolong.ye@intel.com> Tested-by: Aaron Lu <aaron.lu@intel.com> Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Quentin Perret <quentin.perret@arm.com> Cc: Steve Muckle <smuckle@google.com> Cc: Suren Baghdasaryan <surenb@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Todd Kjos <tkjos@google.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Fixes: f9be3e5961c5 (sched/fair: Use util_est in LB and WU paths) Link: https://lore.kernel.org/lkml/20181025093100.GB13236@e110439-lin/ Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-11-05 14:53:58 +00:00
static unsigned long cpu_util_without(int cpu, struct task_struct *p)
sched/fair: Compute task/cpu utilization at wake-up correctly At task wake-up load-tracking isn't updated until the task is enqueued. The task's own view of its utilization contribution may therefore not be aligned with its contribution to the cfs_rq load-tracking which may have been updated in the meantime. Basically, the task's own utilization hasn't yet accounted for the sleep decay, while the cfs_rq may have (partially). Estimating the cfs_rq utilization in case the task is migrated at wake-up as task_rq(p)->cfs.avg.util_avg - p->se.avg.util_avg is therefore incorrect as the two load-tracking signals aren't time synchronized (different last update). To solve this problem, this patch synchronizes the task utilization with its previous rq before the task utilization is used in the wake-up path. Currently the update/synchronization is done _after_ the task has been placed by select_task_rq_fair(). The synchronization is done without having to take the rq lock using the existing mechanism used in remove_entity_load_avg(). Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: freedom.tan@mediatek.com Cc: keita.kobayashi.ym@renesas.com Cc: mgalbraith@suse.de Cc: sgurrappadi@nvidia.com Cc: vincent.guittot@linaro.org Cc: yuyang.du@intel.com Link: http://lkml.kernel.org/r/1476452472-24740-2-git-send-email-morten.rasmussen@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-10-14 13:41:07 +00:00
{
sched/fair: Use util_est in LB and WU paths When the scheduler looks at the CPU utilization, the current PELT value for a CPU is returned straight away. In certain scenarios this can have undesired side effects on task placement. For example, since the task utilization is decayed at wakeup time, when a long sleeping big task is enqueued it does not add immediately a significant contribution to the target CPU. As a result we generate a race condition where other tasks can be placed on the same CPU while it is still considered relatively empty. In order to reduce this kind of race conditions, this patch introduces the required support to integrate the usage of the CPU's estimated utilization in the wakeup path, via cpu_util_wake(), as well as in the load-balance path, via cpu_util() which is used by update_sg_lb_stats(). The estimated utilization of a CPU is defined to be the maximum between its PELT's utilization and the sum of the estimated utilization (at previous dequeue time) of all the tasks currently RUNNABLE on that CPU. This allows to properly represent the spare capacity of a CPU which, for example, has just got a big task running since a long sleep period. Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Joel Fernandes <joelaf@google.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rafael J . Wysocki <rafael.j.wysocki@intel.com> Cc: Steve Muckle <smuckle@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Todd Kjos <tkjos@android.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Link: http://lkml.kernel.org/r/20180309095245.11071-3-patrick.bellasi@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-03-09 09:52:43 +00:00
struct cfs_rq *cfs_rq;
unsigned int util;
sched/fair: Compute task/cpu utilization at wake-up correctly At task wake-up load-tracking isn't updated until the task is enqueued. The task's own view of its utilization contribution may therefore not be aligned with its contribution to the cfs_rq load-tracking which may have been updated in the meantime. Basically, the task's own utilization hasn't yet accounted for the sleep decay, while the cfs_rq may have (partially). Estimating the cfs_rq utilization in case the task is migrated at wake-up as task_rq(p)->cfs.avg.util_avg - p->se.avg.util_avg is therefore incorrect as the two load-tracking signals aren't time synchronized (different last update). To solve this problem, this patch synchronizes the task utilization with its previous rq before the task utilization is used in the wake-up path. Currently the update/synchronization is done _after_ the task has been placed by select_task_rq_fair(). The synchronization is done without having to take the rq lock using the existing mechanism used in remove_entity_load_avg(). Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: freedom.tan@mediatek.com Cc: keita.kobayashi.ym@renesas.com Cc: mgalbraith@suse.de Cc: sgurrappadi@nvidia.com Cc: vincent.guittot@linaro.org Cc: yuyang.du@intel.com Link: http://lkml.kernel.org/r/1476452472-24740-2-git-send-email-morten.rasmussen@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-10-14 13:41:07 +00:00
/* Task has no contribution or is new */
sched/fair: Use util_est in LB and WU paths When the scheduler looks at the CPU utilization, the current PELT value for a CPU is returned straight away. In certain scenarios this can have undesired side effects on task placement. For example, since the task utilization is decayed at wakeup time, when a long sleeping big task is enqueued it does not add immediately a significant contribution to the target CPU. As a result we generate a race condition where other tasks can be placed on the same CPU while it is still considered relatively empty. In order to reduce this kind of race conditions, this patch introduces the required support to integrate the usage of the CPU's estimated utilization in the wakeup path, via cpu_util_wake(), as well as in the load-balance path, via cpu_util() which is used by update_sg_lb_stats(). The estimated utilization of a CPU is defined to be the maximum between its PELT's utilization and the sum of the estimated utilization (at previous dequeue time) of all the tasks currently RUNNABLE on that CPU. This allows to properly represent the spare capacity of a CPU which, for example, has just got a big task running since a long sleep period. Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Joel Fernandes <joelaf@google.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rafael J . Wysocki <rafael.j.wysocki@intel.com> Cc: Steve Muckle <smuckle@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Todd Kjos <tkjos@android.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Link: http://lkml.kernel.org/r/20180309095245.11071-3-patrick.bellasi@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-03-09 09:52:43 +00:00
if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
sched/fair: Compute task/cpu utilization at wake-up correctly At task wake-up load-tracking isn't updated until the task is enqueued. The task's own view of its utilization contribution may therefore not be aligned with its contribution to the cfs_rq load-tracking which may have been updated in the meantime. Basically, the task's own utilization hasn't yet accounted for the sleep decay, while the cfs_rq may have (partially). Estimating the cfs_rq utilization in case the task is migrated at wake-up as task_rq(p)->cfs.avg.util_avg - p->se.avg.util_avg is therefore incorrect as the two load-tracking signals aren't time synchronized (different last update). To solve this problem, this patch synchronizes the task utilization with its previous rq before the task utilization is used in the wake-up path. Currently the update/synchronization is done _after_ the task has been placed by select_task_rq_fair(). The synchronization is done without having to take the rq lock using the existing mechanism used in remove_entity_load_avg(). Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: freedom.tan@mediatek.com Cc: keita.kobayashi.ym@renesas.com Cc: mgalbraith@suse.de Cc: sgurrappadi@nvidia.com Cc: vincent.guittot@linaro.org Cc: yuyang.du@intel.com Link: http://lkml.kernel.org/r/1476452472-24740-2-git-send-email-morten.rasmussen@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-10-14 13:41:07 +00:00
return cpu_util(cpu);
sched/fair: Use util_est in LB and WU paths When the scheduler looks at the CPU utilization, the current PELT value for a CPU is returned straight away. In certain scenarios this can have undesired side effects on task placement. For example, since the task utilization is decayed at wakeup time, when a long sleeping big task is enqueued it does not add immediately a significant contribution to the target CPU. As a result we generate a race condition where other tasks can be placed on the same CPU while it is still considered relatively empty. In order to reduce this kind of race conditions, this patch introduces the required support to integrate the usage of the CPU's estimated utilization in the wakeup path, via cpu_util_wake(), as well as in the load-balance path, via cpu_util() which is used by update_sg_lb_stats(). The estimated utilization of a CPU is defined to be the maximum between its PELT's utilization and the sum of the estimated utilization (at previous dequeue time) of all the tasks currently RUNNABLE on that CPU. This allows to properly represent the spare capacity of a CPU which, for example, has just got a big task running since a long sleep period. Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Joel Fernandes <joelaf@google.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rafael J . Wysocki <rafael.j.wysocki@intel.com> Cc: Steve Muckle <smuckle@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Todd Kjos <tkjos@android.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Link: http://lkml.kernel.org/r/20180309095245.11071-3-patrick.bellasi@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-03-09 09:52:43 +00:00
cfs_rq = &cpu_rq(cpu)->cfs;
util = READ_ONCE(cfs_rq->avg.util_avg);
sched/fair: Fix cpu_util_wake() for 'execl' type workloads A ~10% regression has been reported for UnixBench's execl throughput test by Aaron Lu and Ye Xiaolong: https://lkml.org/lkml/2018/10/30/765 That test is pretty simple, it does a "recursive" execve() syscall on the same binary. Starting from the syscall, this sequence is possible: do_execve() do_execveat_common() __do_execve_file() sched_exec() select_task_rq_fair() <==| Task already enqueued find_idlest_cpu() find_idlest_group() capacity_spare_wake() <==| Functions not called from cpu_util_wake() | the wakeup path which means we can end up calling cpu_util_wake() not only from the "wakeup path", as its name would suggest. Indeed, the task doing an execve() syscall is already enqueued on the CPU we want to get the cpu_util_wake() for. The estimated utilization for a CPU computed in cpu_util_wake() was written under the assumption that function can be called only from the wakeup path. If instead the task is already enqueued, we end up with a utilization which does not remove the current task's contribution from the estimated utilization of the CPU. This will wrongly assume a reduced spare capacity on the current CPU and increase the chances to migrate the task on execve. The regression is tracked down to: commit d519329f72a6 ("sched/fair: Update util_est only on util_avg updates") because in that patch we turn on by default the UTIL_EST sched feature. However, the real issue is introduced by: commit f9be3e5961c5 ("sched/fair: Use util_est in LB and WU paths") Let's fix this by ensuring to always discount the task estimated utilization from the CPU's estimated utilization when the task is also the current one. The same benchmark of the bug report, executed on a dual socket 40 CPUs Intel(R) Xeon(R) CPU E5-2690 v2 @ 3.00GHz machine, reports these "Execl Throughput" figures (higher the better): mainline : 48136.5 lps mainline+fix : 55376.5 lps which correspond to a 15% speedup. Moreover, since {cpu_util,capacity_spare}_wake() are not really only used from the wakeup path, let's remove this ambiguity by using a better matching name: {cpu_util,capacity_spare}_without(). Since we are at that, let's also improve the existing documentation. Reported-by: Aaron Lu <aaron.lu@intel.com> Reported-by: Ye Xiaolong <xiaolong.ye@intel.com> Tested-by: Aaron Lu <aaron.lu@intel.com> Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Quentin Perret <quentin.perret@arm.com> Cc: Steve Muckle <smuckle@google.com> Cc: Suren Baghdasaryan <surenb@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Todd Kjos <tkjos@google.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Fixes: f9be3e5961c5 (sched/fair: Use util_est in LB and WU paths) Link: https://lore.kernel.org/lkml/20181025093100.GB13236@e110439-lin/ Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-11-05 14:53:58 +00:00
/* Discount task's util from CPU's util */
lsub_positive(&util, task_util(p));
sched/fair: Compute task/cpu utilization at wake-up correctly At task wake-up load-tracking isn't updated until the task is enqueued. The task's own view of its utilization contribution may therefore not be aligned with its contribution to the cfs_rq load-tracking which may have been updated in the meantime. Basically, the task's own utilization hasn't yet accounted for the sleep decay, while the cfs_rq may have (partially). Estimating the cfs_rq utilization in case the task is migrated at wake-up as task_rq(p)->cfs.avg.util_avg - p->se.avg.util_avg is therefore incorrect as the two load-tracking signals aren't time synchronized (different last update). To solve this problem, this patch synchronizes the task utilization with its previous rq before the task utilization is used in the wake-up path. Currently the update/synchronization is done _after_ the task has been placed by select_task_rq_fair(). The synchronization is done without having to take the rq lock using the existing mechanism used in remove_entity_load_avg(). Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: freedom.tan@mediatek.com Cc: keita.kobayashi.ym@renesas.com Cc: mgalbraith@suse.de Cc: sgurrappadi@nvidia.com Cc: vincent.guittot@linaro.org Cc: yuyang.du@intel.com Link: http://lkml.kernel.org/r/1476452472-24740-2-git-send-email-morten.rasmussen@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-10-14 13:41:07 +00:00
sched/fair: Use util_est in LB and WU paths When the scheduler looks at the CPU utilization, the current PELT value for a CPU is returned straight away. In certain scenarios this can have undesired side effects on task placement. For example, since the task utilization is decayed at wakeup time, when a long sleeping big task is enqueued it does not add immediately a significant contribution to the target CPU. As a result we generate a race condition where other tasks can be placed on the same CPU while it is still considered relatively empty. In order to reduce this kind of race conditions, this patch introduces the required support to integrate the usage of the CPU's estimated utilization in the wakeup path, via cpu_util_wake(), as well as in the load-balance path, via cpu_util() which is used by update_sg_lb_stats(). The estimated utilization of a CPU is defined to be the maximum between its PELT's utilization and the sum of the estimated utilization (at previous dequeue time) of all the tasks currently RUNNABLE on that CPU. This allows to properly represent the spare capacity of a CPU which, for example, has just got a big task running since a long sleep period. Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Joel Fernandes <joelaf@google.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rafael J . Wysocki <rafael.j.wysocki@intel.com> Cc: Steve Muckle <smuckle@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Todd Kjos <tkjos@android.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Link: http://lkml.kernel.org/r/20180309095245.11071-3-patrick.bellasi@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-03-09 09:52:43 +00:00
/*
* Covered cases:
*
* a) if *p is the only task sleeping on this CPU, then:
* cpu_util (== task_util) > util_est (== 0)
* and thus we return:
sched/fair: Fix cpu_util_wake() for 'execl' type workloads A ~10% regression has been reported for UnixBench's execl throughput test by Aaron Lu and Ye Xiaolong: https://lkml.org/lkml/2018/10/30/765 That test is pretty simple, it does a "recursive" execve() syscall on the same binary. Starting from the syscall, this sequence is possible: do_execve() do_execveat_common() __do_execve_file() sched_exec() select_task_rq_fair() <==| Task already enqueued find_idlest_cpu() find_idlest_group() capacity_spare_wake() <==| Functions not called from cpu_util_wake() | the wakeup path which means we can end up calling cpu_util_wake() not only from the "wakeup path", as its name would suggest. Indeed, the task doing an execve() syscall is already enqueued on the CPU we want to get the cpu_util_wake() for. The estimated utilization for a CPU computed in cpu_util_wake() was written under the assumption that function can be called only from the wakeup path. If instead the task is already enqueued, we end up with a utilization which does not remove the current task's contribution from the estimated utilization of the CPU. This will wrongly assume a reduced spare capacity on the current CPU and increase the chances to migrate the task on execve. The regression is tracked down to: commit d519329f72a6 ("sched/fair: Update util_est only on util_avg updates") because in that patch we turn on by default the UTIL_EST sched feature. However, the real issue is introduced by: commit f9be3e5961c5 ("sched/fair: Use util_est in LB and WU paths") Let's fix this by ensuring to always discount the task estimated utilization from the CPU's estimated utilization when the task is also the current one. The same benchmark of the bug report, executed on a dual socket 40 CPUs Intel(R) Xeon(R) CPU E5-2690 v2 @ 3.00GHz machine, reports these "Execl Throughput" figures (higher the better): mainline : 48136.5 lps mainline+fix : 55376.5 lps which correspond to a 15% speedup. Moreover, since {cpu_util,capacity_spare}_wake() are not really only used from the wakeup path, let's remove this ambiguity by using a better matching name: {cpu_util,capacity_spare}_without(). Since we are at that, let's also improve the existing documentation. Reported-by: Aaron Lu <aaron.lu@intel.com> Reported-by: Ye Xiaolong <xiaolong.ye@intel.com> Tested-by: Aaron Lu <aaron.lu@intel.com> Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Quentin Perret <quentin.perret@arm.com> Cc: Steve Muckle <smuckle@google.com> Cc: Suren Baghdasaryan <surenb@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Todd Kjos <tkjos@google.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Fixes: f9be3e5961c5 (sched/fair: Use util_est in LB and WU paths) Link: https://lore.kernel.org/lkml/20181025093100.GB13236@e110439-lin/ Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-11-05 14:53:58 +00:00
* cpu_util_without = (cpu_util - task_util) = 0
sched/fair: Use util_est in LB and WU paths When the scheduler looks at the CPU utilization, the current PELT value for a CPU is returned straight away. In certain scenarios this can have undesired side effects on task placement. For example, since the task utilization is decayed at wakeup time, when a long sleeping big task is enqueued it does not add immediately a significant contribution to the target CPU. As a result we generate a race condition where other tasks can be placed on the same CPU while it is still considered relatively empty. In order to reduce this kind of race conditions, this patch introduces the required support to integrate the usage of the CPU's estimated utilization in the wakeup path, via cpu_util_wake(), as well as in the load-balance path, via cpu_util() which is used by update_sg_lb_stats(). The estimated utilization of a CPU is defined to be the maximum between its PELT's utilization and the sum of the estimated utilization (at previous dequeue time) of all the tasks currently RUNNABLE on that CPU. This allows to properly represent the spare capacity of a CPU which, for example, has just got a big task running since a long sleep period. Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Joel Fernandes <joelaf@google.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rafael J . Wysocki <rafael.j.wysocki@intel.com> Cc: Steve Muckle <smuckle@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Todd Kjos <tkjos@android.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Link: http://lkml.kernel.org/r/20180309095245.11071-3-patrick.bellasi@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-03-09 09:52:43 +00:00
*
* b) if other tasks are SLEEPING on this CPU, which is now exiting
* IDLE, then:
* cpu_util >= task_util
* cpu_util > util_est (== 0)
* and thus we discount *p's blocked utilization to return:
sched/fair: Fix cpu_util_wake() for 'execl' type workloads A ~10% regression has been reported for UnixBench's execl throughput test by Aaron Lu and Ye Xiaolong: https://lkml.org/lkml/2018/10/30/765 That test is pretty simple, it does a "recursive" execve() syscall on the same binary. Starting from the syscall, this sequence is possible: do_execve() do_execveat_common() __do_execve_file() sched_exec() select_task_rq_fair() <==| Task already enqueued find_idlest_cpu() find_idlest_group() capacity_spare_wake() <==| Functions not called from cpu_util_wake() | the wakeup path which means we can end up calling cpu_util_wake() not only from the "wakeup path", as its name would suggest. Indeed, the task doing an execve() syscall is already enqueued on the CPU we want to get the cpu_util_wake() for. The estimated utilization for a CPU computed in cpu_util_wake() was written under the assumption that function can be called only from the wakeup path. If instead the task is already enqueued, we end up with a utilization which does not remove the current task's contribution from the estimated utilization of the CPU. This will wrongly assume a reduced spare capacity on the current CPU and increase the chances to migrate the task on execve. The regression is tracked down to: commit d519329f72a6 ("sched/fair: Update util_est only on util_avg updates") because in that patch we turn on by default the UTIL_EST sched feature. However, the real issue is introduced by: commit f9be3e5961c5 ("sched/fair: Use util_est in LB and WU paths") Let's fix this by ensuring to always discount the task estimated utilization from the CPU's estimated utilization when the task is also the current one. The same benchmark of the bug report, executed on a dual socket 40 CPUs Intel(R) Xeon(R) CPU E5-2690 v2 @ 3.00GHz machine, reports these "Execl Throughput" figures (higher the better): mainline : 48136.5 lps mainline+fix : 55376.5 lps which correspond to a 15% speedup. Moreover, since {cpu_util,capacity_spare}_wake() are not really only used from the wakeup path, let's remove this ambiguity by using a better matching name: {cpu_util,capacity_spare}_without(). Since we are at that, let's also improve the existing documentation. Reported-by: Aaron Lu <aaron.lu@intel.com> Reported-by: Ye Xiaolong <xiaolong.ye@intel.com> Tested-by: Aaron Lu <aaron.lu@intel.com> Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Quentin Perret <quentin.perret@arm.com> Cc: Steve Muckle <smuckle@google.com> Cc: Suren Baghdasaryan <surenb@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Todd Kjos <tkjos@google.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Fixes: f9be3e5961c5 (sched/fair: Use util_est in LB and WU paths) Link: https://lore.kernel.org/lkml/20181025093100.GB13236@e110439-lin/ Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-11-05 14:53:58 +00:00
* cpu_util_without = (cpu_util - task_util) >= 0
sched/fair: Use util_est in LB and WU paths When the scheduler looks at the CPU utilization, the current PELT value for a CPU is returned straight away. In certain scenarios this can have undesired side effects on task placement. For example, since the task utilization is decayed at wakeup time, when a long sleeping big task is enqueued it does not add immediately a significant contribution to the target CPU. As a result we generate a race condition where other tasks can be placed on the same CPU while it is still considered relatively empty. In order to reduce this kind of race conditions, this patch introduces the required support to integrate the usage of the CPU's estimated utilization in the wakeup path, via cpu_util_wake(), as well as in the load-balance path, via cpu_util() which is used by update_sg_lb_stats(). The estimated utilization of a CPU is defined to be the maximum between its PELT's utilization and the sum of the estimated utilization (at previous dequeue time) of all the tasks currently RUNNABLE on that CPU. This allows to properly represent the spare capacity of a CPU which, for example, has just got a big task running since a long sleep period. Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Joel Fernandes <joelaf@google.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rafael J . Wysocki <rafael.j.wysocki@intel.com> Cc: Steve Muckle <smuckle@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Todd Kjos <tkjos@android.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Link: http://lkml.kernel.org/r/20180309095245.11071-3-patrick.bellasi@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-03-09 09:52:43 +00:00
*
* c) if other tasks are RUNNABLE on that CPU and
* util_est > cpu_util
* then we use util_est since it returns a more restrictive
* estimation of the spare capacity on that CPU, by just
* considering the expected utilization of tasks already
* runnable on that CPU.
*
* Cases a) and b) are covered by the above code, while case c) is
* covered by the following code when estimated utilization is
* enabled.
*/
sched/fair: Fix cpu_util_wake() for 'execl' type workloads A ~10% regression has been reported for UnixBench's execl throughput test by Aaron Lu and Ye Xiaolong: https://lkml.org/lkml/2018/10/30/765 That test is pretty simple, it does a "recursive" execve() syscall on the same binary. Starting from the syscall, this sequence is possible: do_execve() do_execveat_common() __do_execve_file() sched_exec() select_task_rq_fair() <==| Task already enqueued find_idlest_cpu() find_idlest_group() capacity_spare_wake() <==| Functions not called from cpu_util_wake() | the wakeup path which means we can end up calling cpu_util_wake() not only from the "wakeup path", as its name would suggest. Indeed, the task doing an execve() syscall is already enqueued on the CPU we want to get the cpu_util_wake() for. The estimated utilization for a CPU computed in cpu_util_wake() was written under the assumption that function can be called only from the wakeup path. If instead the task is already enqueued, we end up with a utilization which does not remove the current task's contribution from the estimated utilization of the CPU. This will wrongly assume a reduced spare capacity on the current CPU and increase the chances to migrate the task on execve. The regression is tracked down to: commit d519329f72a6 ("sched/fair: Update util_est only on util_avg updates") because in that patch we turn on by default the UTIL_EST sched feature. However, the real issue is introduced by: commit f9be3e5961c5 ("sched/fair: Use util_est in LB and WU paths") Let's fix this by ensuring to always discount the task estimated utilization from the CPU's estimated utilization when the task is also the current one. The same benchmark of the bug report, executed on a dual socket 40 CPUs Intel(R) Xeon(R) CPU E5-2690 v2 @ 3.00GHz machine, reports these "Execl Throughput" figures (higher the better): mainline : 48136.5 lps mainline+fix : 55376.5 lps which correspond to a 15% speedup. Moreover, since {cpu_util,capacity_spare}_wake() are not really only used from the wakeup path, let's remove this ambiguity by using a better matching name: {cpu_util,capacity_spare}_without(). Since we are at that, let's also improve the existing documentation. Reported-by: Aaron Lu <aaron.lu@intel.com> Reported-by: Ye Xiaolong <xiaolong.ye@intel.com> Tested-by: Aaron Lu <aaron.lu@intel.com> Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Quentin Perret <quentin.perret@arm.com> Cc: Steve Muckle <smuckle@google.com> Cc: Suren Baghdasaryan <surenb@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Todd Kjos <tkjos@google.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Fixes: f9be3e5961c5 (sched/fair: Use util_est in LB and WU paths) Link: https://lore.kernel.org/lkml/20181025093100.GB13236@e110439-lin/ Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-11-05 14:53:58 +00:00
if (sched_feat(UTIL_EST)) {
unsigned int estimated =
READ_ONCE(cfs_rq->avg.util_est.enqueued);
/*
* Despite the following checks we still have a small window
* for a possible race, when an execl's select_task_rq_fair()
* races with LB's detach_task():
*
* detach_task()
* p->on_rq = TASK_ON_RQ_MIGRATING;
* ---------------------------------- A
* deactivate_task() \
* dequeue_task() + RaceTime
* util_est_dequeue() /
* ---------------------------------- B
*
* The additional check on "current == p" it's required to
* properly fix the execl regression and it helps in further
* reducing the chances for the above race.
*/
if (unlikely(task_on_rq_queued(p) || current == p))
lsub_positive(&estimated, _task_util_est(p));
sched/fair: Fix cpu_util_wake() for 'execl' type workloads A ~10% regression has been reported for UnixBench's execl throughput test by Aaron Lu and Ye Xiaolong: https://lkml.org/lkml/2018/10/30/765 That test is pretty simple, it does a "recursive" execve() syscall on the same binary. Starting from the syscall, this sequence is possible: do_execve() do_execveat_common() __do_execve_file() sched_exec() select_task_rq_fair() <==| Task already enqueued find_idlest_cpu() find_idlest_group() capacity_spare_wake() <==| Functions not called from cpu_util_wake() | the wakeup path which means we can end up calling cpu_util_wake() not only from the "wakeup path", as its name would suggest. Indeed, the task doing an execve() syscall is already enqueued on the CPU we want to get the cpu_util_wake() for. The estimated utilization for a CPU computed in cpu_util_wake() was written under the assumption that function can be called only from the wakeup path. If instead the task is already enqueued, we end up with a utilization which does not remove the current task's contribution from the estimated utilization of the CPU. This will wrongly assume a reduced spare capacity on the current CPU and increase the chances to migrate the task on execve. The regression is tracked down to: commit d519329f72a6 ("sched/fair: Update util_est only on util_avg updates") because in that patch we turn on by default the UTIL_EST sched feature. However, the real issue is introduced by: commit f9be3e5961c5 ("sched/fair: Use util_est in LB and WU paths") Let's fix this by ensuring to always discount the task estimated utilization from the CPU's estimated utilization when the task is also the current one. The same benchmark of the bug report, executed on a dual socket 40 CPUs Intel(R) Xeon(R) CPU E5-2690 v2 @ 3.00GHz machine, reports these "Execl Throughput" figures (higher the better): mainline : 48136.5 lps mainline+fix : 55376.5 lps which correspond to a 15% speedup. Moreover, since {cpu_util,capacity_spare}_wake() are not really only used from the wakeup path, let's remove this ambiguity by using a better matching name: {cpu_util,capacity_spare}_without(). Since we are at that, let's also improve the existing documentation. Reported-by: Aaron Lu <aaron.lu@intel.com> Reported-by: Ye Xiaolong <xiaolong.ye@intel.com> Tested-by: Aaron Lu <aaron.lu@intel.com> Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Quentin Perret <quentin.perret@arm.com> Cc: Steve Muckle <smuckle@google.com> Cc: Suren Baghdasaryan <surenb@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Todd Kjos <tkjos@google.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Fixes: f9be3e5961c5 (sched/fair: Use util_est in LB and WU paths) Link: https://lore.kernel.org/lkml/20181025093100.GB13236@e110439-lin/ Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-11-05 14:53:58 +00:00
util = max(util, estimated);
}
sched/fair: Use util_est in LB and WU paths When the scheduler looks at the CPU utilization, the current PELT value for a CPU is returned straight away. In certain scenarios this can have undesired side effects on task placement. For example, since the task utilization is decayed at wakeup time, when a long sleeping big task is enqueued it does not add immediately a significant contribution to the target CPU. As a result we generate a race condition where other tasks can be placed on the same CPU while it is still considered relatively empty. In order to reduce this kind of race conditions, this patch introduces the required support to integrate the usage of the CPU's estimated utilization in the wakeup path, via cpu_util_wake(), as well as in the load-balance path, via cpu_util() which is used by update_sg_lb_stats(). The estimated utilization of a CPU is defined to be the maximum between its PELT's utilization and the sum of the estimated utilization (at previous dequeue time) of all the tasks currently RUNNABLE on that CPU. This allows to properly represent the spare capacity of a CPU which, for example, has just got a big task running since a long sleep period. Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Joel Fernandes <joelaf@google.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rafael J . Wysocki <rafael.j.wysocki@intel.com> Cc: Steve Muckle <smuckle@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Todd Kjos <tkjos@android.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Link: http://lkml.kernel.org/r/20180309095245.11071-3-patrick.bellasi@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-03-09 09:52:43 +00:00
/*
* Utilization (estimated) can exceed the CPU capacity, thus let's
* clamp to the maximum CPU capacity to ensure consistency with
* the cpu_util call.
*/
return min_t(unsigned long, util, capacity_orig_of(cpu));
sched/fair: Compute task/cpu utilization at wake-up correctly At task wake-up load-tracking isn't updated until the task is enqueued. The task's own view of its utilization contribution may therefore not be aligned with its contribution to the cfs_rq load-tracking which may have been updated in the meantime. Basically, the task's own utilization hasn't yet accounted for the sleep decay, while the cfs_rq may have (partially). Estimating the cfs_rq utilization in case the task is migrated at wake-up as task_rq(p)->cfs.avg.util_avg - p->se.avg.util_avg is therefore incorrect as the two load-tracking signals aren't time synchronized (different last update). To solve this problem, this patch synchronizes the task utilization with its previous rq before the task utilization is used in the wake-up path. Currently the update/synchronization is done _after_ the task has been placed by select_task_rq_fair(). The synchronization is done without having to take the rq lock using the existing mechanism used in remove_entity_load_avg(). Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: freedom.tan@mediatek.com Cc: keita.kobayashi.ym@renesas.com Cc: mgalbraith@suse.de Cc: sgurrappadi@nvidia.com Cc: vincent.guittot@linaro.org Cc: yuyang.du@intel.com Link: http://lkml.kernel.org/r/1476452472-24740-2-git-send-email-morten.rasmussen@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-10-14 13:41:07 +00:00
}
/*
* Predicts what cpu_util(@cpu) would return if @p was migrated (and enqueued)
* to @dst_cpu.
*/
static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu)
{
struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
unsigned long util_est, util = READ_ONCE(cfs_rq->avg.util_avg);
/*
* If @p migrates from @cpu to another, remove its contribution. Or,
* if @p migrates from another CPU to @cpu, add its contribution. In
* the other cases, @cpu is not impacted by the migration, so the
* util_avg should already be correct.
*/
if (task_cpu(p) == cpu && dst_cpu != cpu)
sub_positive(&util, task_util(p));
else if (task_cpu(p) != cpu && dst_cpu == cpu)
util += task_util(p);
if (sched_feat(UTIL_EST)) {
util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued);
/*
* During wake-up, the task isn't enqueued yet and doesn't
* appear in the cfs_rq->avg.util_est.enqueued of any rq,
* so just add it (if needed) to "simulate" what will be
* cpu_util() after the task has been enqueued.
*/
if (dst_cpu == cpu)
util_est += _task_util_est(p);
util = max(util, util_est);
}
return min(util, capacity_orig_of(cpu));
}
/*
sched/fair: Speed-up energy-aware wake-ups EAS computes the energy impact of migrating a waking task when deciding on which CPU it should run. However, the current approach is known to have a high algorithmic complexity, which can result in prohibitively high wake-up latencies on systems with complex energy models, such as systems with per-CPU DVFS. On such systems, the algorithm complexity is in O(n^2) (ignoring the cost of searching for performance states in the EM) with 'n' the number of CPUs. To address this, re-factor the EAS wake-up path to compute the energy 'delta' (with and without the task) on a per-performance domain basis, rather than system-wide, which brings the complexity down to O(n). No functional changes intended. Test results ~~~~~~~~~~~~ * Setup: Tested on a Google Pixel 3, with a Snapdragon 845 (4+4 CPUs, A55/A75). Base kernel is 5.3-rc5 + Pixel3 specific patches. Android userspace, no graphics. * Test case: Run a periodic rt-app task, with 16ms period, ramping down from 70% to 10%, in 5% steps of 500 ms each (json avail. at [1]). Frequencies of all CPUs are pinned to max (using scaling_min_freq CPUFreq sysfs entries) to reduce variability. The time to run select_task_rq_fair() is measured using the function profiler (/sys/kernel/debug/tracing/trace_stat/function*). See the test script for more details [2]. Test 1: I hacked the DT to 'fake' per-CPU DVFS. That is, we end up with one CPUFreq policy per CPU (8 policies in total). Since all frequencies are pinned to max for the test, this should have no impact on the actual frequency selection, but it does in the EAS calculation. +---------------------------+----------------------------------+ | Without patch | With patch | +-----+-----+----------+----------+-----+-----------------+----------+ | CPU | Hit | Avg (us) | s^2 (us) | Hit | Avg (us) | s^2 (us) | |-----+-----+----------+----------+-----+-----------------+----------+ | 0 | 274 | 38.303 | 1750.239 | 401 | 14.126 (-63.1%) | 146.625 | | 1 | 197 | 49.529 | 1695.852 | 314 | 16.135 (-67.4%) | 167.525 | | 2 | 142 | 34.296 | 1758.665 | 302 | 14.133 (-58.8%) | 130.071 | | 3 | 172 | 31.734 | 1490.975 | 641 | 14.637 (-53.9%) | 139.189 | | 4 | 316 | 7.834 | 178.217 | 425 | 5.413 (-30.9%) | 20.803 | | 5 | 447 | 8.424 | 144.638 | 556 | 5.929 (-29.6%) | 27.301 | | 6 | 581 | 14.886 | 346.793 | 456 | 5.711 (-61.6%) | 23.124 | | 7 | 456 | 10.005 | 211.187 | 997 | 4.708 (-52.9%) | 21.144 | +-----+-----+----------+----------+-----+-----------------+----------+ * Hit, Avg and s^2 are as reported by the function profiler Test 2: I also ran the same test with a normal DT, with 2 CPUFreq policies, to see if this causes regressions in the most common case. +---------------------------+----------------------------------+ | Without patch | With patch | +-----+-----+----------+----------+-----+-----------------+----------+ | CPU | Hit | Avg (us) | s^2 (us) | Hit | Avg (us) | s^2 (us) | |-----+-----+----------+----------+-----+-----------------+----------+ | 0 | 345 | 22.184 | 215.321 | 580 | 18.635 (-16.0%) | 146.892 | | 1 | 358 | 18.597 | 200.596 | 438 | 12.934 (-30.5%) | 104.604 | | 2 | 359 | 25.566 | 200.217 | 397 | 10.826 (-57.7%) | 74.021 | | 3 | 362 | 16.881 | 200.291 | 718 | 11.455 (-32.1%) | 102.280 | | 4 | 457 | 3.822 | 9.895 | 757 | 4.616 (+20.8%) | 13.369 | | 5 | 344 | 4.301 | 7.121 | 594 | 5.320 (+23.7%) | 18.798 | | 6 | 472 | 4.326 | 7.849 | 464 | 5.648 (+30.6%) | 22.022 | | 7 | 331 | 4.630 | 13.937 | 408 | 5.299 (+14.4%) | 18.273 | +-----+-----+----------+----------+-----+-----------------+----------+ * Hit, Avg and s^2 are as reported by the function profiler In addition to these two tests, I also ran 50 iterations of the Lisa EAS functional test suite [3] with this patch applied on Arm Juno r0, Arm Juno r2, Arm TC2 and Hikey960, and could not see any regressions (all EAS functional tests are passing). [1] https://paste.debian.net/1100055/ [2] https://paste.debian.net/1100057/ [3] https://github.com/ARM-software/lisa/blob/master/lisa/tests/scheduler/eas_behaviour.py Signed-off-by: Quentin Perret <quentin.perret@arm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: juri.lelli@redhat.com Cc: morten.rasmussen@arm.com Cc: qais.yousef@arm.com Cc: qperret@qperret.net Cc: rjw@rjwysocki.net Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Link: https://lkml.kernel.org/r/20190912094404.13802-1-qperret@qperret.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-09-12 09:44:04 +00:00
* compute_energy(): Estimates the energy that @pd would consume if @p was
* migrated to @dst_cpu. compute_energy() predicts what will be the utilization
sched/fair: Speed-up energy-aware wake-ups EAS computes the energy impact of migrating a waking task when deciding on which CPU it should run. However, the current approach is known to have a high algorithmic complexity, which can result in prohibitively high wake-up latencies on systems with complex energy models, such as systems with per-CPU DVFS. On such systems, the algorithm complexity is in O(n^2) (ignoring the cost of searching for performance states in the EM) with 'n' the number of CPUs. To address this, re-factor the EAS wake-up path to compute the energy 'delta' (with and without the task) on a per-performance domain basis, rather than system-wide, which brings the complexity down to O(n). No functional changes intended. Test results ~~~~~~~~~~~~ * Setup: Tested on a Google Pixel 3, with a Snapdragon 845 (4+4 CPUs, A55/A75). Base kernel is 5.3-rc5 + Pixel3 specific patches. Android userspace, no graphics. * Test case: Run a periodic rt-app task, with 16ms period, ramping down from 70% to 10%, in 5% steps of 500 ms each (json avail. at [1]). Frequencies of all CPUs are pinned to max (using scaling_min_freq CPUFreq sysfs entries) to reduce variability. The time to run select_task_rq_fair() is measured using the function profiler (/sys/kernel/debug/tracing/trace_stat/function*). See the test script for more details [2]. Test 1: I hacked the DT to 'fake' per-CPU DVFS. That is, we end up with one CPUFreq policy per CPU (8 policies in total). Since all frequencies are pinned to max for the test, this should have no impact on the actual frequency selection, but it does in the EAS calculation. +---------------------------+----------------------------------+ | Without patch | With patch | +-----+-----+----------+----------+-----+-----------------+----------+ | CPU | Hit | Avg (us) | s^2 (us) | Hit | Avg (us) | s^2 (us) | |-----+-----+----------+----------+-----+-----------------+----------+ | 0 | 274 | 38.303 | 1750.239 | 401 | 14.126 (-63.1%) | 146.625 | | 1 | 197 | 49.529 | 1695.852 | 314 | 16.135 (-67.4%) | 167.525 | | 2 | 142 | 34.296 | 1758.665 | 302 | 14.133 (-58.8%) | 130.071 | | 3 | 172 | 31.734 | 1490.975 | 641 | 14.637 (-53.9%) | 139.189 | | 4 | 316 | 7.834 | 178.217 | 425 | 5.413 (-30.9%) | 20.803 | | 5 | 447 | 8.424 | 144.638 | 556 | 5.929 (-29.6%) | 27.301 | | 6 | 581 | 14.886 | 346.793 | 456 | 5.711 (-61.6%) | 23.124 | | 7 | 456 | 10.005 | 211.187 | 997 | 4.708 (-52.9%) | 21.144 | +-----+-----+----------+----------+-----+-----------------+----------+ * Hit, Avg and s^2 are as reported by the function profiler Test 2: I also ran the same test with a normal DT, with 2 CPUFreq policies, to see if this causes regressions in the most common case. +---------------------------+----------------------------------+ | Without patch | With patch | +-----+-----+----------+----------+-----+-----------------+----------+ | CPU | Hit | Avg (us) | s^2 (us) | Hit | Avg (us) | s^2 (us) | |-----+-----+----------+----------+-----+-----------------+----------+ | 0 | 345 | 22.184 | 215.321 | 580 | 18.635 (-16.0%) | 146.892 | | 1 | 358 | 18.597 | 200.596 | 438 | 12.934 (-30.5%) | 104.604 | | 2 | 359 | 25.566 | 200.217 | 397 | 10.826 (-57.7%) | 74.021 | | 3 | 362 | 16.881 | 200.291 | 718 | 11.455 (-32.1%) | 102.280 | | 4 | 457 | 3.822 | 9.895 | 757 | 4.616 (+20.8%) | 13.369 | | 5 | 344 | 4.301 | 7.121 | 594 | 5.320 (+23.7%) | 18.798 | | 6 | 472 | 4.326 | 7.849 | 464 | 5.648 (+30.6%) | 22.022 | | 7 | 331 | 4.630 | 13.937 | 408 | 5.299 (+14.4%) | 18.273 | +-----+-----+----------+----------+-----+-----------------+----------+ * Hit, Avg and s^2 are as reported by the function profiler In addition to these two tests, I also ran 50 iterations of the Lisa EAS functional test suite [3] with this patch applied on Arm Juno r0, Arm Juno r2, Arm TC2 and Hikey960, and could not see any regressions (all EAS functional tests are passing). [1] https://paste.debian.net/1100055/ [2] https://paste.debian.net/1100057/ [3] https://github.com/ARM-software/lisa/blob/master/lisa/tests/scheduler/eas_behaviour.py Signed-off-by: Quentin Perret <quentin.perret@arm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: juri.lelli@redhat.com Cc: morten.rasmussen@arm.com Cc: qais.yousef@arm.com Cc: qperret@qperret.net Cc: rjw@rjwysocki.net Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Link: https://lkml.kernel.org/r/20190912094404.13802-1-qperret@qperret.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-09-12 09:44:04 +00:00
* landscape of @pd's CPUs after the task migration, and uses the Energy Model
* to compute what would be the energy if we decided to actually migrate that
* task.
*/
static long
compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd)
{
sched/fair: Speed-up energy-aware wake-ups EAS computes the energy impact of migrating a waking task when deciding on which CPU it should run. However, the current approach is known to have a high algorithmic complexity, which can result in prohibitively high wake-up latencies on systems with complex energy models, such as systems with per-CPU DVFS. On such systems, the algorithm complexity is in O(n^2) (ignoring the cost of searching for performance states in the EM) with 'n' the number of CPUs. To address this, re-factor the EAS wake-up path to compute the energy 'delta' (with and without the task) on a per-performance domain basis, rather than system-wide, which brings the complexity down to O(n). No functional changes intended. Test results ~~~~~~~~~~~~ * Setup: Tested on a Google Pixel 3, with a Snapdragon 845 (4+4 CPUs, A55/A75). Base kernel is 5.3-rc5 + Pixel3 specific patches. Android userspace, no graphics. * Test case: Run a periodic rt-app task, with 16ms period, ramping down from 70% to 10%, in 5% steps of 500 ms each (json avail. at [1]). Frequencies of all CPUs are pinned to max (using scaling_min_freq CPUFreq sysfs entries) to reduce variability. The time to run select_task_rq_fair() is measured using the function profiler (/sys/kernel/debug/tracing/trace_stat/function*). See the test script for more details [2]. Test 1: I hacked the DT to 'fake' per-CPU DVFS. That is, we end up with one CPUFreq policy per CPU (8 policies in total). Since all frequencies are pinned to max for the test, this should have no impact on the actual frequency selection, but it does in the EAS calculation. +---------------------------+----------------------------------+ | Without patch | With patch | +-----+-----+----------+----------+-----+-----------------+----------+ | CPU | Hit | Avg (us) | s^2 (us) | Hit | Avg (us) | s^2 (us) | |-----+-----+----------+----------+-----+-----------------+----------+ | 0 | 274 | 38.303 | 1750.239 | 401 | 14.126 (-63.1%) | 146.625 | | 1 | 197 | 49.529 | 1695.852 | 314 | 16.135 (-67.4%) | 167.525 | | 2 | 142 | 34.296 | 1758.665 | 302 | 14.133 (-58.8%) | 130.071 | | 3 | 172 | 31.734 | 1490.975 | 641 | 14.637 (-53.9%) | 139.189 | | 4 | 316 | 7.834 | 178.217 | 425 | 5.413 (-30.9%) | 20.803 | | 5 | 447 | 8.424 | 144.638 | 556 | 5.929 (-29.6%) | 27.301 | | 6 | 581 | 14.886 | 346.793 | 456 | 5.711 (-61.6%) | 23.124 | | 7 | 456 | 10.005 | 211.187 | 997 | 4.708 (-52.9%) | 21.144 | +-----+-----+----------+----------+-----+-----------------+----------+ * Hit, Avg and s^2 are as reported by the function profiler Test 2: I also ran the same test with a normal DT, with 2 CPUFreq policies, to see if this causes regressions in the most common case. +---------------------------+----------------------------------+ | Without patch | With patch | +-----+-----+----------+----------+-----+-----------------+----------+ | CPU | Hit | Avg (us) | s^2 (us) | Hit | Avg (us) | s^2 (us) | |-----+-----+----------+----------+-----+-----------------+----------+ | 0 | 345 | 22.184 | 215.321 | 580 | 18.635 (-16.0%) | 146.892 | | 1 | 358 | 18.597 | 200.596 | 438 | 12.934 (-30.5%) | 104.604 | | 2 | 359 | 25.566 | 200.217 | 397 | 10.826 (-57.7%) | 74.021 | | 3 | 362 | 16.881 | 200.291 | 718 | 11.455 (-32.1%) | 102.280 | | 4 | 457 | 3.822 | 9.895 | 757 | 4.616 (+20.8%) | 13.369 | | 5 | 344 | 4.301 | 7.121 | 594 | 5.320 (+23.7%) | 18.798 | | 6 | 472 | 4.326 | 7.849 | 464 | 5.648 (+30.6%) | 22.022 | | 7 | 331 | 4.630 | 13.937 | 408 | 5.299 (+14.4%) | 18.273 | +-----+-----+----------+----------+-----+-----------------+----------+ * Hit, Avg and s^2 are as reported by the function profiler In addition to these two tests, I also ran 50 iterations of the Lisa EAS functional test suite [3] with this patch applied on Arm Juno r0, Arm Juno r2, Arm TC2 and Hikey960, and could not see any regressions (all EAS functional tests are passing). [1] https://paste.debian.net/1100055/ [2] https://paste.debian.net/1100057/ [3] https://github.com/ARM-software/lisa/blob/master/lisa/tests/scheduler/eas_behaviour.py Signed-off-by: Quentin Perret <quentin.perret@arm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: juri.lelli@redhat.com Cc: morten.rasmussen@arm.com Cc: qais.yousef@arm.com Cc: qperret@qperret.net Cc: rjw@rjwysocki.net Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Link: https://lkml.kernel.org/r/20190912094404.13802-1-qperret@qperret.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-09-12 09:44:04 +00:00
struct cpumask *pd_mask = perf_domain_span(pd);
unsigned long cpu_cap = arch_scale_cpu_capacity(cpumask_first(pd_mask));
unsigned long max_util = 0, sum_util = 0;
int cpu;
sched/fair: Speed-up energy-aware wake-ups EAS computes the energy impact of migrating a waking task when deciding on which CPU it should run. However, the current approach is known to have a high algorithmic complexity, which can result in prohibitively high wake-up latencies on systems with complex energy models, such as systems with per-CPU DVFS. On such systems, the algorithm complexity is in O(n^2) (ignoring the cost of searching for performance states in the EM) with 'n' the number of CPUs. To address this, re-factor the EAS wake-up path to compute the energy 'delta' (with and without the task) on a per-performance domain basis, rather than system-wide, which brings the complexity down to O(n). No functional changes intended. Test results ~~~~~~~~~~~~ * Setup: Tested on a Google Pixel 3, with a Snapdragon 845 (4+4 CPUs, A55/A75). Base kernel is 5.3-rc5 + Pixel3 specific patches. Android userspace, no graphics. * Test case: Run a periodic rt-app task, with 16ms period, ramping down from 70% to 10%, in 5% steps of 500 ms each (json avail. at [1]). Frequencies of all CPUs are pinned to max (using scaling_min_freq CPUFreq sysfs entries) to reduce variability. The time to run select_task_rq_fair() is measured using the function profiler (/sys/kernel/debug/tracing/trace_stat/function*). See the test script for more details [2]. Test 1: I hacked the DT to 'fake' per-CPU DVFS. That is, we end up with one CPUFreq policy per CPU (8 policies in total). Since all frequencies are pinned to max for the test, this should have no impact on the actual frequency selection, but it does in the EAS calculation. +---------------------------+----------------------------------+ | Without patch | With patch | +-----+-----+----------+----------+-----+-----------------+----------+ | CPU | Hit | Avg (us) | s^2 (us) | Hit | Avg (us) | s^2 (us) | |-----+-----+----------+----------+-----+-----------------+----------+ | 0 | 274 | 38.303 | 1750.239 | 401 | 14.126 (-63.1%) | 146.625 | | 1 | 197 | 49.529 | 1695.852 | 314 | 16.135 (-67.4%) | 167.525 | | 2 | 142 | 34.296 | 1758.665 | 302 | 14.133 (-58.8%) | 130.071 | | 3 | 172 | 31.734 | 1490.975 | 641 | 14.637 (-53.9%) | 139.189 | | 4 | 316 | 7.834 | 178.217 | 425 | 5.413 (-30.9%) | 20.803 | | 5 | 447 | 8.424 | 144.638 | 556 | 5.929 (-29.6%) | 27.301 | | 6 | 581 | 14.886 | 346.793 | 456 | 5.711 (-61.6%) | 23.124 | | 7 | 456 | 10.005 | 211.187 | 997 | 4.708 (-52.9%) | 21.144 | +-----+-----+----------+----------+-----+-----------------+----------+ * Hit, Avg and s^2 are as reported by the function profiler Test 2: I also ran the same test with a normal DT, with 2 CPUFreq policies, to see if this causes regressions in the most common case. +---------------------------+----------------------------------+ | Without patch | With patch | +-----+-----+----------+----------+-----+-----------------+----------+ | CPU | Hit | Avg (us) | s^2 (us) | Hit | Avg (us) | s^2 (us) | |-----+-----+----------+----------+-----+-----------------+----------+ | 0 | 345 | 22.184 | 215.321 | 580 | 18.635 (-16.0%) | 146.892 | | 1 | 358 | 18.597 | 200.596 | 438 | 12.934 (-30.5%) | 104.604 | | 2 | 359 | 25.566 | 200.217 | 397 | 10.826 (-57.7%) | 74.021 | | 3 | 362 | 16.881 | 200.291 | 718 | 11.455 (-32.1%) | 102.280 | | 4 | 457 | 3.822 | 9.895 | 757 | 4.616 (+20.8%) | 13.369 | | 5 | 344 | 4.301 | 7.121 | 594 | 5.320 (+23.7%) | 18.798 | | 6 | 472 | 4.326 | 7.849 | 464 | 5.648 (+30.6%) | 22.022 | | 7 | 331 | 4.630 | 13.937 | 408 | 5.299 (+14.4%) | 18.273 | +-----+-----+----------+----------+-----+-----------------+----------+ * Hit, Avg and s^2 are as reported by the function profiler In addition to these two tests, I also ran 50 iterations of the Lisa EAS functional test suite [3] with this patch applied on Arm Juno r0, Arm Juno r2, Arm TC2 and Hikey960, and could not see any regressions (all EAS functional tests are passing). [1] https://paste.debian.net/1100055/ [2] https://paste.debian.net/1100057/ [3] https://github.com/ARM-software/lisa/blob/master/lisa/tests/scheduler/eas_behaviour.py Signed-off-by: Quentin Perret <quentin.perret@arm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: juri.lelli@redhat.com Cc: morten.rasmussen@arm.com Cc: qais.yousef@arm.com Cc: qperret@qperret.net Cc: rjw@rjwysocki.net Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Link: https://lkml.kernel.org/r/20190912094404.13802-1-qperret@qperret.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-09-12 09:44:04 +00:00
/*
* The capacity state of CPUs of the current rd can be driven by CPUs
* of another rd if they belong to the same pd. So, account for the
* utilization of these CPUs too by masking pd with cpu_online_mask
* instead of the rd span.
*
* If an entire pd is outside of the current rd, it will not appear in
* its pd list and will not be accounted by compute_energy().
*/
for_each_cpu_and(cpu, pd_mask, cpu_online_mask) {
unsigned long cpu_util, util_cfs = cpu_util_next(cpu, p, dst_cpu);
struct task_struct *tsk = cpu == dst_cpu ? p : NULL;
sched/uclamp: Add uclamp support to energy_compute() The Energy Aware Scheduler (EAS) estimates the energy impact of waking up a task on a given CPU. This estimation is based on: a) an (active) power consumption defined for each CPU frequency b) an estimation of which frequency will be used on each CPU c) an estimation of the busy time (utilization) of each CPU Utilization clamping can affect both b) and c). A CPU is expected to run: - on an higher than required frequency, but for a shorter time, in case its estimated utilization will be smaller than the minimum utilization enforced by uclamp - on a smaller than required frequency, but for a longer time, in case its estimated utilization is bigger than the maximum utilization enforced by uclamp While compute_energy() already accounts clamping effects on busy time, the clamping effects on frequency selection are currently ignored. Fix it by considering how CPU clamp values will be affected by a task waking up and being RUNNABLE on that CPU. Do that by refactoring schedutil_freq_util() to take an additional task_struct* which allows EAS to evaluate the impact on clamp values of a task being eventually queued in a CPU. Clamp values are applied to the RT+CFS utilization only when a FREQUENCY_UTIL is required by compute_energy(). Do note that switching from ENERGY_UTIL to FREQUENCY_UTIL in the computation of the cpu_util signal implies that we are more likely to estimate the highest OPP when a RT task is running in another CPU of the same performance domain. This can have an impact on energy estimation but: - it's not easy to say which approach is better, since it depends on the use case - the original approach could still be obtained by setting a smaller task-specific util_min whenever required Since we are at that: - rename schedutil_freq_util() into schedutil_cpu_util(), since it's not only used for frequency selection. Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alessio Balsini <balsini@android.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Joel Fernandes <joelaf@google.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Quentin Perret <quentin.perret@arm.com> Cc: Rafael J . Wysocki <rafael.j.wysocki@intel.com> Cc: Steve Muckle <smuckle@google.com> Cc: Suren Baghdasaryan <surenb@google.com> Cc: Tejun Heo <tj@kernel.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Todd Kjos <tkjos@google.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Link: https://lkml.kernel.org/r/20190621084217.8167-12-patrick.bellasi@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-06-21 08:42:12 +00:00
/*
sched/fair: Speed-up energy-aware wake-ups EAS computes the energy impact of migrating a waking task when deciding on which CPU it should run. However, the current approach is known to have a high algorithmic complexity, which can result in prohibitively high wake-up latencies on systems with complex energy models, such as systems with per-CPU DVFS. On such systems, the algorithm complexity is in O(n^2) (ignoring the cost of searching for performance states in the EM) with 'n' the number of CPUs. To address this, re-factor the EAS wake-up path to compute the energy 'delta' (with and without the task) on a per-performance domain basis, rather than system-wide, which brings the complexity down to O(n). No functional changes intended. Test results ~~~~~~~~~~~~ * Setup: Tested on a Google Pixel 3, with a Snapdragon 845 (4+4 CPUs, A55/A75). Base kernel is 5.3-rc5 + Pixel3 specific patches. Android userspace, no graphics. * Test case: Run a periodic rt-app task, with 16ms period, ramping down from 70% to 10%, in 5% steps of 500 ms each (json avail. at [1]). Frequencies of all CPUs are pinned to max (using scaling_min_freq CPUFreq sysfs entries) to reduce variability. The time to run select_task_rq_fair() is measured using the function profiler (/sys/kernel/debug/tracing/trace_stat/function*). See the test script for more details [2]. Test 1: I hacked the DT to 'fake' per-CPU DVFS. That is, we end up with one CPUFreq policy per CPU (8 policies in total). Since all frequencies are pinned to max for the test, this should have no impact on the actual frequency selection, but it does in the EAS calculation. +---------------------------+----------------------------------+ | Without patch | With patch | +-----+-----+----------+----------+-----+-----------------+----------+ | CPU | Hit | Avg (us) | s^2 (us) | Hit | Avg (us) | s^2 (us) | |-----+-----+----------+----------+-----+-----------------+----------+ | 0 | 274 | 38.303 | 1750.239 | 401 | 14.126 (-63.1%) | 146.625 | | 1 | 197 | 49.529 | 1695.852 | 314 | 16.135 (-67.4%) | 167.525 | | 2 | 142 | 34.296 | 1758.665 | 302 | 14.133 (-58.8%) | 130.071 | | 3 | 172 | 31.734 | 1490.975 | 641 | 14.637 (-53.9%) | 139.189 | | 4 | 316 | 7.834 | 178.217 | 425 | 5.413 (-30.9%) | 20.803 | | 5 | 447 | 8.424 | 144.638 | 556 | 5.929 (-29.6%) | 27.301 | | 6 | 581 | 14.886 | 346.793 | 456 | 5.711 (-61.6%) | 23.124 | | 7 | 456 | 10.005 | 211.187 | 997 | 4.708 (-52.9%) | 21.144 | +-----+-----+----------+----------+-----+-----------------+----------+ * Hit, Avg and s^2 are as reported by the function profiler Test 2: I also ran the same test with a normal DT, with 2 CPUFreq policies, to see if this causes regressions in the most common case. +---------------------------+----------------------------------+ | Without patch | With patch | +-----+-----+----------+----------+-----+-----------------+----------+ | CPU | Hit | Avg (us) | s^2 (us) | Hit | Avg (us) | s^2 (us) | |-----+-----+----------+----------+-----+-----------------+----------+ | 0 | 345 | 22.184 | 215.321 | 580 | 18.635 (-16.0%) | 146.892 | | 1 | 358 | 18.597 | 200.596 | 438 | 12.934 (-30.5%) | 104.604 | | 2 | 359 | 25.566 | 200.217 | 397 | 10.826 (-57.7%) | 74.021 | | 3 | 362 | 16.881 | 200.291 | 718 | 11.455 (-32.1%) | 102.280 | | 4 | 457 | 3.822 | 9.895 | 757 | 4.616 (+20.8%) | 13.369 | | 5 | 344 | 4.301 | 7.121 | 594 | 5.320 (+23.7%) | 18.798 | | 6 | 472 | 4.326 | 7.849 | 464 | 5.648 (+30.6%) | 22.022 | | 7 | 331 | 4.630 | 13.937 | 408 | 5.299 (+14.4%) | 18.273 | +-----+-----+----------+----------+-----+-----------------+----------+ * Hit, Avg and s^2 are as reported by the function profiler In addition to these two tests, I also ran 50 iterations of the Lisa EAS functional test suite [3] with this patch applied on Arm Juno r0, Arm Juno r2, Arm TC2 and Hikey960, and could not see any regressions (all EAS functional tests are passing). [1] https://paste.debian.net/1100055/ [2] https://paste.debian.net/1100057/ [3] https://github.com/ARM-software/lisa/blob/master/lisa/tests/scheduler/eas_behaviour.py Signed-off-by: Quentin Perret <quentin.perret@arm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: juri.lelli@redhat.com Cc: morten.rasmussen@arm.com Cc: qais.yousef@arm.com Cc: qperret@qperret.net Cc: rjw@rjwysocki.net Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Link: https://lkml.kernel.org/r/20190912094404.13802-1-qperret@qperret.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-09-12 09:44:04 +00:00
* Busy time computation: utilization clamping is not
* required since the ratio (sum_util / cpu_capacity)
* is already enough to scale the EM reported power
* consumption at the (eventually clamped) cpu_capacity.
sched/uclamp: Add uclamp support to energy_compute() The Energy Aware Scheduler (EAS) estimates the energy impact of waking up a task on a given CPU. This estimation is based on: a) an (active) power consumption defined for each CPU frequency b) an estimation of which frequency will be used on each CPU c) an estimation of the busy time (utilization) of each CPU Utilization clamping can affect both b) and c). A CPU is expected to run: - on an higher than required frequency, but for a shorter time, in case its estimated utilization will be smaller than the minimum utilization enforced by uclamp - on a smaller than required frequency, but for a longer time, in case its estimated utilization is bigger than the maximum utilization enforced by uclamp While compute_energy() already accounts clamping effects on busy time, the clamping effects on frequency selection are currently ignored. Fix it by considering how CPU clamp values will be affected by a task waking up and being RUNNABLE on that CPU. Do that by refactoring schedutil_freq_util() to take an additional task_struct* which allows EAS to evaluate the impact on clamp values of a task being eventually queued in a CPU. Clamp values are applied to the RT+CFS utilization only when a FREQUENCY_UTIL is required by compute_energy(). Do note that switching from ENERGY_UTIL to FREQUENCY_UTIL in the computation of the cpu_util signal implies that we are more likely to estimate the highest OPP when a RT task is running in another CPU of the same performance domain. This can have an impact on energy estimation but: - it's not easy to say which approach is better, since it depends on the use case - the original approach could still be obtained by setting a smaller task-specific util_min whenever required Since we are at that: - rename schedutil_freq_util() into schedutil_cpu_util(), since it's not only used for frequency selection. Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alessio Balsini <balsini@android.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Joel Fernandes <joelaf@google.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Quentin Perret <quentin.perret@arm.com> Cc: Rafael J . Wysocki <rafael.j.wysocki@intel.com> Cc: Steve Muckle <smuckle@google.com> Cc: Suren Baghdasaryan <surenb@google.com> Cc: Tejun Heo <tj@kernel.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Todd Kjos <tkjos@google.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Link: https://lkml.kernel.org/r/20190621084217.8167-12-patrick.bellasi@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-06-21 08:42:12 +00:00
*/
sched/fair: Speed-up energy-aware wake-ups EAS computes the energy impact of migrating a waking task when deciding on which CPU it should run. However, the current approach is known to have a high algorithmic complexity, which can result in prohibitively high wake-up latencies on systems with complex energy models, such as systems with per-CPU DVFS. On such systems, the algorithm complexity is in O(n^2) (ignoring the cost of searching for performance states in the EM) with 'n' the number of CPUs. To address this, re-factor the EAS wake-up path to compute the energy 'delta' (with and without the task) on a per-performance domain basis, rather than system-wide, which brings the complexity down to O(n). No functional changes intended. Test results ~~~~~~~~~~~~ * Setup: Tested on a Google Pixel 3, with a Snapdragon 845 (4+4 CPUs, A55/A75). Base kernel is 5.3-rc5 + Pixel3 specific patches. Android userspace, no graphics. * Test case: Run a periodic rt-app task, with 16ms period, ramping down from 70% to 10%, in 5% steps of 500 ms each (json avail. at [1]). Frequencies of all CPUs are pinned to max (using scaling_min_freq CPUFreq sysfs entries) to reduce variability. The time to run select_task_rq_fair() is measured using the function profiler (/sys/kernel/debug/tracing/trace_stat/function*). See the test script for more details [2]. Test 1: I hacked the DT to 'fake' per-CPU DVFS. That is, we end up with one CPUFreq policy per CPU (8 policies in total). Since all frequencies are pinned to max for the test, this should have no impact on the actual frequency selection, but it does in the EAS calculation. +---------------------------+----------------------------------+ | Without patch | With patch | +-----+-----+----------+----------+-----+-----------------+----------+ | CPU | Hit | Avg (us) | s^2 (us) | Hit | Avg (us) | s^2 (us) | |-----+-----+----------+----------+-----+-----------------+----------+ | 0 | 274 | 38.303 | 1750.239 | 401 | 14.126 (-63.1%) | 146.625 | | 1 | 197 | 49.529 | 1695.852 | 314 | 16.135 (-67.4%) | 167.525 | | 2 | 142 | 34.296 | 1758.665 | 302 | 14.133 (-58.8%) | 130.071 | | 3 | 172 | 31.734 | 1490.975 | 641 | 14.637 (-53.9%) | 139.189 | | 4 | 316 | 7.834 | 178.217 | 425 | 5.413 (-30.9%) | 20.803 | | 5 | 447 | 8.424 | 144.638 | 556 | 5.929 (-29.6%) | 27.301 | | 6 | 581 | 14.886 | 346.793 | 456 | 5.711 (-61.6%) | 23.124 | | 7 | 456 | 10.005 | 211.187 | 997 | 4.708 (-52.9%) | 21.144 | +-----+-----+----------+----------+-----+-----------------+----------+ * Hit, Avg and s^2 are as reported by the function profiler Test 2: I also ran the same test with a normal DT, with 2 CPUFreq policies, to see if this causes regressions in the most common case. +---------------------------+----------------------------------+ | Without patch | With patch | +-----+-----+----------+----------+-----+-----------------+----------+ | CPU | Hit | Avg (us) | s^2 (us) | Hit | Avg (us) | s^2 (us) | |-----+-----+----------+----------+-----+-----------------+----------+ | 0 | 345 | 22.184 | 215.321 | 580 | 18.635 (-16.0%) | 146.892 | | 1 | 358 | 18.597 | 200.596 | 438 | 12.934 (-30.5%) | 104.604 | | 2 | 359 | 25.566 | 200.217 | 397 | 10.826 (-57.7%) | 74.021 | | 3 | 362 | 16.881 | 200.291 | 718 | 11.455 (-32.1%) | 102.280 | | 4 | 457 | 3.822 | 9.895 | 757 | 4.616 (+20.8%) | 13.369 | | 5 | 344 | 4.301 | 7.121 | 594 | 5.320 (+23.7%) | 18.798 | | 6 | 472 | 4.326 | 7.849 | 464 | 5.648 (+30.6%) | 22.022 | | 7 | 331 | 4.630 | 13.937 | 408 | 5.299 (+14.4%) | 18.273 | +-----+-----+----------+----------+-----+-----------------+----------+ * Hit, Avg and s^2 are as reported by the function profiler In addition to these two tests, I also ran 50 iterations of the Lisa EAS functional test suite [3] with this patch applied on Arm Juno r0, Arm Juno r2, Arm TC2 and Hikey960, and could not see any regressions (all EAS functional tests are passing). [1] https://paste.debian.net/1100055/ [2] https://paste.debian.net/1100057/ [3] https://github.com/ARM-software/lisa/blob/master/lisa/tests/scheduler/eas_behaviour.py Signed-off-by: Quentin Perret <quentin.perret@arm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: juri.lelli@redhat.com Cc: morten.rasmussen@arm.com Cc: qais.yousef@arm.com Cc: qperret@qperret.net Cc: rjw@rjwysocki.net Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Link: https://lkml.kernel.org/r/20190912094404.13802-1-qperret@qperret.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-09-12 09:44:04 +00:00
sum_util += schedutil_cpu_util(cpu, util_cfs, cpu_cap,
ENERGY_UTIL, NULL);
sched/uclamp: Add uclamp support to energy_compute() The Energy Aware Scheduler (EAS) estimates the energy impact of waking up a task on a given CPU. This estimation is based on: a) an (active) power consumption defined for each CPU frequency b) an estimation of which frequency will be used on each CPU c) an estimation of the busy time (utilization) of each CPU Utilization clamping can affect both b) and c). A CPU is expected to run: - on an higher than required frequency, but for a shorter time, in case its estimated utilization will be smaller than the minimum utilization enforced by uclamp - on a smaller than required frequency, but for a longer time, in case its estimated utilization is bigger than the maximum utilization enforced by uclamp While compute_energy() already accounts clamping effects on busy time, the clamping effects on frequency selection are currently ignored. Fix it by considering how CPU clamp values will be affected by a task waking up and being RUNNABLE on that CPU. Do that by refactoring schedutil_freq_util() to take an additional task_struct* which allows EAS to evaluate the impact on clamp values of a task being eventually queued in a CPU. Clamp values are applied to the RT+CFS utilization only when a FREQUENCY_UTIL is required by compute_energy(). Do note that switching from ENERGY_UTIL to FREQUENCY_UTIL in the computation of the cpu_util signal implies that we are more likely to estimate the highest OPP when a RT task is running in another CPU of the same performance domain. This can have an impact on energy estimation but: - it's not easy to say which approach is better, since it depends on the use case - the original approach could still be obtained by setting a smaller task-specific util_min whenever required Since we are at that: - rename schedutil_freq_util() into schedutil_cpu_util(), since it's not only used for frequency selection. Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alessio Balsini <balsini@android.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Joel Fernandes <joelaf@google.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Quentin Perret <quentin.perret@arm.com> Cc: Rafael J . Wysocki <rafael.j.wysocki@intel.com> Cc: Steve Muckle <smuckle@google.com> Cc: Suren Baghdasaryan <surenb@google.com> Cc: Tejun Heo <tj@kernel.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Todd Kjos <tkjos@google.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Link: https://lkml.kernel.org/r/20190621084217.8167-12-patrick.bellasi@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-06-21 08:42:12 +00:00
/*
sched/fair: Speed-up energy-aware wake-ups EAS computes the energy impact of migrating a waking task when deciding on which CPU it should run. However, the current approach is known to have a high algorithmic complexity, which can result in prohibitively high wake-up latencies on systems with complex energy models, such as systems with per-CPU DVFS. On such systems, the algorithm complexity is in O(n^2) (ignoring the cost of searching for performance states in the EM) with 'n' the number of CPUs. To address this, re-factor the EAS wake-up path to compute the energy 'delta' (with and without the task) on a per-performance domain basis, rather than system-wide, which brings the complexity down to O(n). No functional changes intended. Test results ~~~~~~~~~~~~ * Setup: Tested on a Google Pixel 3, with a Snapdragon 845 (4+4 CPUs, A55/A75). Base kernel is 5.3-rc5 + Pixel3 specific patches. Android userspace, no graphics. * Test case: Run a periodic rt-app task, with 16ms period, ramping down from 70% to 10%, in 5% steps of 500 ms each (json avail. at [1]). Frequencies of all CPUs are pinned to max (using scaling_min_freq CPUFreq sysfs entries) to reduce variability. The time to run select_task_rq_fair() is measured using the function profiler (/sys/kernel/debug/tracing/trace_stat/function*). See the test script for more details [2]. Test 1: I hacked the DT to 'fake' per-CPU DVFS. That is, we end up with one CPUFreq policy per CPU (8 policies in total). Since all frequencies are pinned to max for the test, this should have no impact on the actual frequency selection, but it does in the EAS calculation. +---------------------------+----------------------------------+ | Without patch | With patch | +-----+-----+----------+----------+-----+-----------------+----------+ | CPU | Hit | Avg (us) | s^2 (us) | Hit | Avg (us) | s^2 (us) | |-----+-----+----------+----------+-----+-----------------+----------+ | 0 | 274 | 38.303 | 1750.239 | 401 | 14.126 (-63.1%) | 146.625 | | 1 | 197 | 49.529 | 1695.852 | 314 | 16.135 (-67.4%) | 167.525 | | 2 | 142 | 34.296 | 1758.665 | 302 | 14.133 (-58.8%) | 130.071 | | 3 | 172 | 31.734 | 1490.975 | 641 | 14.637 (-53.9%) | 139.189 | | 4 | 316 | 7.834 | 178.217 | 425 | 5.413 (-30.9%) | 20.803 | | 5 | 447 | 8.424 | 144.638 | 556 | 5.929 (-29.6%) | 27.301 | | 6 | 581 | 14.886 | 346.793 | 456 | 5.711 (-61.6%) | 23.124 | | 7 | 456 | 10.005 | 211.187 | 997 | 4.708 (-52.9%) | 21.144 | +-----+-----+----------+----------+-----+-----------------+----------+ * Hit, Avg and s^2 are as reported by the function profiler Test 2: I also ran the same test with a normal DT, with 2 CPUFreq policies, to see if this causes regressions in the most common case. +---------------------------+----------------------------------+ | Without patch | With patch | +-----+-----+----------+----------+-----+-----------------+----------+ | CPU | Hit | Avg (us) | s^2 (us) | Hit | Avg (us) | s^2 (us) | |-----+-----+----------+----------+-----+-----------------+----------+ | 0 | 345 | 22.184 | 215.321 | 580 | 18.635 (-16.0%) | 146.892 | | 1 | 358 | 18.597 | 200.596 | 438 | 12.934 (-30.5%) | 104.604 | | 2 | 359 | 25.566 | 200.217 | 397 | 10.826 (-57.7%) | 74.021 | | 3 | 362 | 16.881 | 200.291 | 718 | 11.455 (-32.1%) | 102.280 | | 4 | 457 | 3.822 | 9.895 | 757 | 4.616 (+20.8%) | 13.369 | | 5 | 344 | 4.301 | 7.121 | 594 | 5.320 (+23.7%) | 18.798 | | 6 | 472 | 4.326 | 7.849 | 464 | 5.648 (+30.6%) | 22.022 | | 7 | 331 | 4.630 | 13.937 | 408 | 5.299 (+14.4%) | 18.273 | +-----+-----+----------+----------+-----+-----------------+----------+ * Hit, Avg and s^2 are as reported by the function profiler In addition to these two tests, I also ran 50 iterations of the Lisa EAS functional test suite [3] with this patch applied on Arm Juno r0, Arm Juno r2, Arm TC2 and Hikey960, and could not see any regressions (all EAS functional tests are passing). [1] https://paste.debian.net/1100055/ [2] https://paste.debian.net/1100057/ [3] https://github.com/ARM-software/lisa/blob/master/lisa/tests/scheduler/eas_behaviour.py Signed-off-by: Quentin Perret <quentin.perret@arm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: juri.lelli@redhat.com Cc: morten.rasmussen@arm.com Cc: qais.yousef@arm.com Cc: qperret@qperret.net Cc: rjw@rjwysocki.net Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Link: https://lkml.kernel.org/r/20190912094404.13802-1-qperret@qperret.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-09-12 09:44:04 +00:00
* Performance domain frequency: utilization clamping
* must be considered since it affects the selection
* of the performance domain frequency.
* NOTE: in case RT tasks are running, by default the
* FREQUENCY_UTIL's utilization can be max OPP.
*/
sched/fair: Speed-up energy-aware wake-ups EAS computes the energy impact of migrating a waking task when deciding on which CPU it should run. However, the current approach is known to have a high algorithmic complexity, which can result in prohibitively high wake-up latencies on systems with complex energy models, such as systems with per-CPU DVFS. On such systems, the algorithm complexity is in O(n^2) (ignoring the cost of searching for performance states in the EM) with 'n' the number of CPUs. To address this, re-factor the EAS wake-up path to compute the energy 'delta' (with and without the task) on a per-performance domain basis, rather than system-wide, which brings the complexity down to O(n). No functional changes intended. Test results ~~~~~~~~~~~~ * Setup: Tested on a Google Pixel 3, with a Snapdragon 845 (4+4 CPUs, A55/A75). Base kernel is 5.3-rc5 + Pixel3 specific patches. Android userspace, no graphics. * Test case: Run a periodic rt-app task, with 16ms period, ramping down from 70% to 10%, in 5% steps of 500 ms each (json avail. at [1]). Frequencies of all CPUs are pinned to max (using scaling_min_freq CPUFreq sysfs entries) to reduce variability. The time to run select_task_rq_fair() is measured using the function profiler (/sys/kernel/debug/tracing/trace_stat/function*). See the test script for more details [2]. Test 1: I hacked the DT to 'fake' per-CPU DVFS. That is, we end up with one CPUFreq policy per CPU (8 policies in total). Since all frequencies are pinned to max for the test, this should have no impact on the actual frequency selection, but it does in the EAS calculation. +---------------------------+----------------------------------+ | Without patch | With patch | +-----+-----+----------+----------+-----+-----------------+----------+ | CPU | Hit | Avg (us) | s^2 (us) | Hit | Avg (us) | s^2 (us) | |-----+-----+----------+----------+-----+-----------------+----------+ | 0 | 274 | 38.303 | 1750.239 | 401 | 14.126 (-63.1%) | 146.625 | | 1 | 197 | 49.529 | 1695.852 | 314 | 16.135 (-67.4%) | 167.525 | | 2 | 142 | 34.296 | 1758.665 | 302 | 14.133 (-58.8%) | 130.071 | | 3 | 172 | 31.734 | 1490.975 | 641 | 14.637 (-53.9%) | 139.189 | | 4 | 316 | 7.834 | 178.217 | 425 | 5.413 (-30.9%) | 20.803 | | 5 | 447 | 8.424 | 144.638 | 556 | 5.929 (-29.6%) | 27.301 | | 6 | 581 | 14.886 | 346.793 | 456 | 5.711 (-61.6%) | 23.124 | | 7 | 456 | 10.005 | 211.187 | 997 | 4.708 (-52.9%) | 21.144 | +-----+-----+----------+----------+-----+-----------------+----------+ * Hit, Avg and s^2 are as reported by the function profiler Test 2: I also ran the same test with a normal DT, with 2 CPUFreq policies, to see if this causes regressions in the most common case. +---------------------------+----------------------------------+ | Without patch | With patch | +-----+-----+----------+----------+-----+-----------------+----------+ | CPU | Hit | Avg (us) | s^2 (us) | Hit | Avg (us) | s^2 (us) | |-----+-----+----------+----------+-----+-----------------+----------+ | 0 | 345 | 22.184 | 215.321 | 580 | 18.635 (-16.0%) | 146.892 | | 1 | 358 | 18.597 | 200.596 | 438 | 12.934 (-30.5%) | 104.604 | | 2 | 359 | 25.566 | 200.217 | 397 | 10.826 (-57.7%) | 74.021 | | 3 | 362 | 16.881 | 200.291 | 718 | 11.455 (-32.1%) | 102.280 | | 4 | 457 | 3.822 | 9.895 | 757 | 4.616 (+20.8%) | 13.369 | | 5 | 344 | 4.301 | 7.121 | 594 | 5.320 (+23.7%) | 18.798 | | 6 | 472 | 4.326 | 7.849 | 464 | 5.648 (+30.6%) | 22.022 | | 7 | 331 | 4.630 | 13.937 | 408 | 5.299 (+14.4%) | 18.273 | +-----+-----+----------+----------+-----+-----------------+----------+ * Hit, Avg and s^2 are as reported by the function profiler In addition to these two tests, I also ran 50 iterations of the Lisa EAS functional test suite [3] with this patch applied on Arm Juno r0, Arm Juno r2, Arm TC2 and Hikey960, and could not see any regressions (all EAS functional tests are passing). [1] https://paste.debian.net/1100055/ [2] https://paste.debian.net/1100057/ [3] https://github.com/ARM-software/lisa/blob/master/lisa/tests/scheduler/eas_behaviour.py Signed-off-by: Quentin Perret <quentin.perret@arm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: juri.lelli@redhat.com Cc: morten.rasmussen@arm.com Cc: qais.yousef@arm.com Cc: qperret@qperret.net Cc: rjw@rjwysocki.net Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Link: https://lkml.kernel.org/r/20190912094404.13802-1-qperret@qperret.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-09-12 09:44:04 +00:00
cpu_util = schedutil_cpu_util(cpu, util_cfs, cpu_cap,
FREQUENCY_UTIL, tsk);
max_util = max(max_util, cpu_util);
}
sched/fair: Speed-up energy-aware wake-ups EAS computes the energy impact of migrating a waking task when deciding on which CPU it should run. However, the current approach is known to have a high algorithmic complexity, which can result in prohibitively high wake-up latencies on systems with complex energy models, such as systems with per-CPU DVFS. On such systems, the algorithm complexity is in O(n^2) (ignoring the cost of searching for performance states in the EM) with 'n' the number of CPUs. To address this, re-factor the EAS wake-up path to compute the energy 'delta' (with and without the task) on a per-performance domain basis, rather than system-wide, which brings the complexity down to O(n). No functional changes intended. Test results ~~~~~~~~~~~~ * Setup: Tested on a Google Pixel 3, with a Snapdragon 845 (4+4 CPUs, A55/A75). Base kernel is 5.3-rc5 + Pixel3 specific patches. Android userspace, no graphics. * Test case: Run a periodic rt-app task, with 16ms period, ramping down from 70% to 10%, in 5% steps of 500 ms each (json avail. at [1]). Frequencies of all CPUs are pinned to max (using scaling_min_freq CPUFreq sysfs entries) to reduce variability. The time to run select_task_rq_fair() is measured using the function profiler (/sys/kernel/debug/tracing/trace_stat/function*). See the test script for more details [2]. Test 1: I hacked the DT to 'fake' per-CPU DVFS. That is, we end up with one CPUFreq policy per CPU (8 policies in total). Since all frequencies are pinned to max for the test, this should have no impact on the actual frequency selection, but it does in the EAS calculation. +---------------------------+----------------------------------+ | Without patch | With patch | +-----+-----+----------+----------+-----+-----------------+----------+ | CPU | Hit | Avg (us) | s^2 (us) | Hit | Avg (us) | s^2 (us) | |-----+-----+----------+----------+-----+-----------------+----------+ | 0 | 274 | 38.303 | 1750.239 | 401 | 14.126 (-63.1%) | 146.625 | | 1 | 197 | 49.529 | 1695.852 | 314 | 16.135 (-67.4%) | 167.525 | | 2 | 142 | 34.296 | 1758.665 | 302 | 14.133 (-58.8%) | 130.071 | | 3 | 172 | 31.734 | 1490.975 | 641 | 14.637 (-53.9%) | 139.189 | | 4 | 316 | 7.834 | 178.217 | 425 | 5.413 (-30.9%) | 20.803 | | 5 | 447 | 8.424 | 144.638 | 556 | 5.929 (-29.6%) | 27.301 | | 6 | 581 | 14.886 | 346.793 | 456 | 5.711 (-61.6%) | 23.124 | | 7 | 456 | 10.005 | 211.187 | 997 | 4.708 (-52.9%) | 21.144 | +-----+-----+----------+----------+-----+-----------------+----------+ * Hit, Avg and s^2 are as reported by the function profiler Test 2: I also ran the same test with a normal DT, with 2 CPUFreq policies, to see if this causes regressions in the most common case. +---------------------------+----------------------------------+ | Without patch | With patch | +-----+-----+----------+----------+-----+-----------------+----------+ | CPU | Hit | Avg (us) | s^2 (us) | Hit | Avg (us) | s^2 (us) | |-----+-----+----------+----------+-----+-----------------+----------+ | 0 | 345 | 22.184 | 215.321 | 580 | 18.635 (-16.0%) | 146.892 | | 1 | 358 | 18.597 | 200.596 | 438 | 12.934 (-30.5%) | 104.604 | | 2 | 359 | 25.566 | 200.217 | 397 | 10.826 (-57.7%) | 74.021 | | 3 | 362 | 16.881 | 200.291 | 718 | 11.455 (-32.1%) | 102.280 | | 4 | 457 | 3.822 | 9.895 | 757 | 4.616 (+20.8%) | 13.369 | | 5 | 344 | 4.301 | 7.121 | 594 | 5.320 (+23.7%) | 18.798 | | 6 | 472 | 4.326 | 7.849 | 464 | 5.648 (+30.6%) | 22.022 | | 7 | 331 | 4.630 | 13.937 | 408 | 5.299 (+14.4%) | 18.273 | +-----+-----+----------+----------+-----+-----------------+----------+ * Hit, Avg and s^2 are as reported by the function profiler In addition to these two tests, I also ran 50 iterations of the Lisa EAS functional test suite [3] with this patch applied on Arm Juno r0, Arm Juno r2, Arm TC2 and Hikey960, and could not see any regressions (all EAS functional tests are passing). [1] https://paste.debian.net/1100055/ [2] https://paste.debian.net/1100057/ [3] https://github.com/ARM-software/lisa/blob/master/lisa/tests/scheduler/eas_behaviour.py Signed-off-by: Quentin Perret <quentin.perret@arm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: juri.lelli@redhat.com Cc: morten.rasmussen@arm.com Cc: qais.yousef@arm.com Cc: qperret@qperret.net Cc: rjw@rjwysocki.net Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Link: https://lkml.kernel.org/r/20190912094404.13802-1-qperret@qperret.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-09-12 09:44:04 +00:00
return em_pd_energy(pd->em_pd, max_util, sum_util);
}
sched/fair: Select an energy-efficient CPU on task wake-up If an Energy Model (EM) is available and if the system isn't overutilized, re-route waking tasks into an energy-aware placement algorithm. The selection of an energy-efficient CPU for a task is achieved by estimating the impact on system-level active energy resulting from the placement of the task on the CPU with the highest spare capacity in each performance domain. This strategy spreads tasks in a performance domain and avoids overly aggressive task packing. The best CPU energy-wise is then selected if it saves a large enough amount of energy with respect to prev_cpu. Although it has already shown significant benefits on some existing targets, this approach cannot scale to platforms with numerous CPUs. This is an attempt to do something useful as writing a fast heuristic that performs reasonably well on a broad spectrum of architectures isn't an easy task. As such, the scope of usability of the energy-aware wake-up path is restricted to systems with the SD_ASYM_CPUCAPACITY flag set, and where the EM isn't too complex. Signed-off-by: Quentin Perret <quentin.perret@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: adharmap@codeaurora.org Cc: chris.redpath@arm.com Cc: currojerez@riseup.net Cc: dietmar.eggemann@arm.com Cc: edubezval@gmail.com Cc: gregkh@linuxfoundation.org Cc: javi.merino@kernel.org Cc: joel@joelfernandes.org Cc: juri.lelli@redhat.com Cc: morten.rasmussen@arm.com Cc: patrick.bellasi@arm.com Cc: pkondeti@codeaurora.org Cc: rjw@rjwysocki.net Cc: skannan@codeaurora.org Cc: smuckle@google.com Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Cc: viresh.kumar@linaro.org Link: https://lkml.kernel.org/r/20181203095628.11858-15-quentin.perret@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-12-03 09:56:27 +00:00
/*
* find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
* waking task. find_energy_efficient_cpu() looks for the CPU with maximum
* spare capacity in each performance domain and uses it as a potential
* candidate to execute the task. Then, it uses the Energy Model to figure
* out which of the CPU candidates is the most energy-efficient.
*
* The rationale for this heuristic is as follows. In a performance domain,
* all the most energy efficient CPU candidates (according to the Energy
* Model) are those for which we'll request a low frequency. When there are
* several CPUs for which the frequency request will be the same, we don't
* have enough data to break the tie between them, because the Energy Model
* only includes active power costs. With this model, if we assume that
* frequency requests follow utilization (e.g. using schedutil), the CPU with
* the maximum spare capacity in a performance domain is guaranteed to be among
* the best candidates of the performance domain.
*
* In practice, it could be preferable from an energy standpoint to pack
* small tasks on a CPU in order to let other CPUs go in deeper idle states,
* but that could also hurt our chances to go cluster idle, and we have no
* ways to tell with the current Energy Model if this is actually a good
* idea or not. So, find_energy_efficient_cpu() basically favors
* cluster-packing, and spreading inside a cluster. That should at least be
* a good thing for latency, and this is consistent with the idea that most
* of the energy savings of EAS come from the asymmetry of the system, and
* not so much from breaking the tie between identical CPUs. That's also the
* reason why EAS is enabled in the topology code only for systems where
* SD_ASYM_CPUCAPACITY is set.
*
* NOTE: Forkees are not accepted in the energy-aware wake-up path because
* they don't have any useful utilization data yet and it's not possible to
* forecast their impact on energy consumption. Consequently, they will be
* placed by find_idlest_cpu() on the least loaded CPU, which might turn out
* to be energy-inefficient in some use-cases. The alternative would be to
* bias new tasks towards specific types of CPUs first, or to try to infer
* their util_avg from the parent task, but those heuristics could hurt
* other use-cases too. So, until someone finds a better way to solve this,
* let's keep things simple by re-using the existing slow path.
*/
static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
{
sched/fair: Speed-up energy-aware wake-ups EAS computes the energy impact of migrating a waking task when deciding on which CPU it should run. However, the current approach is known to have a high algorithmic complexity, which can result in prohibitively high wake-up latencies on systems with complex energy models, such as systems with per-CPU DVFS. On such systems, the algorithm complexity is in O(n^2) (ignoring the cost of searching for performance states in the EM) with 'n' the number of CPUs. To address this, re-factor the EAS wake-up path to compute the energy 'delta' (with and without the task) on a per-performance domain basis, rather than system-wide, which brings the complexity down to O(n). No functional changes intended. Test results ~~~~~~~~~~~~ * Setup: Tested on a Google Pixel 3, with a Snapdragon 845 (4+4 CPUs, A55/A75). Base kernel is 5.3-rc5 + Pixel3 specific patches. Android userspace, no graphics. * Test case: Run a periodic rt-app task, with 16ms period, ramping down from 70% to 10%, in 5% steps of 500 ms each (json avail. at [1]). Frequencies of all CPUs are pinned to max (using scaling_min_freq CPUFreq sysfs entries) to reduce variability. The time to run select_task_rq_fair() is measured using the function profiler (/sys/kernel/debug/tracing/trace_stat/function*). See the test script for more details [2]. Test 1: I hacked the DT to 'fake' per-CPU DVFS. That is, we end up with one CPUFreq policy per CPU (8 policies in total). Since all frequencies are pinned to max for the test, this should have no impact on the actual frequency selection, but it does in the EAS calculation. +---------------------------+----------------------------------+ | Without patch | With patch | +-----+-----+----------+----------+-----+-----------------+----------+ | CPU | Hit | Avg (us) | s^2 (us) | Hit | Avg (us) | s^2 (us) | |-----+-----+----------+----------+-----+-----------------+----------+ | 0 | 274 | 38.303 | 1750.239 | 401 | 14.126 (-63.1%) | 146.625 | | 1 | 197 | 49.529 | 1695.852 | 314 | 16.135 (-67.4%) | 167.525 | | 2 | 142 | 34.296 | 1758.665 | 302 | 14.133 (-58.8%) | 130.071 | | 3 | 172 | 31.734 | 1490.975 | 641 | 14.637 (-53.9%) | 139.189 | | 4 | 316 | 7.834 | 178.217 | 425 | 5.413 (-30.9%) | 20.803 | | 5 | 447 | 8.424 | 144.638 | 556 | 5.929 (-29.6%) | 27.301 | | 6 | 581 | 14.886 | 346.793 | 456 | 5.711 (-61.6%) | 23.124 | | 7 | 456 | 10.005 | 211.187 | 997 | 4.708 (-52.9%) | 21.144 | +-----+-----+----------+----------+-----+-----------------+----------+ * Hit, Avg and s^2 are as reported by the function profiler Test 2: I also ran the same test with a normal DT, with 2 CPUFreq policies, to see if this causes regressions in the most common case. +---------------------------+----------------------------------+ | Without patch | With patch | +-----+-----+----------+----------+-----+-----------------+----------+ | CPU | Hit | Avg (us) | s^2 (us) | Hit | Avg (us) | s^2 (us) | |-----+-----+----------+----------+-----+-----------------+----------+ | 0 | 345 | 22.184 | 215.321 | 580 | 18.635 (-16.0%) | 146.892 | | 1 | 358 | 18.597 | 200.596 | 438 | 12.934 (-30.5%) | 104.604 | | 2 | 359 | 25.566 | 200.217 | 397 | 10.826 (-57.7%) | 74.021 | | 3 | 362 | 16.881 | 200.291 | 718 | 11.455 (-32.1%) | 102.280 | | 4 | 457 | 3.822 | 9.895 | 757 | 4.616 (+20.8%) | 13.369 | | 5 | 344 | 4.301 | 7.121 | 594 | 5.320 (+23.7%) | 18.798 | | 6 | 472 | 4.326 | 7.849 | 464 | 5.648 (+30.6%) | 22.022 | | 7 | 331 | 4.630 | 13.937 | 408 | 5.299 (+14.4%) | 18.273 | +-----+-----+----------+----------+-----+-----------------+----------+ * Hit, Avg and s^2 are as reported by the function profiler In addition to these two tests, I also ran 50 iterations of the Lisa EAS functional test suite [3] with this patch applied on Arm Juno r0, Arm Juno r2, Arm TC2 and Hikey960, and could not see any regressions (all EAS functional tests are passing). [1] https://paste.debian.net/1100055/ [2] https://paste.debian.net/1100057/ [3] https://github.com/ARM-software/lisa/blob/master/lisa/tests/scheduler/eas_behaviour.py Signed-off-by: Quentin Perret <quentin.perret@arm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: juri.lelli@redhat.com Cc: morten.rasmussen@arm.com Cc: qais.yousef@arm.com Cc: qperret@qperret.net Cc: rjw@rjwysocki.net Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Link: https://lkml.kernel.org/r/20190912094404.13802-1-qperret@qperret.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-09-12 09:44:04 +00:00
unsigned long prev_delta = ULONG_MAX, best_delta = ULONG_MAX;
sched/fair: Select an energy-efficient CPU on task wake-up If an Energy Model (EM) is available and if the system isn't overutilized, re-route waking tasks into an energy-aware placement algorithm. The selection of an energy-efficient CPU for a task is achieved by estimating the impact on system-level active energy resulting from the placement of the task on the CPU with the highest spare capacity in each performance domain. This strategy spreads tasks in a performance domain and avoids overly aggressive task packing. The best CPU energy-wise is then selected if it saves a large enough amount of energy with respect to prev_cpu. Although it has already shown significant benefits on some existing targets, this approach cannot scale to platforms with numerous CPUs. This is an attempt to do something useful as writing a fast heuristic that performs reasonably well on a broad spectrum of architectures isn't an easy task. As such, the scope of usability of the energy-aware wake-up path is restricted to systems with the SD_ASYM_CPUCAPACITY flag set, and where the EM isn't too complex. Signed-off-by: Quentin Perret <quentin.perret@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: adharmap@codeaurora.org Cc: chris.redpath@arm.com Cc: currojerez@riseup.net Cc: dietmar.eggemann@arm.com Cc: edubezval@gmail.com Cc: gregkh@linuxfoundation.org Cc: javi.merino@kernel.org Cc: joel@joelfernandes.org Cc: juri.lelli@redhat.com Cc: morten.rasmussen@arm.com Cc: patrick.bellasi@arm.com Cc: pkondeti@codeaurora.org Cc: rjw@rjwysocki.net Cc: skannan@codeaurora.org Cc: smuckle@google.com Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Cc: viresh.kumar@linaro.org Link: https://lkml.kernel.org/r/20181203095628.11858-15-quentin.perret@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-12-03 09:56:27 +00:00
struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
sched/fair: Speed-up energy-aware wake-ups EAS computes the energy impact of migrating a waking task when deciding on which CPU it should run. However, the current approach is known to have a high algorithmic complexity, which can result in prohibitively high wake-up latencies on systems with complex energy models, such as systems with per-CPU DVFS. On such systems, the algorithm complexity is in O(n^2) (ignoring the cost of searching for performance states in the EM) with 'n' the number of CPUs. To address this, re-factor the EAS wake-up path to compute the energy 'delta' (with and without the task) on a per-performance domain basis, rather than system-wide, which brings the complexity down to O(n). No functional changes intended. Test results ~~~~~~~~~~~~ * Setup: Tested on a Google Pixel 3, with a Snapdragon 845 (4+4 CPUs, A55/A75). Base kernel is 5.3-rc5 + Pixel3 specific patches. Android userspace, no graphics. * Test case: Run a periodic rt-app task, with 16ms period, ramping down from 70% to 10%, in 5% steps of 500 ms each (json avail. at [1]). Frequencies of all CPUs are pinned to max (using scaling_min_freq CPUFreq sysfs entries) to reduce variability. The time to run select_task_rq_fair() is measured using the function profiler (/sys/kernel/debug/tracing/trace_stat/function*). See the test script for more details [2]. Test 1: I hacked the DT to 'fake' per-CPU DVFS. That is, we end up with one CPUFreq policy per CPU (8 policies in total). Since all frequencies are pinned to max for the test, this should have no impact on the actual frequency selection, but it does in the EAS calculation. +---------------------------+----------------------------------+ | Without patch | With patch | +-----+-----+----------+----------+-----+-----------------+----------+ | CPU | Hit | Avg (us) | s^2 (us) | Hit | Avg (us) | s^2 (us) | |-----+-----+----------+----------+-----+-----------------+----------+ | 0 | 274 | 38.303 | 1750.239 | 401 | 14.126 (-63.1%) | 146.625 | | 1 | 197 | 49.529 | 1695.852 | 314 | 16.135 (-67.4%) | 167.525 | | 2 | 142 | 34.296 | 1758.665 | 302 | 14.133 (-58.8%) | 130.071 | | 3 | 172 | 31.734 | 1490.975 | 641 | 14.637 (-53.9%) | 139.189 | | 4 | 316 | 7.834 | 178.217 | 425 | 5.413 (-30.9%) | 20.803 | | 5 | 447 | 8.424 | 144.638 | 556 | 5.929 (-29.6%) | 27.301 | | 6 | 581 | 14.886 | 346.793 | 456 | 5.711 (-61.6%) | 23.124 | | 7 | 456 | 10.005 | 211.187 | 997 | 4.708 (-52.9%) | 21.144 | +-----+-----+----------+----------+-----+-----------------+----------+ * Hit, Avg and s^2 are as reported by the function profiler Test 2: I also ran the same test with a normal DT, with 2 CPUFreq policies, to see if this causes regressions in the most common case. +---------------------------+----------------------------------+ | Without patch | With patch | +-----+-----+----------+----------+-----+-----------------+----------+ | CPU | Hit | Avg (us) | s^2 (us) | Hit | Avg (us) | s^2 (us) | |-----+-----+----------+----------+-----+-----------------+----------+ | 0 | 345 | 22.184 | 215.321 | 580 | 18.635 (-16.0%) | 146.892 | | 1 | 358 | 18.597 | 200.596 | 438 | 12.934 (-30.5%) | 104.604 | | 2 | 359 | 25.566 | 200.217 | 397 | 10.826 (-57.7%) | 74.021 | | 3 | 362 | 16.881 | 200.291 | 718 | 11.455 (-32.1%) | 102.280 | | 4 | 457 | 3.822 | 9.895 | 757 | 4.616 (+20.8%) | 13.369 | | 5 | 344 | 4.301 | 7.121 | 594 | 5.320 (+23.7%) | 18.798 | | 6 | 472 | 4.326 | 7.849 | 464 | 5.648 (+30.6%) | 22.022 | | 7 | 331 | 4.630 | 13.937 | 408 | 5.299 (+14.4%) | 18.273 | +-----+-----+----------+----------+-----+-----------------+----------+ * Hit, Avg and s^2 are as reported by the function profiler In addition to these two tests, I also ran 50 iterations of the Lisa EAS functional test suite [3] with this patch applied on Arm Juno r0, Arm Juno r2, Arm TC2 and Hikey960, and could not see any regressions (all EAS functional tests are passing). [1] https://paste.debian.net/1100055/ [2] https://paste.debian.net/1100057/ [3] https://github.com/ARM-software/lisa/blob/master/lisa/tests/scheduler/eas_behaviour.py Signed-off-by: Quentin Perret <quentin.perret@arm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: juri.lelli@redhat.com Cc: morten.rasmussen@arm.com Cc: qais.yousef@arm.com Cc: qperret@qperret.net Cc: rjw@rjwysocki.net Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Link: https://lkml.kernel.org/r/20190912094404.13802-1-qperret@qperret.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-09-12 09:44:04 +00:00
unsigned long cpu_cap, util, base_energy = 0;
sched/fair: Select an energy-efficient CPU on task wake-up If an Energy Model (EM) is available and if the system isn't overutilized, re-route waking tasks into an energy-aware placement algorithm. The selection of an energy-efficient CPU for a task is achieved by estimating the impact on system-level active energy resulting from the placement of the task on the CPU with the highest spare capacity in each performance domain. This strategy spreads tasks in a performance domain and avoids overly aggressive task packing. The best CPU energy-wise is then selected if it saves a large enough amount of energy with respect to prev_cpu. Although it has already shown significant benefits on some existing targets, this approach cannot scale to platforms with numerous CPUs. This is an attempt to do something useful as writing a fast heuristic that performs reasonably well on a broad spectrum of architectures isn't an easy task. As such, the scope of usability of the energy-aware wake-up path is restricted to systems with the SD_ASYM_CPUCAPACITY flag set, and where the EM isn't too complex. Signed-off-by: Quentin Perret <quentin.perret@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: adharmap@codeaurora.org Cc: chris.redpath@arm.com Cc: currojerez@riseup.net Cc: dietmar.eggemann@arm.com Cc: edubezval@gmail.com Cc: gregkh@linuxfoundation.org Cc: javi.merino@kernel.org Cc: joel@joelfernandes.org Cc: juri.lelli@redhat.com Cc: morten.rasmussen@arm.com Cc: patrick.bellasi@arm.com Cc: pkondeti@codeaurora.org Cc: rjw@rjwysocki.net Cc: skannan@codeaurora.org Cc: smuckle@google.com Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Cc: viresh.kumar@linaro.org Link: https://lkml.kernel.org/r/20181203095628.11858-15-quentin.perret@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-12-03 09:56:27 +00:00
int cpu, best_energy_cpu = prev_cpu;
struct sched_domain *sd;
sched/fair: Speed-up energy-aware wake-ups EAS computes the energy impact of migrating a waking task when deciding on which CPU it should run. However, the current approach is known to have a high algorithmic complexity, which can result in prohibitively high wake-up latencies on systems with complex energy models, such as systems with per-CPU DVFS. On such systems, the algorithm complexity is in O(n^2) (ignoring the cost of searching for performance states in the EM) with 'n' the number of CPUs. To address this, re-factor the EAS wake-up path to compute the energy 'delta' (with and without the task) on a per-performance domain basis, rather than system-wide, which brings the complexity down to O(n). No functional changes intended. Test results ~~~~~~~~~~~~ * Setup: Tested on a Google Pixel 3, with a Snapdragon 845 (4+4 CPUs, A55/A75). Base kernel is 5.3-rc5 + Pixel3 specific patches. Android userspace, no graphics. * Test case: Run a periodic rt-app task, with 16ms period, ramping down from 70% to 10%, in 5% steps of 500 ms each (json avail. at [1]). Frequencies of all CPUs are pinned to max (using scaling_min_freq CPUFreq sysfs entries) to reduce variability. The time to run select_task_rq_fair() is measured using the function profiler (/sys/kernel/debug/tracing/trace_stat/function*). See the test script for more details [2]. Test 1: I hacked the DT to 'fake' per-CPU DVFS. That is, we end up with one CPUFreq policy per CPU (8 policies in total). Since all frequencies are pinned to max for the test, this should have no impact on the actual frequency selection, but it does in the EAS calculation. +---------------------------+----------------------------------+ | Without patch | With patch | +-----+-----+----------+----------+-----+-----------------+----------+ | CPU | Hit | Avg (us) | s^2 (us) | Hit | Avg (us) | s^2 (us) | |-----+-----+----------+----------+-----+-----------------+----------+ | 0 | 274 | 38.303 | 1750.239 | 401 | 14.126 (-63.1%) | 146.625 | | 1 | 197 | 49.529 | 1695.852 | 314 | 16.135 (-67.4%) | 167.525 | | 2 | 142 | 34.296 | 1758.665 | 302 | 14.133 (-58.8%) | 130.071 | | 3 | 172 | 31.734 | 1490.975 | 641 | 14.637 (-53.9%) | 139.189 | | 4 | 316 | 7.834 | 178.217 | 425 | 5.413 (-30.9%) | 20.803 | | 5 | 447 | 8.424 | 144.638 | 556 | 5.929 (-29.6%) | 27.301 | | 6 | 581 | 14.886 | 346.793 | 456 | 5.711 (-61.6%) | 23.124 | | 7 | 456 | 10.005 | 211.187 | 997 | 4.708 (-52.9%) | 21.144 | +-----+-----+----------+----------+-----+-----------------+----------+ * Hit, Avg and s^2 are as reported by the function profiler Test 2: I also ran the same test with a normal DT, with 2 CPUFreq policies, to see if this causes regressions in the most common case. +---------------------------+----------------------------------+ | Without patch | With patch | +-----+-----+----------+----------+-----+-----------------+----------+ | CPU | Hit | Avg (us) | s^2 (us) | Hit | Avg (us) | s^2 (us) | |-----+-----+----------+----------+-----+-----------------+----------+ | 0 | 345 | 22.184 | 215.321 | 580 | 18.635 (-16.0%) | 146.892 | | 1 | 358 | 18.597 | 200.596 | 438 | 12.934 (-30.5%) | 104.604 | | 2 | 359 | 25.566 | 200.217 | 397 | 10.826 (-57.7%) | 74.021 | | 3 | 362 | 16.881 | 200.291 | 718 | 11.455 (-32.1%) | 102.280 | | 4 | 457 | 3.822 | 9.895 | 757 | 4.616 (+20.8%) | 13.369 | | 5 | 344 | 4.301 | 7.121 | 594 | 5.320 (+23.7%) | 18.798 | | 6 | 472 | 4.326 | 7.849 | 464 | 5.648 (+30.6%) | 22.022 | | 7 | 331 | 4.630 | 13.937 | 408 | 5.299 (+14.4%) | 18.273 | +-----+-----+----------+----------+-----+-----------------+----------+ * Hit, Avg and s^2 are as reported by the function profiler In addition to these two tests, I also ran 50 iterations of the Lisa EAS functional test suite [3] with this patch applied on Arm Juno r0, Arm Juno r2, Arm TC2 and Hikey960, and could not see any regressions (all EAS functional tests are passing). [1] https://paste.debian.net/1100055/ [2] https://paste.debian.net/1100057/ [3] https://github.com/ARM-software/lisa/blob/master/lisa/tests/scheduler/eas_behaviour.py Signed-off-by: Quentin Perret <quentin.perret@arm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: juri.lelli@redhat.com Cc: morten.rasmussen@arm.com Cc: qais.yousef@arm.com Cc: qperret@qperret.net Cc: rjw@rjwysocki.net Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Link: https://lkml.kernel.org/r/20190912094404.13802-1-qperret@qperret.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-09-12 09:44:04 +00:00
struct perf_domain *pd;
sched/fair: Select an energy-efficient CPU on task wake-up If an Energy Model (EM) is available and if the system isn't overutilized, re-route waking tasks into an energy-aware placement algorithm. The selection of an energy-efficient CPU for a task is achieved by estimating the impact on system-level active energy resulting from the placement of the task on the CPU with the highest spare capacity in each performance domain. This strategy spreads tasks in a performance domain and avoids overly aggressive task packing. The best CPU energy-wise is then selected if it saves a large enough amount of energy with respect to prev_cpu. Although it has already shown significant benefits on some existing targets, this approach cannot scale to platforms with numerous CPUs. This is an attempt to do something useful as writing a fast heuristic that performs reasonably well on a broad spectrum of architectures isn't an easy task. As such, the scope of usability of the energy-aware wake-up path is restricted to systems with the SD_ASYM_CPUCAPACITY flag set, and where the EM isn't too complex. Signed-off-by: Quentin Perret <quentin.perret@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: adharmap@codeaurora.org Cc: chris.redpath@arm.com Cc: currojerez@riseup.net Cc: dietmar.eggemann@arm.com Cc: edubezval@gmail.com Cc: gregkh@linuxfoundation.org Cc: javi.merino@kernel.org Cc: joel@joelfernandes.org Cc: juri.lelli@redhat.com Cc: morten.rasmussen@arm.com Cc: patrick.bellasi@arm.com Cc: pkondeti@codeaurora.org Cc: rjw@rjwysocki.net Cc: skannan@codeaurora.org Cc: smuckle@google.com Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Cc: viresh.kumar@linaro.org Link: https://lkml.kernel.org/r/20181203095628.11858-15-quentin.perret@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-12-03 09:56:27 +00:00
rcu_read_lock();
pd = rcu_dereference(rd->pd);
if (!pd || READ_ONCE(rd->overutilized))
goto fail;
/*
* Energy-aware wake-up happens on the lowest sched_domain starting
* from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
*/
sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity));
while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
sd = sd->parent;
if (!sd)
goto fail;
sync_entity_load_avg(&p->se);
if (!task_util_est(p))
goto unlock;
for (; pd; pd = pd->next) {
sched/fair: Speed-up energy-aware wake-ups EAS computes the energy impact of migrating a waking task when deciding on which CPU it should run. However, the current approach is known to have a high algorithmic complexity, which can result in prohibitively high wake-up latencies on systems with complex energy models, such as systems with per-CPU DVFS. On such systems, the algorithm complexity is in O(n^2) (ignoring the cost of searching for performance states in the EM) with 'n' the number of CPUs. To address this, re-factor the EAS wake-up path to compute the energy 'delta' (with and without the task) on a per-performance domain basis, rather than system-wide, which brings the complexity down to O(n). No functional changes intended. Test results ~~~~~~~~~~~~ * Setup: Tested on a Google Pixel 3, with a Snapdragon 845 (4+4 CPUs, A55/A75). Base kernel is 5.3-rc5 + Pixel3 specific patches. Android userspace, no graphics. * Test case: Run a periodic rt-app task, with 16ms period, ramping down from 70% to 10%, in 5% steps of 500 ms each (json avail. at [1]). Frequencies of all CPUs are pinned to max (using scaling_min_freq CPUFreq sysfs entries) to reduce variability. The time to run select_task_rq_fair() is measured using the function profiler (/sys/kernel/debug/tracing/trace_stat/function*). See the test script for more details [2]. Test 1: I hacked the DT to 'fake' per-CPU DVFS. That is, we end up with one CPUFreq policy per CPU (8 policies in total). Since all frequencies are pinned to max for the test, this should have no impact on the actual frequency selection, but it does in the EAS calculation. +---------------------------+----------------------------------+ | Without patch | With patch | +-----+-----+----------+----------+-----+-----------------+----------+ | CPU | Hit | Avg (us) | s^2 (us) | Hit | Avg (us) | s^2 (us) | |-----+-----+----------+----------+-----+-----------------+----------+ | 0 | 274 | 38.303 | 1750.239 | 401 | 14.126 (-63.1%) | 146.625 | | 1 | 197 | 49.529 | 1695.852 | 314 | 16.135 (-67.4%) | 167.525 | | 2 | 142 | 34.296 | 1758.665 | 302 | 14.133 (-58.8%) | 130.071 | | 3 | 172 | 31.734 | 1490.975 | 641 | 14.637 (-53.9%) | 139.189 | | 4 | 316 | 7.834 | 178.217 | 425 | 5.413 (-30.9%) | 20.803 | | 5 | 447 | 8.424 | 144.638 | 556 | 5.929 (-29.6%) | 27.301 | | 6 | 581 | 14.886 | 346.793 | 456 | 5.711 (-61.6%) | 23.124 | | 7 | 456 | 10.005 | 211.187 | 997 | 4.708 (-52.9%) | 21.144 | +-----+-----+----------+----------+-----+-----------------+----------+ * Hit, Avg and s^2 are as reported by the function profiler Test 2: I also ran the same test with a normal DT, with 2 CPUFreq policies, to see if this causes regressions in the most common case. +---------------------------+----------------------------------+ | Without patch | With patch | +-----+-----+----------+----------+-----+-----------------+----------+ | CPU | Hit | Avg (us) | s^2 (us) | Hit | Avg (us) | s^2 (us) | |-----+-----+----------+----------+-----+-----------------+----------+ | 0 | 345 | 22.184 | 215.321 | 580 | 18.635 (-16.0%) | 146.892 | | 1 | 358 | 18.597 | 200.596 | 438 | 12.934 (-30.5%) | 104.604 | | 2 | 359 | 25.566 | 200.217 | 397 | 10.826 (-57.7%) | 74.021 | | 3 | 362 | 16.881 | 200.291 | 718 | 11.455 (-32.1%) | 102.280 | | 4 | 457 | 3.822 | 9.895 | 757 | 4.616 (+20.8%) | 13.369 | | 5 | 344 | 4.301 | 7.121 | 594 | 5.320 (+23.7%) | 18.798 | | 6 | 472 | 4.326 | 7.849 | 464 | 5.648 (+30.6%) | 22.022 | | 7 | 331 | 4.630 | 13.937 | 408 | 5.299 (+14.4%) | 18.273 | +-----+-----+----------+----------+-----+-----------------+----------+ * Hit, Avg and s^2 are as reported by the function profiler In addition to these two tests, I also ran 50 iterations of the Lisa EAS functional test suite [3] with this patch applied on Arm Juno r0, Arm Juno r2, Arm TC2 and Hikey960, and could not see any regressions (all EAS functional tests are passing). [1] https://paste.debian.net/1100055/ [2] https://paste.debian.net/1100057/ [3] https://github.com/ARM-software/lisa/blob/master/lisa/tests/scheduler/eas_behaviour.py Signed-off-by: Quentin Perret <quentin.perret@arm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: juri.lelli@redhat.com Cc: morten.rasmussen@arm.com Cc: qais.yousef@arm.com Cc: qperret@qperret.net Cc: rjw@rjwysocki.net Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Link: https://lkml.kernel.org/r/20190912094404.13802-1-qperret@qperret.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-09-12 09:44:04 +00:00
unsigned long cur_delta, spare_cap, max_spare_cap = 0;
unsigned long base_energy_pd;
sched/fair: Select an energy-efficient CPU on task wake-up If an Energy Model (EM) is available and if the system isn't overutilized, re-route waking tasks into an energy-aware placement algorithm. The selection of an energy-efficient CPU for a task is achieved by estimating the impact on system-level active energy resulting from the placement of the task on the CPU with the highest spare capacity in each performance domain. This strategy spreads tasks in a performance domain and avoids overly aggressive task packing. The best CPU energy-wise is then selected if it saves a large enough amount of energy with respect to prev_cpu. Although it has already shown significant benefits on some existing targets, this approach cannot scale to platforms with numerous CPUs. This is an attempt to do something useful as writing a fast heuristic that performs reasonably well on a broad spectrum of architectures isn't an easy task. As such, the scope of usability of the energy-aware wake-up path is restricted to systems with the SD_ASYM_CPUCAPACITY flag set, and where the EM isn't too complex. Signed-off-by: Quentin Perret <quentin.perret@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: adharmap@codeaurora.org Cc: chris.redpath@arm.com Cc: currojerez@riseup.net Cc: dietmar.eggemann@arm.com Cc: edubezval@gmail.com Cc: gregkh@linuxfoundation.org Cc: javi.merino@kernel.org Cc: joel@joelfernandes.org Cc: juri.lelli@redhat.com Cc: morten.rasmussen@arm.com Cc: patrick.bellasi@arm.com Cc: pkondeti@codeaurora.org Cc: rjw@rjwysocki.net Cc: skannan@codeaurora.org Cc: smuckle@google.com Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Cc: viresh.kumar@linaro.org Link: https://lkml.kernel.org/r/20181203095628.11858-15-quentin.perret@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-12-03 09:56:27 +00:00
int max_spare_cap_cpu = -1;
sched/fair: Speed-up energy-aware wake-ups EAS computes the energy impact of migrating a waking task when deciding on which CPU it should run. However, the current approach is known to have a high algorithmic complexity, which can result in prohibitively high wake-up latencies on systems with complex energy models, such as systems with per-CPU DVFS. On such systems, the algorithm complexity is in O(n^2) (ignoring the cost of searching for performance states in the EM) with 'n' the number of CPUs. To address this, re-factor the EAS wake-up path to compute the energy 'delta' (with and without the task) on a per-performance domain basis, rather than system-wide, which brings the complexity down to O(n). No functional changes intended. Test results ~~~~~~~~~~~~ * Setup: Tested on a Google Pixel 3, with a Snapdragon 845 (4+4 CPUs, A55/A75). Base kernel is 5.3-rc5 + Pixel3 specific patches. Android userspace, no graphics. * Test case: Run a periodic rt-app task, with 16ms period, ramping down from 70% to 10%, in 5% steps of 500 ms each (json avail. at [1]). Frequencies of all CPUs are pinned to max (using scaling_min_freq CPUFreq sysfs entries) to reduce variability. The time to run select_task_rq_fair() is measured using the function profiler (/sys/kernel/debug/tracing/trace_stat/function*). See the test script for more details [2]. Test 1: I hacked the DT to 'fake' per-CPU DVFS. That is, we end up with one CPUFreq policy per CPU (8 policies in total). Since all frequencies are pinned to max for the test, this should have no impact on the actual frequency selection, but it does in the EAS calculation. +---------------------------+----------------------------------+ | Without patch | With patch | +-----+-----+----------+----------+-----+-----------------+----------+ | CPU | Hit | Avg (us) | s^2 (us) | Hit | Avg (us) | s^2 (us) | |-----+-----+----------+----------+-----+-----------------+----------+ | 0 | 274 | 38.303 | 1750.239 | 401 | 14.126 (-63.1%) | 146.625 | | 1 | 197 | 49.529 | 1695.852 | 314 | 16.135 (-67.4%) | 167.525 | | 2 | 142 | 34.296 | 1758.665 | 302 | 14.133 (-58.8%) | 130.071 | | 3 | 172 | 31.734 | 1490.975 | 641 | 14.637 (-53.9%) | 139.189 | | 4 | 316 | 7.834 | 178.217 | 425 | 5.413 (-30.9%) | 20.803 | | 5 | 447 | 8.424 | 144.638 | 556 | 5.929 (-29.6%) | 27.301 | | 6 | 581 | 14.886 | 346.793 | 456 | 5.711 (-61.6%) | 23.124 | | 7 | 456 | 10.005 | 211.187 | 997 | 4.708 (-52.9%) | 21.144 | +-----+-----+----------+----------+-----+-----------------+----------+ * Hit, Avg and s^2 are as reported by the function profiler Test 2: I also ran the same test with a normal DT, with 2 CPUFreq policies, to see if this causes regressions in the most common case. +---------------------------+----------------------------------+ | Without patch | With patch | +-----+-----+----------+----------+-----+-----------------+----------+ | CPU | Hit | Avg (us) | s^2 (us) | Hit | Avg (us) | s^2 (us) | |-----+-----+----------+----------+-----+-----------------+----------+ | 0 | 345 | 22.184 | 215.321 | 580 | 18.635 (-16.0%) | 146.892 | | 1 | 358 | 18.597 | 200.596 | 438 | 12.934 (-30.5%) | 104.604 | | 2 | 359 | 25.566 | 200.217 | 397 | 10.826 (-57.7%) | 74.021 | | 3 | 362 | 16.881 | 200.291 | 718 | 11.455 (-32.1%) | 102.280 | | 4 | 457 | 3.822 | 9.895 | 757 | 4.616 (+20.8%) | 13.369 | | 5 | 344 | 4.301 | 7.121 | 594 | 5.320 (+23.7%) | 18.798 | | 6 | 472 | 4.326 | 7.849 | 464 | 5.648 (+30.6%) | 22.022 | | 7 | 331 | 4.630 | 13.937 | 408 | 5.299 (+14.4%) | 18.273 | +-----+-----+----------+----------+-----+-----------------+----------+ * Hit, Avg and s^2 are as reported by the function profiler In addition to these two tests, I also ran 50 iterations of the Lisa EAS functional test suite [3] with this patch applied on Arm Juno r0, Arm Juno r2, Arm TC2 and Hikey960, and could not see any regressions (all EAS functional tests are passing). [1] https://paste.debian.net/1100055/ [2] https://paste.debian.net/1100057/ [3] https://github.com/ARM-software/lisa/blob/master/lisa/tests/scheduler/eas_behaviour.py Signed-off-by: Quentin Perret <quentin.perret@arm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: juri.lelli@redhat.com Cc: morten.rasmussen@arm.com Cc: qais.yousef@arm.com Cc: qperret@qperret.net Cc: rjw@rjwysocki.net Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Link: https://lkml.kernel.org/r/20190912094404.13802-1-qperret@qperret.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-09-12 09:44:04 +00:00
/* Compute the 'base' energy of the pd, without @p */
base_energy_pd = compute_energy(p, -1, pd);
base_energy += base_energy_pd;
sched/fair: Select an energy-efficient CPU on task wake-up If an Energy Model (EM) is available and if the system isn't overutilized, re-route waking tasks into an energy-aware placement algorithm. The selection of an energy-efficient CPU for a task is achieved by estimating the impact on system-level active energy resulting from the placement of the task on the CPU with the highest spare capacity in each performance domain. This strategy spreads tasks in a performance domain and avoids overly aggressive task packing. The best CPU energy-wise is then selected if it saves a large enough amount of energy with respect to prev_cpu. Although it has already shown significant benefits on some existing targets, this approach cannot scale to platforms with numerous CPUs. This is an attempt to do something useful as writing a fast heuristic that performs reasonably well on a broad spectrum of architectures isn't an easy task. As such, the scope of usability of the energy-aware wake-up path is restricted to systems with the SD_ASYM_CPUCAPACITY flag set, and where the EM isn't too complex. Signed-off-by: Quentin Perret <quentin.perret@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: adharmap@codeaurora.org Cc: chris.redpath@arm.com Cc: currojerez@riseup.net Cc: dietmar.eggemann@arm.com Cc: edubezval@gmail.com Cc: gregkh@linuxfoundation.org Cc: javi.merino@kernel.org Cc: joel@joelfernandes.org Cc: juri.lelli@redhat.com Cc: morten.rasmussen@arm.com Cc: patrick.bellasi@arm.com Cc: pkondeti@codeaurora.org Cc: rjw@rjwysocki.net Cc: skannan@codeaurora.org Cc: smuckle@google.com Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Cc: viresh.kumar@linaro.org Link: https://lkml.kernel.org/r/20181203095628.11858-15-quentin.perret@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-12-03 09:56:27 +00:00
for_each_cpu_and(cpu, perf_domain_span(pd), sched_domain_span(sd)) {
if (!cpumask_test_cpu(cpu, p->cpus_ptr))
sched/fair: Select an energy-efficient CPU on task wake-up If an Energy Model (EM) is available and if the system isn't overutilized, re-route waking tasks into an energy-aware placement algorithm. The selection of an energy-efficient CPU for a task is achieved by estimating the impact on system-level active energy resulting from the placement of the task on the CPU with the highest spare capacity in each performance domain. This strategy spreads tasks in a performance domain and avoids overly aggressive task packing. The best CPU energy-wise is then selected if it saves a large enough amount of energy with respect to prev_cpu. Although it has already shown significant benefits on some existing targets, this approach cannot scale to platforms with numerous CPUs. This is an attempt to do something useful as writing a fast heuristic that performs reasonably well on a broad spectrum of architectures isn't an easy task. As such, the scope of usability of the energy-aware wake-up path is restricted to systems with the SD_ASYM_CPUCAPACITY flag set, and where the EM isn't too complex. Signed-off-by: Quentin Perret <quentin.perret@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: adharmap@codeaurora.org Cc: chris.redpath@arm.com Cc: currojerez@riseup.net Cc: dietmar.eggemann@arm.com Cc: edubezval@gmail.com Cc: gregkh@linuxfoundation.org Cc: javi.merino@kernel.org Cc: joel@joelfernandes.org Cc: juri.lelli@redhat.com Cc: morten.rasmussen@arm.com Cc: patrick.bellasi@arm.com Cc: pkondeti@codeaurora.org Cc: rjw@rjwysocki.net Cc: skannan@codeaurora.org Cc: smuckle@google.com Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Cc: viresh.kumar@linaro.org Link: https://lkml.kernel.org/r/20181203095628.11858-15-quentin.perret@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-12-03 09:56:27 +00:00
continue;
util = cpu_util_next(cpu, p, cpu);
cpu_cap = capacity_of(cpu);
sched/fair: Make EAS wakeup placement consider uclamp restrictions task_fits_capacity() has just been made uclamp-aware, and find_energy_efficient_cpu() needs to go through the same treatment. Things are somewhat different here however - using the task max clamp isn't sufficient. Consider the following setup: The target runqueue, rq: rq.cpu_capacity_orig = 512 rq.cfs.avg.util_avg = 200 rq.uclamp.max = 768 // the max p.uclamp.max of all enqueued p's is 768 The waking task, p (not yet enqueued on rq): p.util_est = 600 p.uclamp.max = 100 Now, consider the following code which doesn't use the rq clamps: util = uclamp_task_util(p); // Does the task fit in the spare CPU capacity? cpu = cpu_of(rq); fits_capacity(util, cpu_capacity(cpu) - cpu_util(cpu)) This would lead to: util = 100; fits_capacity(100, 512 - 200) fits_capacity() would return true. However, enqueuing p on that CPU *will* cause it to become overutilized since rq clamp values are max-aggregated, so we'd remain with rq.uclamp.max = 768 which comes from the other tasks already enqueued on rq. Thus, we could select a high enough frequency to reach beyond 0.8 * 512 utilization (== overutilized) after enqueuing p on rq. What find_energy_efficient_cpu() needs here is uclamp_rq_util_with() which lets us peek at the future utilization landscape, including rq-wide uclamp values. Make find_energy_efficient_cpu() use uclamp_rq_util_with() for its fits_capacity() check. This is in line with what compute_energy() ends up using for estimating utilization. Tested-By: Dietmar Eggemann <dietmar.eggemann@arm.com> Suggested-by: Quentin Perret <qperret@google.com> Signed-off-by: Valentin Schneider <valentin.schneider@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: https://lkml.kernel.org/r/20191211113851.24241-6-valentin.schneider@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-12-11 11:38:51 +00:00
spare_cap = cpu_cap - util;
/*
* Skip CPUs that cannot satisfy the capacity request.
* IOW, placing the task there would make the CPU
* overutilized. Take uclamp into account to see how
* much capacity we can get out of the CPU; this is
* aligned with schedutil_cpu_util().
*/
util = uclamp_rq_util_with(cpu_rq(cpu), util, p);
if (!fits_capacity(util, cpu_cap))
sched/fair: Select an energy-efficient CPU on task wake-up If an Energy Model (EM) is available and if the system isn't overutilized, re-route waking tasks into an energy-aware placement algorithm. The selection of an energy-efficient CPU for a task is achieved by estimating the impact on system-level active energy resulting from the placement of the task on the CPU with the highest spare capacity in each performance domain. This strategy spreads tasks in a performance domain and avoids overly aggressive task packing. The best CPU energy-wise is then selected if it saves a large enough amount of energy with respect to prev_cpu. Although it has already shown significant benefits on some existing targets, this approach cannot scale to platforms with numerous CPUs. This is an attempt to do something useful as writing a fast heuristic that performs reasonably well on a broad spectrum of architectures isn't an easy task. As such, the scope of usability of the energy-aware wake-up path is restricted to systems with the SD_ASYM_CPUCAPACITY flag set, and where the EM isn't too complex. Signed-off-by: Quentin Perret <quentin.perret@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: adharmap@codeaurora.org Cc: chris.redpath@arm.com Cc: currojerez@riseup.net Cc: dietmar.eggemann@arm.com Cc: edubezval@gmail.com Cc: gregkh@linuxfoundation.org Cc: javi.merino@kernel.org Cc: joel@joelfernandes.org Cc: juri.lelli@redhat.com Cc: morten.rasmussen@arm.com Cc: patrick.bellasi@arm.com Cc: pkondeti@codeaurora.org Cc: rjw@rjwysocki.net Cc: skannan@codeaurora.org Cc: smuckle@google.com Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Cc: viresh.kumar@linaro.org Link: https://lkml.kernel.org/r/20181203095628.11858-15-quentin.perret@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-12-03 09:56:27 +00:00
continue;
/* Always use prev_cpu as a candidate. */
if (cpu == prev_cpu) {
sched/fair: Speed-up energy-aware wake-ups EAS computes the energy impact of migrating a waking task when deciding on which CPU it should run. However, the current approach is known to have a high algorithmic complexity, which can result in prohibitively high wake-up latencies on systems with complex energy models, such as systems with per-CPU DVFS. On such systems, the algorithm complexity is in O(n^2) (ignoring the cost of searching for performance states in the EM) with 'n' the number of CPUs. To address this, re-factor the EAS wake-up path to compute the energy 'delta' (with and without the task) on a per-performance domain basis, rather than system-wide, which brings the complexity down to O(n). No functional changes intended. Test results ~~~~~~~~~~~~ * Setup: Tested on a Google Pixel 3, with a Snapdragon 845 (4+4 CPUs, A55/A75). Base kernel is 5.3-rc5 + Pixel3 specific patches. Android userspace, no graphics. * Test case: Run a periodic rt-app task, with 16ms period, ramping down from 70% to 10%, in 5% steps of 500 ms each (json avail. at [1]). Frequencies of all CPUs are pinned to max (using scaling_min_freq CPUFreq sysfs entries) to reduce variability. The time to run select_task_rq_fair() is measured using the function profiler (/sys/kernel/debug/tracing/trace_stat/function*). See the test script for more details [2]. Test 1: I hacked the DT to 'fake' per-CPU DVFS. That is, we end up with one CPUFreq policy per CPU (8 policies in total). Since all frequencies are pinned to max for the test, this should have no impact on the actual frequency selection, but it does in the EAS calculation. +---------------------------+----------------------------------+ | Without patch | With patch | +-----+-----+----------+----------+-----+-----------------+----------+ | CPU | Hit | Avg (us) | s^2 (us) | Hit | Avg (us) | s^2 (us) | |-----+-----+----------+----------+-----+-----------------+----------+ | 0 | 274 | 38.303 | 1750.239 | 401 | 14.126 (-63.1%) | 146.625 | | 1 | 197 | 49.529 | 1695.852 | 314 | 16.135 (-67.4%) | 167.525 | | 2 | 142 | 34.296 | 1758.665 | 302 | 14.133 (-58.8%) | 130.071 | | 3 | 172 | 31.734 | 1490.975 | 641 | 14.637 (-53.9%) | 139.189 | | 4 | 316 | 7.834 | 178.217 | 425 | 5.413 (-30.9%) | 20.803 | | 5 | 447 | 8.424 | 144.638 | 556 | 5.929 (-29.6%) | 27.301 | | 6 | 581 | 14.886 | 346.793 | 456 | 5.711 (-61.6%) | 23.124 | | 7 | 456 | 10.005 | 211.187 | 997 | 4.708 (-52.9%) | 21.144 | +-----+-----+----------+----------+-----+-----------------+----------+ * Hit, Avg and s^2 are as reported by the function profiler Test 2: I also ran the same test with a normal DT, with 2 CPUFreq policies, to see if this causes regressions in the most common case. +---------------------------+----------------------------------+ | Without patch | With patch | +-----+-----+----------+----------+-----+-----------------+----------+ | CPU | Hit | Avg (us) | s^2 (us) | Hit | Avg (us) | s^2 (us) | |-----+-----+----------+----------+-----+-----------------+----------+ | 0 | 345 | 22.184 | 215.321 | 580 | 18.635 (-16.0%) | 146.892 | | 1 | 358 | 18.597 | 200.596 | 438 | 12.934 (-30.5%) | 104.604 | | 2 | 359 | 25.566 | 200.217 | 397 | 10.826 (-57.7%) | 74.021 | | 3 | 362 | 16.881 | 200.291 | 718 | 11.455 (-32.1%) | 102.280 | | 4 | 457 | 3.822 | 9.895 | 757 | 4.616 (+20.8%) | 13.369 | | 5 | 344 | 4.301 | 7.121 | 594 | 5.320 (+23.7%) | 18.798 | | 6 | 472 | 4.326 | 7.849 | 464 | 5.648 (+30.6%) | 22.022 | | 7 | 331 | 4.630 | 13.937 | 408 | 5.299 (+14.4%) | 18.273 | +-----+-----+----------+----------+-----+-----------------+----------+ * Hit, Avg and s^2 are as reported by the function profiler In addition to these two tests, I also ran 50 iterations of the Lisa EAS functional test suite [3] with this patch applied on Arm Juno r0, Arm Juno r2, Arm TC2 and Hikey960, and could not see any regressions (all EAS functional tests are passing). [1] https://paste.debian.net/1100055/ [2] https://paste.debian.net/1100057/ [3] https://github.com/ARM-software/lisa/blob/master/lisa/tests/scheduler/eas_behaviour.py Signed-off-by: Quentin Perret <quentin.perret@arm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: juri.lelli@redhat.com Cc: morten.rasmussen@arm.com Cc: qais.yousef@arm.com Cc: qperret@qperret.net Cc: rjw@rjwysocki.net Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Link: https://lkml.kernel.org/r/20190912094404.13802-1-qperret@qperret.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-09-12 09:44:04 +00:00
prev_delta = compute_energy(p, prev_cpu, pd);
prev_delta -= base_energy_pd;
best_delta = min(best_delta, prev_delta);
sched/fair: Select an energy-efficient CPU on task wake-up If an Energy Model (EM) is available and if the system isn't overutilized, re-route waking tasks into an energy-aware placement algorithm. The selection of an energy-efficient CPU for a task is achieved by estimating the impact on system-level active energy resulting from the placement of the task on the CPU with the highest spare capacity in each performance domain. This strategy spreads tasks in a performance domain and avoids overly aggressive task packing. The best CPU energy-wise is then selected if it saves a large enough amount of energy with respect to prev_cpu. Although it has already shown significant benefits on some existing targets, this approach cannot scale to platforms with numerous CPUs. This is an attempt to do something useful as writing a fast heuristic that performs reasonably well on a broad spectrum of architectures isn't an easy task. As such, the scope of usability of the energy-aware wake-up path is restricted to systems with the SD_ASYM_CPUCAPACITY flag set, and where the EM isn't too complex. Signed-off-by: Quentin Perret <quentin.perret@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: adharmap@codeaurora.org Cc: chris.redpath@arm.com Cc: currojerez@riseup.net Cc: dietmar.eggemann@arm.com Cc: edubezval@gmail.com Cc: gregkh@linuxfoundation.org Cc: javi.merino@kernel.org Cc: joel@joelfernandes.org Cc: juri.lelli@redhat.com Cc: morten.rasmussen@arm.com Cc: patrick.bellasi@arm.com Cc: pkondeti@codeaurora.org Cc: rjw@rjwysocki.net Cc: skannan@codeaurora.org Cc: smuckle@google.com Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Cc: viresh.kumar@linaro.org Link: https://lkml.kernel.org/r/20181203095628.11858-15-quentin.perret@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-12-03 09:56:27 +00:00
}
/*
* Find the CPU with the maximum spare capacity in
* the performance domain
*/
if (spare_cap > max_spare_cap) {
max_spare_cap = spare_cap;
max_spare_cap_cpu = cpu;
}
}
/* Evaluate the energy impact of using this CPU. */
if (max_spare_cap_cpu >= 0 && max_spare_cap_cpu != prev_cpu) {
sched/fair: Speed-up energy-aware wake-ups EAS computes the energy impact of migrating a waking task when deciding on which CPU it should run. However, the current approach is known to have a high algorithmic complexity, which can result in prohibitively high wake-up latencies on systems with complex energy models, such as systems with per-CPU DVFS. On such systems, the algorithm complexity is in O(n^2) (ignoring the cost of searching for performance states in the EM) with 'n' the number of CPUs. To address this, re-factor the EAS wake-up path to compute the energy 'delta' (with and without the task) on a per-performance domain basis, rather than system-wide, which brings the complexity down to O(n). No functional changes intended. Test results ~~~~~~~~~~~~ * Setup: Tested on a Google Pixel 3, with a Snapdragon 845 (4+4 CPUs, A55/A75). Base kernel is 5.3-rc5 + Pixel3 specific patches. Android userspace, no graphics. * Test case: Run a periodic rt-app task, with 16ms period, ramping down from 70% to 10%, in 5% steps of 500 ms each (json avail. at [1]). Frequencies of all CPUs are pinned to max (using scaling_min_freq CPUFreq sysfs entries) to reduce variability. The time to run select_task_rq_fair() is measured using the function profiler (/sys/kernel/debug/tracing/trace_stat/function*). See the test script for more details [2]. Test 1: I hacked the DT to 'fake' per-CPU DVFS. That is, we end up with one CPUFreq policy per CPU (8 policies in total). Since all frequencies are pinned to max for the test, this should have no impact on the actual frequency selection, but it does in the EAS calculation. +---------------------------+----------------------------------+ | Without patch | With patch | +-----+-----+----------+----------+-----+-----------------+----------+ | CPU | Hit | Avg (us) | s^2 (us) | Hit | Avg (us) | s^2 (us) | |-----+-----+----------+----------+-----+-----------------+----------+ | 0 | 274 | 38.303 | 1750.239 | 401 | 14.126 (-63.1%) | 146.625 | | 1 | 197 | 49.529 | 1695.852 | 314 | 16.135 (-67.4%) | 167.525 | | 2 | 142 | 34.296 | 1758.665 | 302 | 14.133 (-58.8%) | 130.071 | | 3 | 172 | 31.734 | 1490.975 | 641 | 14.637 (-53.9%) | 139.189 | | 4 | 316 | 7.834 | 178.217 | 425 | 5.413 (-30.9%) | 20.803 | | 5 | 447 | 8.424 | 144.638 | 556 | 5.929 (-29.6%) | 27.301 | | 6 | 581 | 14.886 | 346.793 | 456 | 5.711 (-61.6%) | 23.124 | | 7 | 456 | 10.005 | 211.187 | 997 | 4.708 (-52.9%) | 21.144 | +-----+-----+----------+----------+-----+-----------------+----------+ * Hit, Avg and s^2 are as reported by the function profiler Test 2: I also ran the same test with a normal DT, with 2 CPUFreq policies, to see if this causes regressions in the most common case. +---------------------------+----------------------------------+ | Without patch | With patch | +-----+-----+----------+----------+-----+-----------------+----------+ | CPU | Hit | Avg (us) | s^2 (us) | Hit | Avg (us) | s^2 (us) | |-----+-----+----------+----------+-----+-----------------+----------+ | 0 | 345 | 22.184 | 215.321 | 580 | 18.635 (-16.0%) | 146.892 | | 1 | 358 | 18.597 | 200.596 | 438 | 12.934 (-30.5%) | 104.604 | | 2 | 359 | 25.566 | 200.217 | 397 | 10.826 (-57.7%) | 74.021 | | 3 | 362 | 16.881 | 200.291 | 718 | 11.455 (-32.1%) | 102.280 | | 4 | 457 | 3.822 | 9.895 | 757 | 4.616 (+20.8%) | 13.369 | | 5 | 344 | 4.301 | 7.121 | 594 | 5.320 (+23.7%) | 18.798 | | 6 | 472 | 4.326 | 7.849 | 464 | 5.648 (+30.6%) | 22.022 | | 7 | 331 | 4.630 | 13.937 | 408 | 5.299 (+14.4%) | 18.273 | +-----+-----+----------+----------+-----+-----------------+----------+ * Hit, Avg and s^2 are as reported by the function profiler In addition to these two tests, I also ran 50 iterations of the Lisa EAS functional test suite [3] with this patch applied on Arm Juno r0, Arm Juno r2, Arm TC2 and Hikey960, and could not see any regressions (all EAS functional tests are passing). [1] https://paste.debian.net/1100055/ [2] https://paste.debian.net/1100057/ [3] https://github.com/ARM-software/lisa/blob/master/lisa/tests/scheduler/eas_behaviour.py Signed-off-by: Quentin Perret <quentin.perret@arm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: juri.lelli@redhat.com Cc: morten.rasmussen@arm.com Cc: qais.yousef@arm.com Cc: qperret@qperret.net Cc: rjw@rjwysocki.net Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Link: https://lkml.kernel.org/r/20190912094404.13802-1-qperret@qperret.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-09-12 09:44:04 +00:00
cur_delta = compute_energy(p, max_spare_cap_cpu, pd);
cur_delta -= base_energy_pd;
if (cur_delta < best_delta) {
best_delta = cur_delta;
sched/fair: Select an energy-efficient CPU on task wake-up If an Energy Model (EM) is available and if the system isn't overutilized, re-route waking tasks into an energy-aware placement algorithm. The selection of an energy-efficient CPU for a task is achieved by estimating the impact on system-level active energy resulting from the placement of the task on the CPU with the highest spare capacity in each performance domain. This strategy spreads tasks in a performance domain and avoids overly aggressive task packing. The best CPU energy-wise is then selected if it saves a large enough amount of energy with respect to prev_cpu. Although it has already shown significant benefits on some existing targets, this approach cannot scale to platforms with numerous CPUs. This is an attempt to do something useful as writing a fast heuristic that performs reasonably well on a broad spectrum of architectures isn't an easy task. As such, the scope of usability of the energy-aware wake-up path is restricted to systems with the SD_ASYM_CPUCAPACITY flag set, and where the EM isn't too complex. Signed-off-by: Quentin Perret <quentin.perret@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: adharmap@codeaurora.org Cc: chris.redpath@arm.com Cc: currojerez@riseup.net Cc: dietmar.eggemann@arm.com Cc: edubezval@gmail.com Cc: gregkh@linuxfoundation.org Cc: javi.merino@kernel.org Cc: joel@joelfernandes.org Cc: juri.lelli@redhat.com Cc: morten.rasmussen@arm.com Cc: patrick.bellasi@arm.com Cc: pkondeti@codeaurora.org Cc: rjw@rjwysocki.net Cc: skannan@codeaurora.org Cc: smuckle@google.com Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Cc: viresh.kumar@linaro.org Link: https://lkml.kernel.org/r/20181203095628.11858-15-quentin.perret@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-12-03 09:56:27 +00:00
best_energy_cpu = max_spare_cap_cpu;
}
}
}
unlock:
rcu_read_unlock();
/*
* Pick the best CPU if prev_cpu cannot be used, or if it saves at
* least 6% of the energy used by prev_cpu.
*/
sched/fair: Speed-up energy-aware wake-ups EAS computes the energy impact of migrating a waking task when deciding on which CPU it should run. However, the current approach is known to have a high algorithmic complexity, which can result in prohibitively high wake-up latencies on systems with complex energy models, such as systems with per-CPU DVFS. On such systems, the algorithm complexity is in O(n^2) (ignoring the cost of searching for performance states in the EM) with 'n' the number of CPUs. To address this, re-factor the EAS wake-up path to compute the energy 'delta' (with and without the task) on a per-performance domain basis, rather than system-wide, which brings the complexity down to O(n). No functional changes intended. Test results ~~~~~~~~~~~~ * Setup: Tested on a Google Pixel 3, with a Snapdragon 845 (4+4 CPUs, A55/A75). Base kernel is 5.3-rc5 + Pixel3 specific patches. Android userspace, no graphics. * Test case: Run a periodic rt-app task, with 16ms period, ramping down from 70% to 10%, in 5% steps of 500 ms each (json avail. at [1]). Frequencies of all CPUs are pinned to max (using scaling_min_freq CPUFreq sysfs entries) to reduce variability. The time to run select_task_rq_fair() is measured using the function profiler (/sys/kernel/debug/tracing/trace_stat/function*). See the test script for more details [2]. Test 1: I hacked the DT to 'fake' per-CPU DVFS. That is, we end up with one CPUFreq policy per CPU (8 policies in total). Since all frequencies are pinned to max for the test, this should have no impact on the actual frequency selection, but it does in the EAS calculation. +---------------------------+----------------------------------+ | Without patch | With patch | +-----+-----+----------+----------+-----+-----------------+----------+ | CPU | Hit | Avg (us) | s^2 (us) | Hit | Avg (us) | s^2 (us) | |-----+-----+----------+----------+-----+-----------------+----------+ | 0 | 274 | 38.303 | 1750.239 | 401 | 14.126 (-63.1%) | 146.625 | | 1 | 197 | 49.529 | 1695.852 | 314 | 16.135 (-67.4%) | 167.525 | | 2 | 142 | 34.296 | 1758.665 | 302 | 14.133 (-58.8%) | 130.071 | | 3 | 172 | 31.734 | 1490.975 | 641 | 14.637 (-53.9%) | 139.189 | | 4 | 316 | 7.834 | 178.217 | 425 | 5.413 (-30.9%) | 20.803 | | 5 | 447 | 8.424 | 144.638 | 556 | 5.929 (-29.6%) | 27.301 | | 6 | 581 | 14.886 | 346.793 | 456 | 5.711 (-61.6%) | 23.124 | | 7 | 456 | 10.005 | 211.187 | 997 | 4.708 (-52.9%) | 21.144 | +-----+-----+----------+----------+-----+-----------------+----------+ * Hit, Avg and s^2 are as reported by the function profiler Test 2: I also ran the same test with a normal DT, with 2 CPUFreq policies, to see if this causes regressions in the most common case. +---------------------------+----------------------------------+ | Without patch | With patch | +-----+-----+----------+----------+-----+-----------------+----------+ | CPU | Hit | Avg (us) | s^2 (us) | Hit | Avg (us) | s^2 (us) | |-----+-----+----------+----------+-----+-----------------+----------+ | 0 | 345 | 22.184 | 215.321 | 580 | 18.635 (-16.0%) | 146.892 | | 1 | 358 | 18.597 | 200.596 | 438 | 12.934 (-30.5%) | 104.604 | | 2 | 359 | 25.566 | 200.217 | 397 | 10.826 (-57.7%) | 74.021 | | 3 | 362 | 16.881 | 200.291 | 718 | 11.455 (-32.1%) | 102.280 | | 4 | 457 | 3.822 | 9.895 | 757 | 4.616 (+20.8%) | 13.369 | | 5 | 344 | 4.301 | 7.121 | 594 | 5.320 (+23.7%) | 18.798 | | 6 | 472 | 4.326 | 7.849 | 464 | 5.648 (+30.6%) | 22.022 | | 7 | 331 | 4.630 | 13.937 | 408 | 5.299 (+14.4%) | 18.273 | +-----+-----+----------+----------+-----+-----------------+----------+ * Hit, Avg and s^2 are as reported by the function profiler In addition to these two tests, I also ran 50 iterations of the Lisa EAS functional test suite [3] with this patch applied on Arm Juno r0, Arm Juno r2, Arm TC2 and Hikey960, and could not see any regressions (all EAS functional tests are passing). [1] https://paste.debian.net/1100055/ [2] https://paste.debian.net/1100057/ [3] https://github.com/ARM-software/lisa/blob/master/lisa/tests/scheduler/eas_behaviour.py Signed-off-by: Quentin Perret <quentin.perret@arm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: juri.lelli@redhat.com Cc: morten.rasmussen@arm.com Cc: qais.yousef@arm.com Cc: qperret@qperret.net Cc: rjw@rjwysocki.net Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Link: https://lkml.kernel.org/r/20190912094404.13802-1-qperret@qperret.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-09-12 09:44:04 +00:00
if (prev_delta == ULONG_MAX)
sched/fair: Select an energy-efficient CPU on task wake-up If an Energy Model (EM) is available and if the system isn't overutilized, re-route waking tasks into an energy-aware placement algorithm. The selection of an energy-efficient CPU for a task is achieved by estimating the impact on system-level active energy resulting from the placement of the task on the CPU with the highest spare capacity in each performance domain. This strategy spreads tasks in a performance domain and avoids overly aggressive task packing. The best CPU energy-wise is then selected if it saves a large enough amount of energy with respect to prev_cpu. Although it has already shown significant benefits on some existing targets, this approach cannot scale to platforms with numerous CPUs. This is an attempt to do something useful as writing a fast heuristic that performs reasonably well on a broad spectrum of architectures isn't an easy task. As such, the scope of usability of the energy-aware wake-up path is restricted to systems with the SD_ASYM_CPUCAPACITY flag set, and where the EM isn't too complex. Signed-off-by: Quentin Perret <quentin.perret@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: adharmap@codeaurora.org Cc: chris.redpath@arm.com Cc: currojerez@riseup.net Cc: dietmar.eggemann@arm.com Cc: edubezval@gmail.com Cc: gregkh@linuxfoundation.org Cc: javi.merino@kernel.org Cc: joel@joelfernandes.org Cc: juri.lelli@redhat.com Cc: morten.rasmussen@arm.com Cc: patrick.bellasi@arm.com Cc: pkondeti@codeaurora.org Cc: rjw@rjwysocki.net Cc: skannan@codeaurora.org Cc: smuckle@google.com Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Cc: viresh.kumar@linaro.org Link: https://lkml.kernel.org/r/20181203095628.11858-15-quentin.perret@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-12-03 09:56:27 +00:00
return best_energy_cpu;
sched/fair: Speed-up energy-aware wake-ups EAS computes the energy impact of migrating a waking task when deciding on which CPU it should run. However, the current approach is known to have a high algorithmic complexity, which can result in prohibitively high wake-up latencies on systems with complex energy models, such as systems with per-CPU DVFS. On such systems, the algorithm complexity is in O(n^2) (ignoring the cost of searching for performance states in the EM) with 'n' the number of CPUs. To address this, re-factor the EAS wake-up path to compute the energy 'delta' (with and without the task) on a per-performance domain basis, rather than system-wide, which brings the complexity down to O(n). No functional changes intended. Test results ~~~~~~~~~~~~ * Setup: Tested on a Google Pixel 3, with a Snapdragon 845 (4+4 CPUs, A55/A75). Base kernel is 5.3-rc5 + Pixel3 specific patches. Android userspace, no graphics. * Test case: Run a periodic rt-app task, with 16ms period, ramping down from 70% to 10%, in 5% steps of 500 ms each (json avail. at [1]). Frequencies of all CPUs are pinned to max (using scaling_min_freq CPUFreq sysfs entries) to reduce variability. The time to run select_task_rq_fair() is measured using the function profiler (/sys/kernel/debug/tracing/trace_stat/function*). See the test script for more details [2]. Test 1: I hacked the DT to 'fake' per-CPU DVFS. That is, we end up with one CPUFreq policy per CPU (8 policies in total). Since all frequencies are pinned to max for the test, this should have no impact on the actual frequency selection, but it does in the EAS calculation. +---------------------------+----------------------------------+ | Without patch | With patch | +-----+-----+----------+----------+-----+-----------------+----------+ | CPU | Hit | Avg (us) | s^2 (us) | Hit | Avg (us) | s^2 (us) | |-----+-----+----------+----------+-----+-----------------+----------+ | 0 | 274 | 38.303 | 1750.239 | 401 | 14.126 (-63.1%) | 146.625 | | 1 | 197 | 49.529 | 1695.852 | 314 | 16.135 (-67.4%) | 167.525 | | 2 | 142 | 34.296 | 1758.665 | 302 | 14.133 (-58.8%) | 130.071 | | 3 | 172 | 31.734 | 1490.975 | 641 | 14.637 (-53.9%) | 139.189 | | 4 | 316 | 7.834 | 178.217 | 425 | 5.413 (-30.9%) | 20.803 | | 5 | 447 | 8.424 | 144.638 | 556 | 5.929 (-29.6%) | 27.301 | | 6 | 581 | 14.886 | 346.793 | 456 | 5.711 (-61.6%) | 23.124 | | 7 | 456 | 10.005 | 211.187 | 997 | 4.708 (-52.9%) | 21.144 | +-----+-----+----------+----------+-----+-----------------+----------+ * Hit, Avg and s^2 are as reported by the function profiler Test 2: I also ran the same test with a normal DT, with 2 CPUFreq policies, to see if this causes regressions in the most common case. +---------------------------+----------------------------------+ | Without patch | With patch | +-----+-----+----------+----------+-----+-----------------+----------+ | CPU | Hit | Avg (us) | s^2 (us) | Hit | Avg (us) | s^2 (us) | |-----+-----+----------+----------+-----+-----------------+----------+ | 0 | 345 | 22.184 | 215.321 | 580 | 18.635 (-16.0%) | 146.892 | | 1 | 358 | 18.597 | 200.596 | 438 | 12.934 (-30.5%) | 104.604 | | 2 | 359 | 25.566 | 200.217 | 397 | 10.826 (-57.7%) | 74.021 | | 3 | 362 | 16.881 | 200.291 | 718 | 11.455 (-32.1%) | 102.280 | | 4 | 457 | 3.822 | 9.895 | 757 | 4.616 (+20.8%) | 13.369 | | 5 | 344 | 4.301 | 7.121 | 594 | 5.320 (+23.7%) | 18.798 | | 6 | 472 | 4.326 | 7.849 | 464 | 5.648 (+30.6%) | 22.022 | | 7 | 331 | 4.630 | 13.937 | 408 | 5.299 (+14.4%) | 18.273 | +-----+-----+----------+----------+-----+-----------------+----------+ * Hit, Avg and s^2 are as reported by the function profiler In addition to these two tests, I also ran 50 iterations of the Lisa EAS functional test suite [3] with this patch applied on Arm Juno r0, Arm Juno r2, Arm TC2 and Hikey960, and could not see any regressions (all EAS functional tests are passing). [1] https://paste.debian.net/1100055/ [2] https://paste.debian.net/1100057/ [3] https://github.com/ARM-software/lisa/blob/master/lisa/tests/scheduler/eas_behaviour.py Signed-off-by: Quentin Perret <quentin.perret@arm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: juri.lelli@redhat.com Cc: morten.rasmussen@arm.com Cc: qais.yousef@arm.com Cc: qperret@qperret.net Cc: rjw@rjwysocki.net Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Link: https://lkml.kernel.org/r/20190912094404.13802-1-qperret@qperret.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-09-12 09:44:04 +00:00
if ((prev_delta - best_delta) > ((prev_delta + base_energy) >> 4))
sched/fair: Select an energy-efficient CPU on task wake-up If an Energy Model (EM) is available and if the system isn't overutilized, re-route waking tasks into an energy-aware placement algorithm. The selection of an energy-efficient CPU for a task is achieved by estimating the impact on system-level active energy resulting from the placement of the task on the CPU with the highest spare capacity in each performance domain. This strategy spreads tasks in a performance domain and avoids overly aggressive task packing. The best CPU energy-wise is then selected if it saves a large enough amount of energy with respect to prev_cpu. Although it has already shown significant benefits on some existing targets, this approach cannot scale to platforms with numerous CPUs. This is an attempt to do something useful as writing a fast heuristic that performs reasonably well on a broad spectrum of architectures isn't an easy task. As such, the scope of usability of the energy-aware wake-up path is restricted to systems with the SD_ASYM_CPUCAPACITY flag set, and where the EM isn't too complex. Signed-off-by: Quentin Perret <quentin.perret@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: adharmap@codeaurora.org Cc: chris.redpath@arm.com Cc: currojerez@riseup.net Cc: dietmar.eggemann@arm.com Cc: edubezval@gmail.com Cc: gregkh@linuxfoundation.org Cc: javi.merino@kernel.org Cc: joel@joelfernandes.org Cc: juri.lelli@redhat.com Cc: morten.rasmussen@arm.com Cc: patrick.bellasi@arm.com Cc: pkondeti@codeaurora.org Cc: rjw@rjwysocki.net Cc: skannan@codeaurora.org Cc: smuckle@google.com Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Cc: viresh.kumar@linaro.org Link: https://lkml.kernel.org/r/20181203095628.11858-15-quentin.perret@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-12-03 09:56:27 +00:00
return best_energy_cpu;
return prev_cpu;
fail:
rcu_read_unlock();
return -1;
}
/*
* select_task_rq_fair: Select target runqueue for the waking task in domains
* that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
* SD_BALANCE_FORK, or SD_BALANCE_EXEC.
*
* Balances load by selecting the idlest CPU in the idlest group, or under
* certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
*
* Returns the target CPU number.
*
* preempt must be disabled.
*/
static int
select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
{
struct sched_domain *tmp, *sd = NULL;
int cpu = smp_processor_id();
int new_cpu = prev_cpu;
sched: Fix select_idle_sibling() logic in select_task_rq_fair() Issues in the current select_idle_sibling() logic in select_task_rq_fair() in the context of a task wake-up: a) Once we select the idle sibling, we use that domain (spanning the cpu that the task is currently woken-up and the idle sibling that we found) in our wake_affine() decisions. This domain is completely different from the domain(we are supposed to use) that spans the cpu that the task currently woken-up and the cpu where the task previously ran. b) We do select_idle_sibling() check only for the cpu that the task is currently woken-up on. If select_task_rq_fair() selects the previously run cpu for waking the task, doing a select_idle_sibling() check for that cpu also helps and we don't do this currently. c) In the scenarios where the cpu that the task is woken-up is busy but with its HT siblings are idle, we are selecting the task be woken-up on the idle HT sibling instead of a core that it previously ran and currently completely idle. i.e., we are not taking decisions based on wake_affine() but directly selecting an idle sibling that can cause an imbalance at the SMT/MC level which will be later corrected by the periodic load balancer. Fix this by first going through the load imbalance calculations using wake_affine() and once we make a decision of woken-up cpu vs previously-ran cpu, then choose a possible idle sibling for waking up the task on. Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> LKML-Reference: <1270079265.7835.8.camel@sbs-t61.sc.intel.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-03-31 23:47:45 +00:00
int want_affine = 0;
int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
if (sd_flag & SD_BALANCE_WAKE) {
record_wakee(p);
sched/fair: Select an energy-efficient CPU on task wake-up If an Energy Model (EM) is available and if the system isn't overutilized, re-route waking tasks into an energy-aware placement algorithm. The selection of an energy-efficient CPU for a task is achieved by estimating the impact on system-level active energy resulting from the placement of the task on the CPU with the highest spare capacity in each performance domain. This strategy spreads tasks in a performance domain and avoids overly aggressive task packing. The best CPU energy-wise is then selected if it saves a large enough amount of energy with respect to prev_cpu. Although it has already shown significant benefits on some existing targets, this approach cannot scale to platforms with numerous CPUs. This is an attempt to do something useful as writing a fast heuristic that performs reasonably well on a broad spectrum of architectures isn't an easy task. As such, the scope of usability of the energy-aware wake-up path is restricted to systems with the SD_ASYM_CPUCAPACITY flag set, and where the EM isn't too complex. Signed-off-by: Quentin Perret <quentin.perret@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: adharmap@codeaurora.org Cc: chris.redpath@arm.com Cc: currojerez@riseup.net Cc: dietmar.eggemann@arm.com Cc: edubezval@gmail.com Cc: gregkh@linuxfoundation.org Cc: javi.merino@kernel.org Cc: joel@joelfernandes.org Cc: juri.lelli@redhat.com Cc: morten.rasmussen@arm.com Cc: patrick.bellasi@arm.com Cc: pkondeti@codeaurora.org Cc: rjw@rjwysocki.net Cc: skannan@codeaurora.org Cc: smuckle@google.com Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Cc: viresh.kumar@linaro.org Link: https://lkml.kernel.org/r/20181203095628.11858-15-quentin.perret@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-12-03 09:56:27 +00:00
if (sched_energy_enabled()) {
sched/fair: Select an energy-efficient CPU on task wake-up If an Energy Model (EM) is available and if the system isn't overutilized, re-route waking tasks into an energy-aware placement algorithm. The selection of an energy-efficient CPU for a task is achieved by estimating the impact on system-level active energy resulting from the placement of the task on the CPU with the highest spare capacity in each performance domain. This strategy spreads tasks in a performance domain and avoids overly aggressive task packing. The best CPU energy-wise is then selected if it saves a large enough amount of energy with respect to prev_cpu. Although it has already shown significant benefits on some existing targets, this approach cannot scale to platforms with numerous CPUs. This is an attempt to do something useful as writing a fast heuristic that performs reasonably well on a broad spectrum of architectures isn't an easy task. As such, the scope of usability of the energy-aware wake-up path is restricted to systems with the SD_ASYM_CPUCAPACITY flag set, and where the EM isn't too complex. Signed-off-by: Quentin Perret <quentin.perret@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: adharmap@codeaurora.org Cc: chris.redpath@arm.com Cc: currojerez@riseup.net Cc: dietmar.eggemann@arm.com Cc: edubezval@gmail.com Cc: gregkh@linuxfoundation.org Cc: javi.merino@kernel.org Cc: joel@joelfernandes.org Cc: juri.lelli@redhat.com Cc: morten.rasmussen@arm.com Cc: patrick.bellasi@arm.com Cc: pkondeti@codeaurora.org Cc: rjw@rjwysocki.net Cc: skannan@codeaurora.org Cc: smuckle@google.com Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Cc: viresh.kumar@linaro.org Link: https://lkml.kernel.org/r/20181203095628.11858-15-quentin.perret@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-12-03 09:56:27 +00:00
new_cpu = find_energy_efficient_cpu(p, prev_cpu);
if (new_cpu >= 0)
return new_cpu;
new_cpu = prev_cpu;
}
want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, p->cpus_ptr);
}
sched: Dynamically allocate sched_domain/sched_group data-structures Instead of relying on static allocations for the sched_domain and sched_group trees, dynamically allocate and RCU free them. Allocating this dynamically also allows for some build_sched_groups() simplification since we can now (like with other simplifications) rely on the sched_domain tree instead of hard-coded knowledge. One tricky to note is that detach_destroy_domains() needs to hold rcu_read_lock() over the entire tear-down, per-cpu is not sufficient since that can lead to partial sched_group existance (could possibly be solved by doing the tear-down backwards but this is much more robust). A concequence of the above is that we can no longer print the sched_domain debug stuff from cpu_attach_domain() since that might now run with preemption disabled (due to classic RCU etc.) and sched_domain_debug() does some GFP_KERNEL allocations. Another thing to note is that we now fully rely on normal RCU and not RCU-sched, this is because with the new and exiting RCU flavours we grew over the years BH doesn't necessarily hold off RCU-sched grace periods (-rt is known to break this). This would in fact already cause us grief since we do sched_domain/sched_group iterations from softirq context. This patch is somewhat larger than I would like it to be, but I didn't find any means of shrinking/splitting this. Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> Cc: Nick Piggin <npiggin@kernel.dk> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Link: http://lkml.kernel.org/r/20110407122942.245307941@chello.nl Signed-off-by: Ingo Molnar <mingo@elte.hu>
2011-04-07 12:09:50 +00:00
rcu_read_lock();
for_each_domain(cpu, tmp) {
if (!(tmp->flags & SD_LOAD_BALANCE))
break;
/*
* If both 'cpu' and 'prev_cpu' are part of this domain,
sched: Fix select_idle_sibling() logic in select_task_rq_fair() Issues in the current select_idle_sibling() logic in select_task_rq_fair() in the context of a task wake-up: a) Once we select the idle sibling, we use that domain (spanning the cpu that the task is currently woken-up and the idle sibling that we found) in our wake_affine() decisions. This domain is completely different from the domain(we are supposed to use) that spans the cpu that the task currently woken-up and the cpu where the task previously ran. b) We do select_idle_sibling() check only for the cpu that the task is currently woken-up on. If select_task_rq_fair() selects the previously run cpu for waking the task, doing a select_idle_sibling() check for that cpu also helps and we don't do this currently. c) In the scenarios where the cpu that the task is woken-up is busy but with its HT siblings are idle, we are selecting the task be woken-up on the idle HT sibling instead of a core that it previously ran and currently completely idle. i.e., we are not taking decisions based on wake_affine() but directly selecting an idle sibling that can cause an imbalance at the SMT/MC level which will be later corrected by the periodic load balancer. Fix this by first going through the load imbalance calculations using wake_affine() and once we make a decision of woken-up cpu vs previously-ran cpu, then choose a possible idle sibling for waking up the task on. Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> LKML-Reference: <1270079265.7835.8.camel@sbs-t61.sc.intel.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-03-31 23:47:45 +00:00
* cpu is a valid SD_WAKE_AFFINE target.
*/
sched: Fix select_idle_sibling() logic in select_task_rq_fair() Issues in the current select_idle_sibling() logic in select_task_rq_fair() in the context of a task wake-up: a) Once we select the idle sibling, we use that domain (spanning the cpu that the task is currently woken-up and the idle sibling that we found) in our wake_affine() decisions. This domain is completely different from the domain(we are supposed to use) that spans the cpu that the task currently woken-up and the cpu where the task previously ran. b) We do select_idle_sibling() check only for the cpu that the task is currently woken-up on. If select_task_rq_fair() selects the previously run cpu for waking the task, doing a select_idle_sibling() check for that cpu also helps and we don't do this currently. c) In the scenarios where the cpu that the task is woken-up is busy but with its HT siblings are idle, we are selecting the task be woken-up on the idle HT sibling instead of a core that it previously ran and currently completely idle. i.e., we are not taking decisions based on wake_affine() but directly selecting an idle sibling that can cause an imbalance at the SMT/MC level which will be later corrected by the periodic load balancer. Fix this by first going through the load imbalance calculations using wake_affine() and once we make a decision of woken-up cpu vs previously-ran cpu, then choose a possible idle sibling for waking up the task on. Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> LKML-Reference: <1270079265.7835.8.camel@sbs-t61.sc.intel.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-03-31 23:47:45 +00:00
if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
if (cpu != prev_cpu)
new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
sd = NULL; /* Prefer wake_affine over balance flags */
break;
}
if (tmp->flags & sd_flag)
sd = tmp;
else if (!want_affine)
break;
}
if (unlikely(sd)) {
/* Slow path */
new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
} else if (sd_flag & SD_BALANCE_WAKE) { /* XXX always ? */
/* Fast path */
new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
if (want_affine)
current->recent_used_cpu = cpu;
}
sched: Dynamically allocate sched_domain/sched_group data-structures Instead of relying on static allocations for the sched_domain and sched_group trees, dynamically allocate and RCU free them. Allocating this dynamically also allows for some build_sched_groups() simplification since we can now (like with other simplifications) rely on the sched_domain tree instead of hard-coded knowledge. One tricky to note is that detach_destroy_domains() needs to hold rcu_read_lock() over the entire tear-down, per-cpu is not sufficient since that can lead to partial sched_group existance (could possibly be solved by doing the tear-down backwards but this is much more robust). A concequence of the above is that we can no longer print the sched_domain debug stuff from cpu_attach_domain() since that might now run with preemption disabled (due to classic RCU etc.) and sched_domain_debug() does some GFP_KERNEL allocations. Another thing to note is that we now fully rely on normal RCU and not RCU-sched, this is because with the new and exiting RCU flavours we grew over the years BH doesn't necessarily hold off RCU-sched grace periods (-rt is known to break this). This would in fact already cause us grief since we do sched_domain/sched_group iterations from softirq context. This patch is somewhat larger than I would like it to be, but I didn't find any means of shrinking/splitting this. Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> Cc: Nick Piggin <npiggin@kernel.dk> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Link: http://lkml.kernel.org/r/20110407122942.245307941@chello.nl Signed-off-by: Ingo Molnar <mingo@elte.hu>
2011-04-07 12:09:50 +00:00
rcu_read_unlock();
return new_cpu;
}
static void detach_entity_cfs_rq(struct sched_entity *se);
/*
* Called immediately before a task is migrated to a new CPU; task_cpu(p) and
* cfs_rq_of(p) references at time of call are still valid and identify the
* previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
*/
sched/numa: Reset scan rate whenever task moves across nodes Currently task scan rate is reset when NUMA balancer migrates the task to a different node. If NUMA balancer initiates a swap, reset is only applicable to the task that initiates the swap. Similarly no scan rate reset is done if the task is migrated across nodes by traditional load balancer. Instead move the scan reset to the migrate_task_rq. This ensures the task moved out of its preferred node, either gets back to its preferred node quickly or finds a new preferred node. Doing so, would be fair to all tasks migrating across nodes. Specjbb2005 results (8 warehouses) Higher bops are better 2 Socket - 2 Node Haswell - X86 JVMS Prev Current %Change 4 200668 203370 1.3465 1 321791 328431 2.06345 2 Socket - 4 Node Power8 - PowerNV JVMS Prev Current %Change 1 204848 206070 0.59654 2 Socket - 2 Node Power9 - PowerNV JVMS Prev Current %Change 4 188098 188386 0.153112 1 200351 201566 0.606436 4 Socket - 4 Node Power7 - PowerVM JVMS Prev Current %Change 8 58145.9 59157.4 1.73959 1 103798 105495 1.63491 Some events stats before and after applying the patch. perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 13,912,183 13,825,492 migrations 1,155,931 1,152,509 faults 367,139 371,948 cache-misses 54,240,196,814 55,654,206,041 sched:sched_move_numa 1,571 1,856 sched:sched_stick_numa 9 4 sched:sched_swap_numa 463 428 migrate:mm_migrate_pages 703 898 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 50155 57146 numa_hint_faults_local 45264 51612 numa_hit 239652 238164 numa_huge_pte_updates 36 16 numa_interleave 68 63 numa_local 239576 238085 numa_other 76 79 numa_pages_migrated 680 883 numa_pte_updates 71146 67540 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 3,156,720 3,288,525 migrations 30,354 38,652 faults 97,261 111,678 cache-misses 12,400,026,826 12,111,197,376 sched:sched_move_numa 4 900 sched:sched_stick_numa 0 0 sched:sched_swap_numa 1 5 migrate:mm_migrate_pages 20 714 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 272 18572 numa_hint_faults_local 186 14850 numa_hit 71362 73197 numa_huge_pte_updates 0 11 numa_interleave 23 25 numa_local 71299 73138 numa_other 63 59 numa_pages_migrated 2 712 numa_pte_updates 0 24021 perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 8,606,824 8,451,543 migrations 155,352 202,804 faults 301,409 310,024 cache-misses 157,759,224 253,522,507 sched:sched_move_numa 168 213 sched:sched_stick_numa 0 0 sched:sched_swap_numa 3 2 migrate:mm_migrate_pages 125 88 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 4650 11830 numa_hint_faults_local 3946 11301 numa_hit 90489 90038 numa_huge_pte_updates 0 0 numa_interleave 892 855 numa_local 90034 89796 numa_other 455 242 numa_pages_migrated 124 88 numa_pte_updates 4818 12039 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 2,113,167 2,049,153 migrations 10,533 11,405 faults 142,727 162,309 cache-misses 5,594,192 7,203,343 sched:sched_move_numa 10 22 sched:sched_stick_numa 0 0 sched:sched_swap_numa 0 0 migrate:mm_migrate_pages 6 1 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 744 1693 numa_hint_faults_local 584 1669 numa_hit 25551 25177 numa_huge_pte_updates 0 0 numa_interleave 263 194 numa_local 25302 24993 numa_other 249 184 numa_pages_migrated 6 1 numa_pte_updates 744 1577 perf stats 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 101,227,352 94,515,937 migrations 4,151,829 4,203,554 faults 745,233 832,697 cache-misses 224,669,561,766 226,248,698,331 sched:sched_move_numa 617 1,730 sched:sched_stick_numa 2 14 sched:sched_swap_numa 187 432 migrate:mm_migrate_pages 316 1,398 vmstat 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 24195 80079 numa_hint_faults_local 21639 68620 numa_hit 238331 241187 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 238331 241186 numa_other 0 1 numa_pages_migrated 204 1347 numa_pte_updates 24561 80729 perf stats 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 62,738,978 63,704,961 migrations 562,702 573,404 faults 228,465 230,878 cache-misses 75,778,067,952 76,568,222,781 sched:sched_move_numa 648 509 sched:sched_stick_numa 13 31 sched:sched_swap_numa 137 182 migrate:mm_migrate_pages 733 541 vmstat 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 10281 8501 numa_hint_faults_local 3242 2960 numa_hit 36338 35526 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 36338 35526 numa_other 0 0 numa_pages_migrated 706 539 numa_pte_updates 10176 8433 Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Jirka Hladky <jhladky@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1537552141-27815-4-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-09-21 17:48:58 +00:00
static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
{
/*
* As blocked tasks retain absolute vruntime the migration needs to
* deal with this by subtracting the old and adding the new
* min_vruntime -- the latter is done by enqueue_entity() when placing
* the task on the new runqueue.
*/
if (p->state == TASK_WAKING) {
struct sched_entity *se = &p->se;
struct cfs_rq *cfs_rq = cfs_rq_of(se);
u64 min_vruntime;
#ifndef CONFIG_64BIT
u64 min_vruntime_copy;
do {
min_vruntime_copy = cfs_rq->min_vruntime_copy;
smp_rmb();
min_vruntime = cfs_rq->min_vruntime;
} while (min_vruntime != min_vruntime_copy);
#else
min_vruntime = cfs_rq->min_vruntime;
#endif
se->vruntime -= min_vruntime;
}
if (p->on_rq == TASK_ON_RQ_MIGRATING) {
/*
* In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
* rq->lock and can modify state directly.
*/
lockdep_assert_held(&task_rq(p)->lock);
detach_entity_cfs_rq(&p->se);
} else {
/*
* We are supposed to update the task to "current" time, then
* its up to date and ready to go to new CPU/cfs_rq. But we
* have difficulty in getting what current time is, so simply
* throw away the out-of-date time. This will result in the
* wakee task is less decayed, but giving the wakee more load
* sounds not bad.
*/
remove_entity_load_avg(&p->se);
}
sched/fair: Rewrite runnable load and utilization average tracking The idea of runnable load average (let runnable time contribute to weight) was proposed by Paul Turner and Ben Segall, and it is still followed by this rewrite. This rewrite aims to solve the following issues: 1. cfs_rq's load average (namely runnable_load_avg and blocked_load_avg) is updated at the granularity of an entity at a time, which results in the cfs_rq's load average is stale or partially updated: at any time, only one entity is up to date, all other entities are effectively lagging behind. This is undesirable. To illustrate, if we have n runnable entities in the cfs_rq, as time elapses, they certainly become outdated: t0: cfs_rq { e1_old, e2_old, ..., en_old } and when we update: t1: update e1, then we have cfs_rq { e1_new, e2_old, ..., en_old } t2: update e2, then we have cfs_rq { e1_old, e2_new, ..., en_old } ... We solve this by combining all runnable entities' load averages together in cfs_rq's avg, and update the cfs_rq's avg as a whole. This is based on the fact that if we regard the update as a function, then: w * update(e) = update(w * e) and update(e1) + update(e2) = update(e1 + e2), then w1 * update(e1) + w2 * update(e2) = update(w1 * e1 + w2 * e2) therefore, by this rewrite, we have an entirely updated cfs_rq at the time we update it: t1: update cfs_rq { e1_new, e2_new, ..., en_new } t2: update cfs_rq { e1_new, e2_new, ..., en_new } ... 2. cfs_rq's load average is different between top rq->cfs_rq and other task_group's per CPU cfs_rqs in whether or not blocked_load_average contributes to the load. The basic idea behind runnable load average (the same for utilization) is that the blocked state is taken into account as opposed to only accounting for the currently runnable state. Therefore, the average should include both the runnable/running and blocked load averages. This rewrite does that. In addition, we also combine runnable/running and blocked averages of all entities into the cfs_rq's average, and update it together at once. This is based on the fact that: update(runnable) + update(blocked) = update(runnable + blocked) This significantly reduces the code as we don't need to separately maintain/update runnable/running load and blocked load. 3. How task_group entities' share is calculated is complex and imprecise. We reduce the complexity in this rewrite to allow a very simple rule: the task_group's load_avg is aggregated from its per CPU cfs_rqs's load_avgs. Then group entity's weight is simply proportional to its own cfs_rq's load_avg / task_group's load_avg. To illustrate, if a task_group has { cfs_rq1, cfs_rq2, ..., cfs_rqn }, then, task_group_avg = cfs_rq1_avg + cfs_rq2_avg + ... + cfs_rqn_avg, then cfs_rqx's entity's share = cfs_rqx_avg / task_group_avg * task_group's share To sum up, this rewrite in principle is equivalent to the current one, but fixes the issues described above. Turns out, it significantly reduces the code complexity and hence increases clarity and efficiency. In addition, the new averages are more smooth/continuous (no spurious spikes and valleys) and updated more consistently and quickly to reflect the load dynamics. As a result, we have less load tracking overhead, better performance, and especially better power efficiency due to more balanced load. Signed-off-by: Yuyang Du <yuyang.du@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: arjan@linux.intel.com Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: fengguang.wu@intel.com Cc: len.brown@intel.com Cc: morten.rasmussen@arm.com Cc: pjt@google.com Cc: rafael.j.wysocki@intel.com Cc: umgwanakikbuti@gmail.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1436918682-4971-3-git-send-email-yuyang.du@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-07-15 00:04:37 +00:00
/* Tell new CPU we are migrated */
p->se.avg.last_update_time = 0;
/* We have migrated, no longer consider this task hot */
sched/fair: Rewrite runnable load and utilization average tracking The idea of runnable load average (let runnable time contribute to weight) was proposed by Paul Turner and Ben Segall, and it is still followed by this rewrite. This rewrite aims to solve the following issues: 1. cfs_rq's load average (namely runnable_load_avg and blocked_load_avg) is updated at the granularity of an entity at a time, which results in the cfs_rq's load average is stale or partially updated: at any time, only one entity is up to date, all other entities are effectively lagging behind. This is undesirable. To illustrate, if we have n runnable entities in the cfs_rq, as time elapses, they certainly become outdated: t0: cfs_rq { e1_old, e2_old, ..., en_old } and when we update: t1: update e1, then we have cfs_rq { e1_new, e2_old, ..., en_old } t2: update e2, then we have cfs_rq { e1_old, e2_new, ..., en_old } ... We solve this by combining all runnable entities' load averages together in cfs_rq's avg, and update the cfs_rq's avg as a whole. This is based on the fact that if we regard the update as a function, then: w * update(e) = update(w * e) and update(e1) + update(e2) = update(e1 + e2), then w1 * update(e1) + w2 * update(e2) = update(w1 * e1 + w2 * e2) therefore, by this rewrite, we have an entirely updated cfs_rq at the time we update it: t1: update cfs_rq { e1_new, e2_new, ..., en_new } t2: update cfs_rq { e1_new, e2_new, ..., en_new } ... 2. cfs_rq's load average is different between top rq->cfs_rq and other task_group's per CPU cfs_rqs in whether or not blocked_load_average contributes to the load. The basic idea behind runnable load average (the same for utilization) is that the blocked state is taken into account as opposed to only accounting for the currently runnable state. Therefore, the average should include both the runnable/running and blocked load averages. This rewrite does that. In addition, we also combine runnable/running and blocked averages of all entities into the cfs_rq's average, and update it together at once. This is based on the fact that: update(runnable) + update(blocked) = update(runnable + blocked) This significantly reduces the code as we don't need to separately maintain/update runnable/running load and blocked load. 3. How task_group entities' share is calculated is complex and imprecise. We reduce the complexity in this rewrite to allow a very simple rule: the task_group's load_avg is aggregated from its per CPU cfs_rqs's load_avgs. Then group entity's weight is simply proportional to its own cfs_rq's load_avg / task_group's load_avg. To illustrate, if a task_group has { cfs_rq1, cfs_rq2, ..., cfs_rqn }, then, task_group_avg = cfs_rq1_avg + cfs_rq2_avg + ... + cfs_rqn_avg, then cfs_rqx's entity's share = cfs_rqx_avg / task_group_avg * task_group's share To sum up, this rewrite in principle is equivalent to the current one, but fixes the issues described above. Turns out, it significantly reduces the code complexity and hence increases clarity and efficiency. In addition, the new averages are more smooth/continuous (no spurious spikes and valleys) and updated more consistently and quickly to reflect the load dynamics. As a result, we have less load tracking overhead, better performance, and especially better power efficiency due to more balanced load. Signed-off-by: Yuyang Du <yuyang.du@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: arjan@linux.intel.com Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: fengguang.wu@intel.com Cc: len.brown@intel.com Cc: morten.rasmussen@arm.com Cc: pjt@google.com Cc: rafael.j.wysocki@intel.com Cc: umgwanakikbuti@gmail.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1436918682-4971-3-git-send-email-yuyang.du@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-07-15 00:04:37 +00:00
p->se.exec_start = 0;
sched/numa: Reset scan rate whenever task moves across nodes Currently task scan rate is reset when NUMA balancer migrates the task to a different node. If NUMA balancer initiates a swap, reset is only applicable to the task that initiates the swap. Similarly no scan rate reset is done if the task is migrated across nodes by traditional load balancer. Instead move the scan reset to the migrate_task_rq. This ensures the task moved out of its preferred node, either gets back to its preferred node quickly or finds a new preferred node. Doing so, would be fair to all tasks migrating across nodes. Specjbb2005 results (8 warehouses) Higher bops are better 2 Socket - 2 Node Haswell - X86 JVMS Prev Current %Change 4 200668 203370 1.3465 1 321791 328431 2.06345 2 Socket - 4 Node Power8 - PowerNV JVMS Prev Current %Change 1 204848 206070 0.59654 2 Socket - 2 Node Power9 - PowerNV JVMS Prev Current %Change 4 188098 188386 0.153112 1 200351 201566 0.606436 4 Socket - 4 Node Power7 - PowerVM JVMS Prev Current %Change 8 58145.9 59157.4 1.73959 1 103798 105495 1.63491 Some events stats before and after applying the patch. perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 13,912,183 13,825,492 migrations 1,155,931 1,152,509 faults 367,139 371,948 cache-misses 54,240,196,814 55,654,206,041 sched:sched_move_numa 1,571 1,856 sched:sched_stick_numa 9 4 sched:sched_swap_numa 463 428 migrate:mm_migrate_pages 703 898 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 50155 57146 numa_hint_faults_local 45264 51612 numa_hit 239652 238164 numa_huge_pte_updates 36 16 numa_interleave 68 63 numa_local 239576 238085 numa_other 76 79 numa_pages_migrated 680 883 numa_pte_updates 71146 67540 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After cs 3,156,720 3,288,525 migrations 30,354 38,652 faults 97,261 111,678 cache-misses 12,400,026,826 12,111,197,376 sched:sched_move_numa 4 900 sched:sched_stick_numa 0 0 sched:sched_swap_numa 1 5 migrate:mm_migrate_pages 20 714 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Haswell - X86 Event Before After numa_hint_faults 272 18572 numa_hint_faults_local 186 14850 numa_hit 71362 73197 numa_huge_pte_updates 0 11 numa_interleave 23 25 numa_local 71299 73138 numa_other 63 59 numa_pages_migrated 2 712 numa_pte_updates 0 24021 perf stats 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 8,606,824 8,451,543 migrations 155,352 202,804 faults 301,409 310,024 cache-misses 157,759,224 253,522,507 sched:sched_move_numa 168 213 sched:sched_stick_numa 0 0 sched:sched_swap_numa 3 2 migrate:mm_migrate_pages 125 88 vmstat 8th warehouse Multi JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 4650 11830 numa_hint_faults_local 3946 11301 numa_hit 90489 90038 numa_huge_pte_updates 0 0 numa_interleave 892 855 numa_local 90034 89796 numa_other 455 242 numa_pages_migrated 124 88 numa_pte_updates 4818 12039 perf stats 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After cs 2,113,167 2,049,153 migrations 10,533 11,405 faults 142,727 162,309 cache-misses 5,594,192 7,203,343 sched:sched_move_numa 10 22 sched:sched_stick_numa 0 0 sched:sched_swap_numa 0 0 migrate:mm_migrate_pages 6 1 vmstat 8th warehouse Single JVM 2 Socket - 2 Node Power9 - PowerNV Event Before After numa_hint_faults 744 1693 numa_hint_faults_local 584 1669 numa_hit 25551 25177 numa_huge_pte_updates 0 0 numa_interleave 263 194 numa_local 25302 24993 numa_other 249 184 numa_pages_migrated 6 1 numa_pte_updates 744 1577 perf stats 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 101,227,352 94,515,937 migrations 4,151,829 4,203,554 faults 745,233 832,697 cache-misses 224,669,561,766 226,248,698,331 sched:sched_move_numa 617 1,730 sched:sched_stick_numa 2 14 sched:sched_swap_numa 187 432 migrate:mm_migrate_pages 316 1,398 vmstat 8th warehouse Multi JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 24195 80079 numa_hint_faults_local 21639 68620 numa_hit 238331 241187 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 238331 241186 numa_other 0 1 numa_pages_migrated 204 1347 numa_pte_updates 24561 80729 perf stats 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After cs 62,738,978 63,704,961 migrations 562,702 573,404 faults 228,465 230,878 cache-misses 75,778,067,952 76,568,222,781 sched:sched_move_numa 648 509 sched:sched_stick_numa 13 31 sched:sched_swap_numa 137 182 migrate:mm_migrate_pages 733 541 vmstat 8th warehouse Single JVM 4 Socket - 4 Node Power7 - PowerVM Event Before After numa_hint_faults 10281 8501 numa_hint_faults_local 3242 2960 numa_hit 36338 35526 numa_huge_pte_updates 0 0 numa_interleave 0 0 numa_local 36338 35526 numa_other 0 0 numa_pages_migrated 706 539 numa_pte_updates 10176 8433 Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Jirka Hladky <jhladky@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1537552141-27815-4-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-09-21 17:48:58 +00:00
update_scan_period(p, new_cpu);
}
static void task_dead_fair(struct task_struct *p)
{
remove_entity_load_avg(&p->se);
}
sched: Fix pick_next_task() vs 'change' pattern race Commit 67692435c411 ("sched: Rework pick_next_task() slow-path") inadvertly introduced a race because it changed a previously unexplored dependency between dropping the rq->lock and sched_class::put_prev_task(). The comments about dropping rq->lock, in for example newidle_balance(), only mentions the task being current and ->on_cpu being set. But when we look at the 'change' pattern (in for example sched_setnuma()): queued = task_on_rq_queued(p); /* p->on_rq == TASK_ON_RQ_QUEUED */ running = task_current(rq, p); /* rq->curr == p */ if (queued) dequeue_task(...); if (running) put_prev_task(...); /* change task properties */ if (queued) enqueue_task(...); if (running) set_next_task(...); It becomes obvious that if we do this after put_prev_task() has already been called on @p, things go sideways. This is exactly what the commit in question allows to happen when it does: prev->sched_class->put_prev_task(rq, prev, rf); if (!rq->nr_running) newidle_balance(rq, rf); The newidle_balance() call will drop rq->lock after we've called put_prev_task() and that allows the above 'change' pattern to interleave and mess up the state. Furthermore, it turns out we lost the RT-pull when we put the last DL task. Fix both problems by extracting the balancing from put_prev_task() and doing a multi-class balance() pass before put_prev_task(). Fixes: 67692435c411 ("sched: Rework pick_next_task() slow-path") Reported-by: Quentin Perret <qperret@google.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Tested-by: Quentin Perret <qperret@google.com> Tested-by: Valentin Schneider <valentin.schneider@arm.com>
2019-11-08 10:11:52 +00:00
static int
balance_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
{
if (rq->nr_running)
return 1;
return newidle_balance(rq, rf) != 0;
}
#endif /* CONFIG_SMP */
static unsigned long wakeup_gran(struct sched_entity *se)
{
unsigned long gran = sysctl_sched_wakeup_granularity;
/*
* Since its curr running now, convert the gran from real-time
* to virtual-time in his units.
*
* By using 'se' instead of 'curr' we penalize light tasks, so
* they get preempted easier. That is, if 'se' < 'curr' then
* the resulting gran will be larger, therefore penalizing the
* lighter, if otoh 'se' > 'curr' then the resulting gran will
* be smaller, again penalizing the lighter task.
*
* This is especially important for buddies when the leftmost
* task is higher priority than the buddy.
*/
return calc_delta_fair(gran, se);
}
/*
* Should 'se' preempt 'curr'.
*
* |s1
* |s2
* |s3
* g
* |<--->|c
*
* w(c, s1) = -1
* w(c, s2) = 0
* w(c, s3) = 1
*
*/
static int
wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
{
s64 gran, vdiff = curr->vruntime - se->vruntime;
if (vdiff <= 0)
return -1;
gran = wakeup_gran(se);
if (vdiff > gran)
return 1;
return 0;
}
static void set_last_buddy(struct sched_entity *se)
{
if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
return;
for_each_sched_entity(se) {
if (SCHED_WARN_ON(!se->on_rq))
return;
cfs_rq_of(se)->last = se;
}
}
static void set_next_buddy(struct sched_entity *se)
{
if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
return;
for_each_sched_entity(se) {
if (SCHED_WARN_ON(!se->on_rq))
return;
cfs_rq_of(se)->next = se;
}
}
static void set_skip_buddy(struct sched_entity *se)
{
for_each_sched_entity(se)
cfs_rq_of(se)->skip = se;
}
/*
* Preempt the current task with a newly woken task if needed:
*/
static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
{
struct task_struct *curr = rq->curr;
struct sched_entity *se = &curr->se, *pse = &p->se;
struct cfs_rq *cfs_rq = task_cfs_rq(curr);
sched: Strengthen buddies and mitigate buddy induced latencies This patch restores the effectiveness of LAST_BUDDY in preventing pgsql+oltp from collapsing due to wakeup preemption. It also switches LAST_BUDDY to exclusively do what it does best, namely mitigate the effects of aggressive wakeup preemption, which improves vmark throughput markedly, and restores mysql+oltp scalability. Since buddies are about scalability, enable them beginning at the point where we begin expanding sched_latency, namely sched_nr_latency. Previously, buddies were cleared aggressively, which seriously reduced their effectiveness. Not clearing aggressively however, produces a small drop in mysql+oltp throughput immediately after peak, indicating that LAST_BUDDY is actually doing some harm. This is right at the point where X on the desktop in competition with another load wants low latency service. Ergo, do not enable until we need to scale. To mitigate latency induced by buddies, or by a task just missing wakeup preemption, check latency at tick time. Last hunk prevents buddies from stymieing BALANCE_NEWIDLE via CACHE_HOT_BUDDY. Supporting performance tests: tip = v2.6.32-rc5-1497-ga525b32 tipx = NO_GENTLE_FAIR_SLEEPERS NEXT_BUDDY granularity knobs = 31 knobs + 31 buddies tip+x = NO_GENTLE_FAIR_SLEEPERS granularity knobs = 31 knobs (Three run averages except where noted.) vmark: ------ tip 108466 messages per second tip+ 125307 messages per second tip+x 125335 messages per second tipx 117781 messages per second 2.6.31.3 122729 messages per second mysql+oltp: ----------- clients 1 2 4 8 16 32 64 128 256 .......................................................................................... tip 9949.89 18690.20 34801.24 34460.04 32682.88 30765.97 28305.27 25059.64 19548.08 tip+ 10013.90 18526.84 34900.38 34420.14 33069.83 32083.40 30578.30 28010.71 25605.47 tipx 9698.71 18002.70 34477.56 33420.01 32634.30 31657.27 29932.67 26827.52 21487.18 2.6.31.3 8243.11 18784.20 34404.83 33148.38 31900.32 31161.90 29663.81 25995.94 18058.86 pgsql+oltp: ----------- clients 1 2 4 8 16 32 64 128 256 .......................................................................................... tip 13686.37 26609.25 51934.28 51347.81 49479.51 45312.65 36691.91 26851.57 24145.35 tip+ (1x) 13907.85 27135.87 52951.98 52514.04 51742.52 50705.43 49947.97 48374.19 46227.94 tip+x 13906.78 27065.81 52951.19 52542.59 52176.11 51815.94 50838.90 49439.46 46891.00 tipx 13742.46 26769.81 52351.99 51891.73 51320.79 50938.98 50248.65 48908.70 46553.84 2.6.31.3 13815.35 26906.46 52683.34 52061.31 51937.10 51376.80 50474.28 49394.47 47003.25 Signed-off-by: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> LKML-Reference: <new-submission> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-10-23 21:09:22 +00:00
int scale = cfs_rq->nr_running >= sched_nr_latency;
sched: Next buddy hint on sleep and preempt path When a task in a taskgroup sleeps, pick_next_task starts all the way back at the root and picks the task/taskgroup with the min vruntime across all runnable tasks. But when there are many frequently sleeping tasks across different taskgroups, it makes better sense to stay with same taskgroup for its slice period (or until all tasks in the taskgroup sleeps) instead of switching cross taskgroup on each sleep after a short runtime. This helps specifically where taskgroups corresponds to a process with multiple threads. The change reduces the number of CR3 switches in this case. Example: Two taskgroups with 2 threads each which are running for 2ms and sleeping for 1ms. Looking at sched:sched_switch shows: BEFORE: taskgroup_1 threads [5004, 5005], taskgroup_2 threads [5016, 5017] cpu-soaker-5004 [003] 3683.391089 cpu-soaker-5016 [003] 3683.393106 cpu-soaker-5005 [003] 3683.395119 cpu-soaker-5017 [003] 3683.397130 cpu-soaker-5004 [003] 3683.399143 cpu-soaker-5016 [003] 3683.401155 cpu-soaker-5005 [003] 3683.403168 cpu-soaker-5017 [003] 3683.405170 AFTER: taskgroup_1 threads [21890, 21891], taskgroup_2 threads [21934, 21935] cpu-soaker-21890 [003] 865.895494 cpu-soaker-21935 [003] 865.897506 cpu-soaker-21934 [003] 865.899520 cpu-soaker-21935 [003] 865.901532 cpu-soaker-21934 [003] 865.903543 cpu-soaker-21935 [003] 865.905546 cpu-soaker-21891 [003] 865.907548 cpu-soaker-21890 [003] 865.909560 cpu-soaker-21891 [003] 865.911571 cpu-soaker-21890 [003] 865.913582 cpu-soaker-21891 [003] 865.915594 cpu-soaker-21934 [003] 865.917606 Similar problem is there when there are multiple taskgroups and say a task A preempts currently running task B of taskgroup_1. On schedule, pick_next_task can pick an unrelated task on taskgroup_2. Here it would be better to give some preference to task B on pick_next_task. A simple (may be extreme case) benchmark I tried was tbench with 2 tbench client processes with 2 threads each running on a single CPU. Avg throughput across 5 50 sec runs was: BEFORE: 105.84 MB/sec AFTER: 112.42 MB/sec Signed-off-by: Venkatesh Pallipadi <venki@google.com> Acked-by: Rik van Riel <riel@redhat.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/1302802253-25760-1-git-send-email-venki@google.com Signed-off-by: Ingo Molnar <mingo@elte.hu>
2011-04-14 17:30:53 +00:00
int next_buddy_marked = 0;
if (unlikely(se == pse))
return;
/*
* This is possible from callers such as attach_tasks(), in which we
* unconditionally check_prempt_curr() after an enqueue (which may have
* lead to a throttle). This both saves work and prevents false
* next-buddy nomination below.
*/
if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
return;
sched: Next buddy hint on sleep and preempt path When a task in a taskgroup sleeps, pick_next_task starts all the way back at the root and picks the task/taskgroup with the min vruntime across all runnable tasks. But when there are many frequently sleeping tasks across different taskgroups, it makes better sense to stay with same taskgroup for its slice period (or until all tasks in the taskgroup sleeps) instead of switching cross taskgroup on each sleep after a short runtime. This helps specifically where taskgroups corresponds to a process with multiple threads. The change reduces the number of CR3 switches in this case. Example: Two taskgroups with 2 threads each which are running for 2ms and sleeping for 1ms. Looking at sched:sched_switch shows: BEFORE: taskgroup_1 threads [5004, 5005], taskgroup_2 threads [5016, 5017] cpu-soaker-5004 [003] 3683.391089 cpu-soaker-5016 [003] 3683.393106 cpu-soaker-5005 [003] 3683.395119 cpu-soaker-5017 [003] 3683.397130 cpu-soaker-5004 [003] 3683.399143 cpu-soaker-5016 [003] 3683.401155 cpu-soaker-5005 [003] 3683.403168 cpu-soaker-5017 [003] 3683.405170 AFTER: taskgroup_1 threads [21890, 21891], taskgroup_2 threads [21934, 21935] cpu-soaker-21890 [003] 865.895494 cpu-soaker-21935 [003] 865.897506 cpu-soaker-21934 [003] 865.899520 cpu-soaker-21935 [003] 865.901532 cpu-soaker-21934 [003] 865.903543 cpu-soaker-21935 [003] 865.905546 cpu-soaker-21891 [003] 865.907548 cpu-soaker-21890 [003] 865.909560 cpu-soaker-21891 [003] 865.911571 cpu-soaker-21890 [003] 865.913582 cpu-soaker-21891 [003] 865.915594 cpu-soaker-21934 [003] 865.917606 Similar problem is there when there are multiple taskgroups and say a task A preempts currently running task B of taskgroup_1. On schedule, pick_next_task can pick an unrelated task on taskgroup_2. Here it would be better to give some preference to task B on pick_next_task. A simple (may be extreme case) benchmark I tried was tbench with 2 tbench client processes with 2 threads each running on a single CPU. Avg throughput across 5 50 sec runs was: BEFORE: 105.84 MB/sec AFTER: 112.42 MB/sec Signed-off-by: Venkatesh Pallipadi <venki@google.com> Acked-by: Rik van Riel <riel@redhat.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/1302802253-25760-1-git-send-email-venki@google.com Signed-off-by: Ingo Molnar <mingo@elte.hu>
2011-04-14 17:30:53 +00:00
if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
set_next_buddy(pse);
sched: Next buddy hint on sleep and preempt path When a task in a taskgroup sleeps, pick_next_task starts all the way back at the root and picks the task/taskgroup with the min vruntime across all runnable tasks. But when there are many frequently sleeping tasks across different taskgroups, it makes better sense to stay with same taskgroup for its slice period (or until all tasks in the taskgroup sleeps) instead of switching cross taskgroup on each sleep after a short runtime. This helps specifically where taskgroups corresponds to a process with multiple threads. The change reduces the number of CR3 switches in this case. Example: Two taskgroups with 2 threads each which are running for 2ms and sleeping for 1ms. Looking at sched:sched_switch shows: BEFORE: taskgroup_1 threads [5004, 5005], taskgroup_2 threads [5016, 5017] cpu-soaker-5004 [003] 3683.391089 cpu-soaker-5016 [003] 3683.393106 cpu-soaker-5005 [003] 3683.395119 cpu-soaker-5017 [003] 3683.397130 cpu-soaker-5004 [003] 3683.399143 cpu-soaker-5016 [003] 3683.401155 cpu-soaker-5005 [003] 3683.403168 cpu-soaker-5017 [003] 3683.405170 AFTER: taskgroup_1 threads [21890, 21891], taskgroup_2 threads [21934, 21935] cpu-soaker-21890 [003] 865.895494 cpu-soaker-21935 [003] 865.897506 cpu-soaker-21934 [003] 865.899520 cpu-soaker-21935 [003] 865.901532 cpu-soaker-21934 [003] 865.903543 cpu-soaker-21935 [003] 865.905546 cpu-soaker-21891 [003] 865.907548 cpu-soaker-21890 [003] 865.909560 cpu-soaker-21891 [003] 865.911571 cpu-soaker-21890 [003] 865.913582 cpu-soaker-21891 [003] 865.915594 cpu-soaker-21934 [003] 865.917606 Similar problem is there when there are multiple taskgroups and say a task A preempts currently running task B of taskgroup_1. On schedule, pick_next_task can pick an unrelated task on taskgroup_2. Here it would be better to give some preference to task B on pick_next_task. A simple (may be extreme case) benchmark I tried was tbench with 2 tbench client processes with 2 threads each running on a single CPU. Avg throughput across 5 50 sec runs was: BEFORE: 105.84 MB/sec AFTER: 112.42 MB/sec Signed-off-by: Venkatesh Pallipadi <venki@google.com> Acked-by: Rik van Riel <riel@redhat.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/1302802253-25760-1-git-send-email-venki@google.com Signed-off-by: Ingo Molnar <mingo@elte.hu>
2011-04-14 17:30:53 +00:00
next_buddy_marked = 1;
}
/*
* We can come here with TIF_NEED_RESCHED already set from new task
* wake up path.
*
* Note: this also catches the edge-case of curr being in a throttled
* group (e.g. via set_curr_task), since update_curr() (in the
* enqueue of curr) will have resulted in resched being set. This
* prevents us from potentially nominating it as a false LAST_BUDDY
* below.
*/
if (test_tsk_need_resched(curr))
return;
/* Idle tasks are by definition preempted by non-idle tasks. */
if (unlikely(task_has_idle_policy(curr)) &&
likely(!task_has_idle_policy(p)))
goto preempt;
/*
* Batch and idle tasks do not preempt non-idle tasks (their preemption
* is driven by the tick):
*/
if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
return;
find_matching_se(&se, &pse);
update_curr(cfs_rq_of(se));
BUG_ON(!pse);
sched: Next buddy hint on sleep and preempt path When a task in a taskgroup sleeps, pick_next_task starts all the way back at the root and picks the task/taskgroup with the min vruntime across all runnable tasks. But when there are many frequently sleeping tasks across different taskgroups, it makes better sense to stay with same taskgroup for its slice period (or until all tasks in the taskgroup sleeps) instead of switching cross taskgroup on each sleep after a short runtime. This helps specifically where taskgroups corresponds to a process with multiple threads. The change reduces the number of CR3 switches in this case. Example: Two taskgroups with 2 threads each which are running for 2ms and sleeping for 1ms. Looking at sched:sched_switch shows: BEFORE: taskgroup_1 threads [5004, 5005], taskgroup_2 threads [5016, 5017] cpu-soaker-5004 [003] 3683.391089 cpu-soaker-5016 [003] 3683.393106 cpu-soaker-5005 [003] 3683.395119 cpu-soaker-5017 [003] 3683.397130 cpu-soaker-5004 [003] 3683.399143 cpu-soaker-5016 [003] 3683.401155 cpu-soaker-5005 [003] 3683.403168 cpu-soaker-5017 [003] 3683.405170 AFTER: taskgroup_1 threads [21890, 21891], taskgroup_2 threads [21934, 21935] cpu-soaker-21890 [003] 865.895494 cpu-soaker-21935 [003] 865.897506 cpu-soaker-21934 [003] 865.899520 cpu-soaker-21935 [003] 865.901532 cpu-soaker-21934 [003] 865.903543 cpu-soaker-21935 [003] 865.905546 cpu-soaker-21891 [003] 865.907548 cpu-soaker-21890 [003] 865.909560 cpu-soaker-21891 [003] 865.911571 cpu-soaker-21890 [003] 865.913582 cpu-soaker-21891 [003] 865.915594 cpu-soaker-21934 [003] 865.917606 Similar problem is there when there are multiple taskgroups and say a task A preempts currently running task B of taskgroup_1. On schedule, pick_next_task can pick an unrelated task on taskgroup_2. Here it would be better to give some preference to task B on pick_next_task. A simple (may be extreme case) benchmark I tried was tbench with 2 tbench client processes with 2 threads each running on a single CPU. Avg throughput across 5 50 sec runs was: BEFORE: 105.84 MB/sec AFTER: 112.42 MB/sec Signed-off-by: Venkatesh Pallipadi <venki@google.com> Acked-by: Rik van Riel <riel@redhat.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/1302802253-25760-1-git-send-email-venki@google.com Signed-off-by: Ingo Molnar <mingo@elte.hu>
2011-04-14 17:30:53 +00:00
if (wakeup_preempt_entity(se, pse) == 1) {
/*
* Bias pick_next to pick the sched entity that is
* triggering this preemption.
*/
if (!next_buddy_marked)
set_next_buddy(pse);
goto preempt;
sched: Next buddy hint on sleep and preempt path When a task in a taskgroup sleeps, pick_next_task starts all the way back at the root and picks the task/taskgroup with the min vruntime across all runnable tasks. But when there are many frequently sleeping tasks across different taskgroups, it makes better sense to stay with same taskgroup for its slice period (or until all tasks in the taskgroup sleeps) instead of switching cross taskgroup on each sleep after a short runtime. This helps specifically where taskgroups corresponds to a process with multiple threads. The change reduces the number of CR3 switches in this case. Example: Two taskgroups with 2 threads each which are running for 2ms and sleeping for 1ms. Looking at sched:sched_switch shows: BEFORE: taskgroup_1 threads [5004, 5005], taskgroup_2 threads [5016, 5017] cpu-soaker-5004 [003] 3683.391089 cpu-soaker-5016 [003] 3683.393106 cpu-soaker-5005 [003] 3683.395119 cpu-soaker-5017 [003] 3683.397130 cpu-soaker-5004 [003] 3683.399143 cpu-soaker-5016 [003] 3683.401155 cpu-soaker-5005 [003] 3683.403168 cpu-soaker-5017 [003] 3683.405170 AFTER: taskgroup_1 threads [21890, 21891], taskgroup_2 threads [21934, 21935] cpu-soaker-21890 [003] 865.895494 cpu-soaker-21935 [003] 865.897506 cpu-soaker-21934 [003] 865.899520 cpu-soaker-21935 [003] 865.901532 cpu-soaker-21934 [003] 865.903543 cpu-soaker-21935 [003] 865.905546 cpu-soaker-21891 [003] 865.907548 cpu-soaker-21890 [003] 865.909560 cpu-soaker-21891 [003] 865.911571 cpu-soaker-21890 [003] 865.913582 cpu-soaker-21891 [003] 865.915594 cpu-soaker-21934 [003] 865.917606 Similar problem is there when there are multiple taskgroups and say a task A preempts currently running task B of taskgroup_1. On schedule, pick_next_task can pick an unrelated task on taskgroup_2. Here it would be better to give some preference to task B on pick_next_task. A simple (may be extreme case) benchmark I tried was tbench with 2 tbench client processes with 2 threads each running on a single CPU. Avg throughput across 5 50 sec runs was: BEFORE: 105.84 MB/sec AFTER: 112.42 MB/sec Signed-off-by: Venkatesh Pallipadi <venki@google.com> Acked-by: Rik van Riel <riel@redhat.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/1302802253-25760-1-git-send-email-venki@google.com Signed-off-by: Ingo Molnar <mingo@elte.hu>
2011-04-14 17:30:53 +00:00
}
return;
preempt:
resched_curr(rq);
/*
* Only set the backward buddy when the current task is still
* on the rq. This can happen when a wakeup gets interleaved
* with schedule on the ->pre_schedule() or idle_balance()
* point, either of which can * drop the rq lock.
*
* Also, during early boot the idle thread is in the fair class,
* for obvious reasons its a bad idea to schedule back to it.
*/
if (unlikely(!se->on_rq || curr == rq->idle))
return;
if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
set_last_buddy(se);
}
struct task_struct *
pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
{
struct cfs_rq *cfs_rq = &rq->cfs;
struct sched_entity *se;
struct task_struct *p;
int new_tasks;
again:
sched: Fix pick_next_task() vs 'change' pattern race Commit 67692435c411 ("sched: Rework pick_next_task() slow-path") inadvertly introduced a race because it changed a previously unexplored dependency between dropping the rq->lock and sched_class::put_prev_task(). The comments about dropping rq->lock, in for example newidle_balance(), only mentions the task being current and ->on_cpu being set. But when we look at the 'change' pattern (in for example sched_setnuma()): queued = task_on_rq_queued(p); /* p->on_rq == TASK_ON_RQ_QUEUED */ running = task_current(rq, p); /* rq->curr == p */ if (queued) dequeue_task(...); if (running) put_prev_task(...); /* change task properties */ if (queued) enqueue_task(...); if (running) set_next_task(...); It becomes obvious that if we do this after put_prev_task() has already been called on @p, things go sideways. This is exactly what the commit in question allows to happen when it does: prev->sched_class->put_prev_task(rq, prev, rf); if (!rq->nr_running) newidle_balance(rq, rf); The newidle_balance() call will drop rq->lock after we've called put_prev_task() and that allows the above 'change' pattern to interleave and mess up the state. Furthermore, it turns out we lost the RT-pull when we put the last DL task. Fix both problems by extracting the balancing from put_prev_task() and doing a multi-class balance() pass before put_prev_task(). Fixes: 67692435c411 ("sched: Rework pick_next_task() slow-path") Reported-by: Quentin Perret <qperret@google.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Tested-by: Quentin Perret <qperret@google.com> Tested-by: Valentin Schneider <valentin.schneider@arm.com>
2019-11-08 10:11:52 +00:00
if (!sched_fair_runnable(rq))
goto idle;
#ifdef CONFIG_FAIR_GROUP_SCHED
if (!prev || prev->sched_class != &fair_sched_class)
goto simple;
/*
* Because of the set_next_buddy() in dequeue_task_fair() it is rather
* likely that a next task is from the same cgroup as the current.
*
* Therefore attempt to avoid putting and setting the entire cgroup
* hierarchy, only change the part that actually changes.
*/
do {
struct sched_entity *curr = cfs_rq->curr;
/*
* Since we got here without doing put_prev_entity() we also
* have to consider cfs_rq->curr. If it is still a runnable
* entity, update_curr() will update its vruntime, otherwise
* forget we've ever seen it.
*/
sched/fair: Prevent throttling in early pick_next_task_fair() The optimized task selection logic optimistically selects a new task to run without first doing a full put_prev_task(). This is so that we can avoid a put/set on the common ancestors of the old and new task. Similarly, we should only call check_cfs_rq_runtime() to throttle eligible groups if they're part of the common ancestry, otherwise it is possible to end up with no eligible task in the simple task selection. Imagine: /root /prev /next /A /B If our optimistic selection ends up throttling /next, we goto simple and our put_prev_task() ends up throttling /prev, after which we're going to bug out in set_next_entity() because there aren't any tasks left. Avoid this scenario by only throttling common ancestors. Reported-by: Mohammed Naser <mnaser@vexxhost.com> Reported-by: Konstantin Khlebnikov <khlebnikov@yandex-team.ru> Signed-off-by: Ben Segall <bsegall@google.com> [ munged Changelog ] Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Roman Gushchin <klamm@yandex-team.ru> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: pjt@google.com Fixes: 678d5718d8d0 ("sched/fair: Optimize cgroup pick_next_task_fair()") Link: http://lkml.kernel.org/r/xm26wq1oswoq.fsf@sword-of-the-dawn.mtv.corp.google.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-04-06 22:28:10 +00:00
if (curr) {
if (curr->on_rq)
update_curr(cfs_rq);
else
curr = NULL;
sched/fair: Prevent throttling in early pick_next_task_fair() The optimized task selection logic optimistically selects a new task to run without first doing a full put_prev_task(). This is so that we can avoid a put/set on the common ancestors of the old and new task. Similarly, we should only call check_cfs_rq_runtime() to throttle eligible groups if they're part of the common ancestry, otherwise it is possible to end up with no eligible task in the simple task selection. Imagine: /root /prev /next /A /B If our optimistic selection ends up throttling /next, we goto simple and our put_prev_task() ends up throttling /prev, after which we're going to bug out in set_next_entity() because there aren't any tasks left. Avoid this scenario by only throttling common ancestors. Reported-by: Mohammed Naser <mnaser@vexxhost.com> Reported-by: Konstantin Khlebnikov <khlebnikov@yandex-team.ru> Signed-off-by: Ben Segall <bsegall@google.com> [ munged Changelog ] Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Roman Gushchin <klamm@yandex-team.ru> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: pjt@google.com Fixes: 678d5718d8d0 ("sched/fair: Optimize cgroup pick_next_task_fair()") Link: http://lkml.kernel.org/r/xm26wq1oswoq.fsf@sword-of-the-dawn.mtv.corp.google.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-04-06 22:28:10 +00:00
/*
* This call to check_cfs_rq_runtime() will do the
* throttle and dequeue its entity in the parent(s).
* Therefore the nr_running test will indeed
sched/fair: Prevent throttling in early pick_next_task_fair() The optimized task selection logic optimistically selects a new task to run without first doing a full put_prev_task(). This is so that we can avoid a put/set on the common ancestors of the old and new task. Similarly, we should only call check_cfs_rq_runtime() to throttle eligible groups if they're part of the common ancestry, otherwise it is possible to end up with no eligible task in the simple task selection. Imagine: /root /prev /next /A /B If our optimistic selection ends up throttling /next, we goto simple and our put_prev_task() ends up throttling /prev, after which we're going to bug out in set_next_entity() because there aren't any tasks left. Avoid this scenario by only throttling common ancestors. Reported-by: Mohammed Naser <mnaser@vexxhost.com> Reported-by: Konstantin Khlebnikov <khlebnikov@yandex-team.ru> Signed-off-by: Ben Segall <bsegall@google.com> [ munged Changelog ] Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Roman Gushchin <klamm@yandex-team.ru> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: pjt@google.com Fixes: 678d5718d8d0 ("sched/fair: Optimize cgroup pick_next_task_fair()") Link: http://lkml.kernel.org/r/xm26wq1oswoq.fsf@sword-of-the-dawn.mtv.corp.google.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-04-06 22:28:10 +00:00
* be correct.
*/
if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
cfs_rq = &rq->cfs;
if (!cfs_rq->nr_running)
goto idle;
sched/fair: Prevent throttling in early pick_next_task_fair() The optimized task selection logic optimistically selects a new task to run without first doing a full put_prev_task(). This is so that we can avoid a put/set on the common ancestors of the old and new task. Similarly, we should only call check_cfs_rq_runtime() to throttle eligible groups if they're part of the common ancestry, otherwise it is possible to end up with no eligible task in the simple task selection. Imagine: /root /prev /next /A /B If our optimistic selection ends up throttling /next, we goto simple and our put_prev_task() ends up throttling /prev, after which we're going to bug out in set_next_entity() because there aren't any tasks left. Avoid this scenario by only throttling common ancestors. Reported-by: Mohammed Naser <mnaser@vexxhost.com> Reported-by: Konstantin Khlebnikov <khlebnikov@yandex-team.ru> Signed-off-by: Ben Segall <bsegall@google.com> [ munged Changelog ] Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Roman Gushchin <klamm@yandex-team.ru> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: pjt@google.com Fixes: 678d5718d8d0 ("sched/fair: Optimize cgroup pick_next_task_fair()") Link: http://lkml.kernel.org/r/xm26wq1oswoq.fsf@sword-of-the-dawn.mtv.corp.google.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-04-06 22:28:10 +00:00
goto simple;
}
sched/fair: Prevent throttling in early pick_next_task_fair() The optimized task selection logic optimistically selects a new task to run without first doing a full put_prev_task(). This is so that we can avoid a put/set on the common ancestors of the old and new task. Similarly, we should only call check_cfs_rq_runtime() to throttle eligible groups if they're part of the common ancestry, otherwise it is possible to end up with no eligible task in the simple task selection. Imagine: /root /prev /next /A /B If our optimistic selection ends up throttling /next, we goto simple and our put_prev_task() ends up throttling /prev, after which we're going to bug out in set_next_entity() because there aren't any tasks left. Avoid this scenario by only throttling common ancestors. Reported-by: Mohammed Naser <mnaser@vexxhost.com> Reported-by: Konstantin Khlebnikov <khlebnikov@yandex-team.ru> Signed-off-by: Ben Segall <bsegall@google.com> [ munged Changelog ] Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Roman Gushchin <klamm@yandex-team.ru> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: pjt@google.com Fixes: 678d5718d8d0 ("sched/fair: Optimize cgroup pick_next_task_fair()") Link: http://lkml.kernel.org/r/xm26wq1oswoq.fsf@sword-of-the-dawn.mtv.corp.google.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-04-06 22:28:10 +00:00
}
se = pick_next_entity(cfs_rq, curr);
cfs_rq = group_cfs_rq(se);
} while (cfs_rq);
p = task_of(se);
/*
* Since we haven't yet done put_prev_entity and if the selected task
* is a different task than we started out with, try and touch the
* least amount of cfs_rqs.
*/
if (prev != p) {
struct sched_entity *pse = &prev->se;
while (!(cfs_rq = is_same_group(se, pse))) {
int se_depth = se->depth;
int pse_depth = pse->depth;
if (se_depth <= pse_depth) {
put_prev_entity(cfs_rq_of(pse), pse);
pse = parent_entity(pse);
}
if (se_depth >= pse_depth) {
set_next_entity(cfs_rq_of(se), se);
se = parent_entity(se);
}
}
put_prev_entity(cfs_rq, pse);
set_next_entity(cfs_rq, se);
}
sched/fair: Search a task from the tail of the queue As a first step this patch makes cfs_tasks list as MRU one. It means, that when a next task is picked to run on physical CPU it is moved to the front of the list. Therefore, the cfs_tasks list is more or less sorted (except woken tasks) starting from recently given CPU time tasks toward tasks with max wait time in a run-queue, i.e. MRU list. Second, as part of the load balance operation, this approach starts detach_tasks()/detach_one_task() from the tail of the queue instead of the head, giving some advantages: - tends to pick a task with highest wait time; - tasks located in the tail are less likely cache-hot, therefore the can_migrate_task() decision is higher. hackbench illustrates slightly better performance. For example doing 1000 samples and 40 groups on i5-3320M CPU, it shows below figures: default: 0.657 avg patched: 0.646 avg Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Kirill Tkhai <tkhai@yandex.ru> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Mike Galbraith <umgwanakikbuti@gmail.com> Cc: Nicolas Pitre <nicolas.pitre@linaro.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Paul Turner <pjt@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Link: http://lkml.kernel.org/r/20170913102430.8985-2-urezki@gmail.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-09-13 10:24:30 +00:00
goto done;
simple:
#endif
if (prev)
put_prev_task(rq, prev);
do {
se = pick_next_entity(cfs_rq, NULL);
set_next_entity(cfs_rq, se);
cfs_rq = group_cfs_rq(se);
} while (cfs_rq);
p = task_of(se);
done: __maybe_unused;
sched/fair: Search a task from the tail of the queue As a first step this patch makes cfs_tasks list as MRU one. It means, that when a next task is picked to run on physical CPU it is moved to the front of the list. Therefore, the cfs_tasks list is more or less sorted (except woken tasks) starting from recently given CPU time tasks toward tasks with max wait time in a run-queue, i.e. MRU list. Second, as part of the load balance operation, this approach starts detach_tasks()/detach_one_task() from the tail of the queue instead of the head, giving some advantages: - tends to pick a task with highest wait time; - tasks located in the tail are less likely cache-hot, therefore the can_migrate_task() decision is higher. hackbench illustrates slightly better performance. For example doing 1000 samples and 40 groups on i5-3320M CPU, it shows below figures: default: 0.657 avg patched: 0.646 avg Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Kirill Tkhai <tkhai@yandex.ru> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Mike Galbraith <umgwanakikbuti@gmail.com> Cc: Nicolas Pitre <nicolas.pitre@linaro.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Paul Turner <pjt@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Link: http://lkml.kernel.org/r/20170913102430.8985-2-urezki@gmail.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-09-13 10:24:30 +00:00
#ifdef CONFIG_SMP
/*
* Move the next running task to the front of
* the list, so our cfs_tasks list becomes MRU
* one.
*/
list_move(&p->se.group_node, &rq->cfs_tasks);
#endif
sched: Save some hrtick_start_fair cycles hrtick_start_fair() shows up in profiles even when disabled. v3.0.6 taskset -c 3 pipe-test PerfTop: 997 irqs/sec kernel:89.5% exact: 0.0% [1000Hz cycles], (all, CPU: 3) ------------------------------------------------------------------------------------------------ Virgin Patched samples pcnt function samples pcnt function _______ _____ ___________________________ _______ _____ ___________________________ 2880.00 10.2% __schedule 3136.00 11.3% __schedule 1634.00 5.8% pipe_read 1615.00 5.8% pipe_read 1458.00 5.2% system_call 1534.00 5.5% system_call 1382.00 4.9% _raw_spin_lock_irqsave 1412.00 5.1% _raw_spin_lock_irqsave 1202.00 4.3% pipe_write 1255.00 4.5% copy_user_generic_string 1164.00 4.1% copy_user_generic_string 1241.00 4.5% __switch_to 1097.00 3.9% __switch_to 929.00 3.3% mutex_lock 872.00 3.1% mutex_lock 846.00 3.0% mutex_unlock 687.00 2.4% mutex_unlock 804.00 2.9% pipe_write 682.00 2.4% native_sched_clock 713.00 2.6% native_sched_clock 643.00 2.3% system_call_after_swapgs 653.00 2.3% _raw_spin_unlock_irqrestore 617.00 2.2% sched_clock_local 633.00 2.3% fsnotify 612.00 2.2% fsnotify 605.00 2.2% sched_clock_local 596.00 2.1% _raw_spin_unlock_irqrestore 593.00 2.1% system_call_after_swapgs 542.00 1.9% sysret_check 559.00 2.0% sysret_check 467.00 1.7% fget_light 472.00 1.7% fget_light 462.00 1.6% finish_task_switch 461.00 1.7% finish_task_switch 437.00 1.5% vfs_write 442.00 1.6% vfs_write 431.00 1.5% do_sync_write 428.00 1.5% do_sync_write 413.00 1.5% select_task_rq_fair 404.00 1.5% _raw_spin_lock_irq 386.00 1.4% update_curr 402.00 1.4% update_curr 385.00 1.4% rw_verify_area 389.00 1.4% do_sync_read 377.00 1.3% _raw_spin_lock_irq 378.00 1.4% vfs_read 369.00 1.3% do_sync_read 340.00 1.2% pipe_iov_copy_from_user 360.00 1.3% vfs_read 316.00 1.1% __wake_up_sync_key * 342.00 1.2% hrtick_start_fair 313.00 1.1% __wake_up_common Signed-off-by: Mike Galbraith <efault@gmx.de> [ fixed !CONFIG_SCHED_HRTICK borkage ] Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/1321971607.6855.17.camel@marge.simson.net Signed-off-by: Ingo Molnar <mingo@elte.hu>
2011-11-22 14:20:07 +00:00
if (hrtick_enabled(rq))
hrtick_start_fair(rq, p);
sched/fair: Add 'group_misfit_task' load-balance type To maximize throughput in systems with asymmetric CPU capacities (e.g. ARM big.LITTLE) load-balancing has to consider task and CPU utilization as well as per-CPU compute capacity when load-balancing in addition to the current average load based load-balancing policy. Tasks with high utilization that are scheduled on a lower capacity CPU need to be identified and migrated to a higher capacity CPU if possible to maximize throughput. To implement this additional policy an additional group_type (load-balance scenario) is added: 'group_misfit_task'. This represents scenarios where a sched_group has one or more tasks that are not suitable for its per-CPU capacity. 'group_misfit_task' is only considered if the system is not overloaded or imbalanced ('group_imbalanced' or 'group_overloaded'). Identifying misfit tasks requires the rq lock to be held. To avoid taking remote rq locks to examine source sched_groups for misfit tasks, each CPU is responsible for tracking misfit tasks themselves and update the rq->misfit_task flag. This means checking task utilization when tasks are scheduled and on sched_tick. Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: gaku.inami.xh@renesas.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1530699470-29808-3-git-send-email-morten.rasmussen@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-07-04 10:17:40 +00:00
update_misfit_status(p, rq);
return p;
idle:
if (!rf)
return NULL;
new_tasks = newidle_balance(rq, rf);
/*
* Because newidle_balance() releases (and re-acquires) rq->lock, it is
* possible for any higher priority task to appear. In that case we
* must re-start the pick_next_entity() loop.
*/
if (new_tasks < 0)
return RETRY_TASK;
if (new_tasks > 0)
goto again;
sched/fair: Update scale invariance of PELT The current implementation of load tracking invariance scales the contribution with current frequency and uarch performance (only for utilization) of the CPU. One main result of this formula is that the figures are capped by current capacity of CPU. Another one is that the load_avg is not invariant because not scaled with uarch. The util_avg of a periodic task that runs r time slots every p time slots varies in the range : U * (1-y^r)/(1-y^p) * y^i < Utilization < U * (1-y^r)/(1-y^p) with U is the max util_avg value = SCHED_CAPACITY_SCALE At a lower capacity, the range becomes: U * C * (1-y^r')/(1-y^p) * y^i' < Utilization < U * C * (1-y^r')/(1-y^p) with C reflecting the compute capacity ratio between current capacity and max capacity. so C tries to compensate changes in (1-y^r') but it can't be accurate. Instead of scaling the contribution value of PELT algo, we should scale the running time. The PELT signal aims to track the amount of computation of tasks and/or rq so it seems more correct to scale the running time to reflect the effective amount of computation done since the last update. In order to be fully invariant, we need to apply the same amount of running time and idle time whatever the current capacity. Because running at lower capacity implies that the task will run longer, we have to ensure that the same amount of idle time will be applied when system becomes idle and no idle time has been "stolen". But reaching the maximum utilization value (SCHED_CAPACITY_SCALE) means that the task is seen as an always-running task whatever the capacity of the CPU (even at max compute capacity). In this case, we can discard this "stolen" idle times which becomes meaningless. In order to achieve this time scaling, a new clock_pelt is created per rq. The increase of this clock scales with current capacity when something is running on rq and synchronizes with clock_task when rq is idle. With this mechanism, we ensure the same running and idle time whatever the current capacity. This also enables to simplify the pelt algorithm by removing all references of uarch and frequency and applying the same contribution to utilization and loads. Furthermore, the scaling is done only once per update of clock (update_rq_clock_task()) instead of during each update of sched_entities and cfs/rt/dl_rq of the rq like the current implementation. This is interesting when cgroup are involved as shown in the results below: On a hikey (octo Arm64 platform). Performance cpufreq governor and only shallowest c-state to remove variance generated by those power features so we only track the impact of pelt algo. each test runs 16 times: ./perf bench sched pipe (higher is better) kernel tip/sched/core + patch ops/seconds ops/seconds diff cgroup root 59652(+/- 0.18%) 59876(+/- 0.24%) +0.38% level1 55608(+/- 0.27%) 55923(+/- 0.24%) +0.57% level2 52115(+/- 0.29%) 52564(+/- 0.22%) +0.86% hackbench -l 1000 (lower is better) kernel tip/sched/core + patch duration(sec) duration(sec) diff cgroup root 4.453(+/- 2.37%) 4.383(+/- 2.88%) -1.57% level1 4.859(+/- 8.50%) 4.830(+/- 7.07%) -0.60% level2 5.063(+/- 9.83%) 4.928(+/- 9.66%) -2.66% Then, the responsiveness of PELT is improved when CPU is not running at max capacity with this new algorithm. I have put below some examples of duration to reach some typical load values according to the capacity of the CPU with current implementation and with this patch. These values has been computed based on the geometric series and the half period value: Util (%) max capacity half capacity(mainline) half capacity(w/ patch) 972 (95%) 138ms not reachable 276ms 486 (47.5%) 30ms 138ms 60ms 256 (25%) 13ms 32ms 26ms On my hikey (octo Arm64 platform) with schedutil governor, the time to reach max OPP when starting from a null utilization, decreases from 223ms with current scale invariance down to 121ms with the new algorithm. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: patrick.bellasi@arm.com Cc: pjt@google.com Cc: pkondeti@codeaurora.org Cc: quentin.perret@arm.com Cc: rjw@rjwysocki.net Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Link: https://lkml.kernel.org/r/1548257214-13745-3-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-01-23 15:26:53 +00:00
/*
* rq is about to be idle, check if we need to update the
* lost_idle_time of clock_pelt
*/
update_idle_rq_clock_pelt(rq);
return NULL;
}
static struct task_struct *__pick_next_task_fair(struct rq *rq)
{
return pick_next_task_fair(rq, NULL, NULL);
}
/*
* Account for a descheduled task:
*/
sched: Fix pick_next_task() vs 'change' pattern race Commit 67692435c411 ("sched: Rework pick_next_task() slow-path") inadvertly introduced a race because it changed a previously unexplored dependency between dropping the rq->lock and sched_class::put_prev_task(). The comments about dropping rq->lock, in for example newidle_balance(), only mentions the task being current and ->on_cpu being set. But when we look at the 'change' pattern (in for example sched_setnuma()): queued = task_on_rq_queued(p); /* p->on_rq == TASK_ON_RQ_QUEUED */ running = task_current(rq, p); /* rq->curr == p */ if (queued) dequeue_task(...); if (running) put_prev_task(...); /* change task properties */ if (queued) enqueue_task(...); if (running) set_next_task(...); It becomes obvious that if we do this after put_prev_task() has already been called on @p, things go sideways. This is exactly what the commit in question allows to happen when it does: prev->sched_class->put_prev_task(rq, prev, rf); if (!rq->nr_running) newidle_balance(rq, rf); The newidle_balance() call will drop rq->lock after we've called put_prev_task() and that allows the above 'change' pattern to interleave and mess up the state. Furthermore, it turns out we lost the RT-pull when we put the last DL task. Fix both problems by extracting the balancing from put_prev_task() and doing a multi-class balance() pass before put_prev_task(). Fixes: 67692435c411 ("sched: Rework pick_next_task() slow-path") Reported-by: Quentin Perret <qperret@google.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Tested-by: Quentin Perret <qperret@google.com> Tested-by: Valentin Schneider <valentin.schneider@arm.com>
2019-11-08 10:11:52 +00:00
static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
{
struct sched_entity *se = &prev->se;
struct cfs_rq *cfs_rq;
for_each_sched_entity(se) {
cfs_rq = cfs_rq_of(se);
put_prev_entity(cfs_rq, se);
}
}
/*
* sched_yield() is very simple
*
* The magic of dealing with the ->skip buddy is in pick_next_entity.
*/
static void yield_task_fair(struct rq *rq)
{
struct task_struct *curr = rq->curr;
struct cfs_rq *cfs_rq = task_cfs_rq(curr);
struct sched_entity *se = &curr->se;
/*
* Are we the only task in the tree?
*/
if (unlikely(rq->nr_running == 1))
return;
clear_buddies(cfs_rq, se);
if (curr->policy != SCHED_BATCH) {
update_rq_clock(rq);
/*
* Update run-time statistics of the 'current'.
*/
update_curr(cfs_rq);
/*
* Tell update_rq_clock() that we've just updated,
* so we don't do microscopic update in schedule()
* and double the fastpath cost.
*/
rq_clock_skip_update(rq);
}
set_skip_buddy(se);
}
static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
{
struct sched_entity *se = &p->se;
/* throttled hierarchies are not runnable */
if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
return false;
/* Tell the scheduler that we'd really like pse to run next. */
set_next_buddy(se);
yield_task_fair(rq);
return true;
}
#ifdef CONFIG_SMP
/**************************************************
* Fair scheduling class load-balancing methods.
*
* BASICS
*
* The purpose of load-balancing is to achieve the same basic fairness the
* per-CPU scheduler provides, namely provide a proportional amount of compute
* time to each task. This is expressed in the following equation:
*
* W_i,n/P_i == W_j,n/P_j for all i,j (1)
*
* Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
* W_i,0 is defined as:
*
* W_i,0 = \Sum_j w_i,j (2)
*
* Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
* is derived from the nice value as per sched_prio_to_weight[].
*
* The weight average is an exponential decay average of the instantaneous
* weight:
*
* W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
*
* C_i is the compute capacity of CPU i, typically it is the
* fraction of 'recent' time available for SCHED_OTHER task execution. But it
* can also include other factors [XXX].
*
* To achieve this balance we define a measure of imbalance which follows
* directly from (1):
*
* imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
*
* We them move tasks around to minimize the imbalance. In the continuous
* function space it is obvious this converges, in the discrete case we get
* a few fun cases generally called infeasible weight scenarios.
*
* [XXX expand on:
* - infeasible weights;
* - local vs global optima in the discrete case. ]
*
*
* SCHED DOMAINS
*
* In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
* for all i,j solution, we create a tree of CPUs that follows the hardware
* topology where each level pairs two lower groups (or better). This results
* in O(log n) layers. Furthermore we reduce the number of CPUs going up the
* tree to only the first of the previous level and we decrease the frequency
* of load-balance at each level inv. proportional to the number of CPUs in
* the groups.
*
* This yields:
*
* log_2 n 1 n
* \Sum { --- * --- * 2^i } = O(n) (5)
* i = 0 2^i 2^i
* `- size of each group
* | | `- number of CPUs doing load-balance
* | `- freq
* `- sum over all levels
*
* Coupled with a limit on how many tasks we can migrate every balance pass,
* this makes (5) the runtime complexity of the balancer.
*
* An important property here is that each CPU is still (indirectly) connected
* to every other CPU in at most O(log n) steps:
*
* The adjacency matrix of the resulting graph is given by:
*
* log_2 n
* A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
* k = 0
*
* And you'll find that:
*
* A^(log_2 n)_i,j != 0 for all i,j (7)
*
* Showing there's indeed a path between every CPU in at most O(log n) steps.
* The task movement gives a factor of O(m), giving a convergence complexity
* of:
*
* O(nm log n), n := nr_cpus, m := nr_tasks (8)
*
*
* WORK CONSERVING
*
* In order to avoid CPUs going idle while there's still work to do, new idle
* balancing is more aggressive and has the newly idle CPU iterate up the domain
* tree itself instead of relying on other CPUs to bring it work.
*
* This adds some complexity to both (5) and (8) but it reduces the total idle
* time.
*
* [XXX more?]
*
*
* CGROUPS
*
* Cgroups make a horror show out of (2), instead of a simple sum we get:
*
* s_k,i
* W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
* S_k
*
* Where
*
* s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
*
* w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
*
* The big problem is S_k, its a global sum needed to compute a local (W_i)
* property.
*
* [XXX write more on how we solve this.. _after_ merging pjt's patches that
* rewrite all of this once again.]
*/
static unsigned long __read_mostly max_load_balance_interval = HZ/10;
enum fbq_type { regular, remote, all };
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
/*
* 'group_type' describes the group of CPUs at the moment of load balancing.
*
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
* The enum is ordered by pulling priority, with the group with lowest priority
* first so the group_type can simply be compared when selecting the busiest
* group. See update_sd_pick_busiest().
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
*/
sched/fair: Add 'group_misfit_task' load-balance type To maximize throughput in systems with asymmetric CPU capacities (e.g. ARM big.LITTLE) load-balancing has to consider task and CPU utilization as well as per-CPU compute capacity when load-balancing in addition to the current average load based load-balancing policy. Tasks with high utilization that are scheduled on a lower capacity CPU need to be identified and migrated to a higher capacity CPU if possible to maximize throughput. To implement this additional policy an additional group_type (load-balance scenario) is added: 'group_misfit_task'. This represents scenarios where a sched_group has one or more tasks that are not suitable for its per-CPU capacity. 'group_misfit_task' is only considered if the system is not overloaded or imbalanced ('group_imbalanced' or 'group_overloaded'). Identifying misfit tasks requires the rq lock to be held. To avoid taking remote rq locks to examine source sched_groups for misfit tasks, each CPU is responsible for tracking misfit tasks themselves and update the rq->misfit_task flag. This means checking task utilization when tasks are scheduled and on sched_tick. Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: gaku.inami.xh@renesas.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1530699470-29808-3-git-send-email-morten.rasmussen@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-07-04 10:17:40 +00:00
enum group_type {
/* The group has spare capacity that can be used to run more tasks. */
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
group_has_spare = 0,
/*
* The group is fully used and the tasks don't compete for more CPU
* cycles. Nevertheless, some tasks might wait before running.
*/
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
group_fully_busy,
/*
* SD_ASYM_CPUCAPACITY only: One task doesn't fit with CPU's capacity
* and must be migrated to a more powerful CPU.
*/
sched/fair: Add 'group_misfit_task' load-balance type To maximize throughput in systems with asymmetric CPU capacities (e.g. ARM big.LITTLE) load-balancing has to consider task and CPU utilization as well as per-CPU compute capacity when load-balancing in addition to the current average load based load-balancing policy. Tasks with high utilization that are scheduled on a lower capacity CPU need to be identified and migrated to a higher capacity CPU if possible to maximize throughput. To implement this additional policy an additional group_type (load-balance scenario) is added: 'group_misfit_task'. This represents scenarios where a sched_group has one or more tasks that are not suitable for its per-CPU capacity. 'group_misfit_task' is only considered if the system is not overloaded or imbalanced ('group_imbalanced' or 'group_overloaded'). Identifying misfit tasks requires the rq lock to be held. To avoid taking remote rq locks to examine source sched_groups for misfit tasks, each CPU is responsible for tracking misfit tasks themselves and update the rq->misfit_task flag. This means checking task utilization when tasks are scheduled and on sched_tick. Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: gaku.inami.xh@renesas.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1530699470-29808-3-git-send-email-morten.rasmussen@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-07-04 10:17:40 +00:00
group_misfit_task,
/*
* SD_ASYM_PACKING only: One local CPU with higher capacity is available,
* and the task should be migrated to it instead of running on the
* current CPU.
*/
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
group_asym_packing,
/*
* The tasks' affinity constraints previously prevented the scheduler
* from balancing the load across the system.
*/
sched/fair: Add 'group_misfit_task' load-balance type To maximize throughput in systems with asymmetric CPU capacities (e.g. ARM big.LITTLE) load-balancing has to consider task and CPU utilization as well as per-CPU compute capacity when load-balancing in addition to the current average load based load-balancing policy. Tasks with high utilization that are scheduled on a lower capacity CPU need to be identified and migrated to a higher capacity CPU if possible to maximize throughput. To implement this additional policy an additional group_type (load-balance scenario) is added: 'group_misfit_task'. This represents scenarios where a sched_group has one or more tasks that are not suitable for its per-CPU capacity. 'group_misfit_task' is only considered if the system is not overloaded or imbalanced ('group_imbalanced' or 'group_overloaded'). Identifying misfit tasks requires the rq lock to be held. To avoid taking remote rq locks to examine source sched_groups for misfit tasks, each CPU is responsible for tracking misfit tasks themselves and update the rq->misfit_task flag. This means checking task utilization when tasks are scheduled and on sched_tick. Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: gaku.inami.xh@renesas.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1530699470-29808-3-git-send-email-morten.rasmussen@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-07-04 10:17:40 +00:00
group_imbalanced,
/*
* The CPU is overloaded and can't provide expected CPU cycles to all
* tasks.
*/
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
group_overloaded
};
enum migration_type {
migrate_load = 0,
migrate_util,
migrate_task,
migrate_misfit
sched/fair: Add 'group_misfit_task' load-balance type To maximize throughput in systems with asymmetric CPU capacities (e.g. ARM big.LITTLE) load-balancing has to consider task and CPU utilization as well as per-CPU compute capacity when load-balancing in addition to the current average load based load-balancing policy. Tasks with high utilization that are scheduled on a lower capacity CPU need to be identified and migrated to a higher capacity CPU if possible to maximize throughput. To implement this additional policy an additional group_type (load-balance scenario) is added: 'group_misfit_task'. This represents scenarios where a sched_group has one or more tasks that are not suitable for its per-CPU capacity. 'group_misfit_task' is only considered if the system is not overloaded or imbalanced ('group_imbalanced' or 'group_overloaded'). Identifying misfit tasks requires the rq lock to be held. To avoid taking remote rq locks to examine source sched_groups for misfit tasks, each CPU is responsible for tracking misfit tasks themselves and update the rq->misfit_task flag. This means checking task utilization when tasks are scheduled and on sched_tick. Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: gaku.inami.xh@renesas.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1530699470-29808-3-git-send-email-morten.rasmussen@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-07-04 10:17:40 +00:00
};
#define LBF_ALL_PINNED 0x01
#define LBF_NEED_BREAK 0x02
#define LBF_DST_PINNED 0x04
#define LBF_SOME_PINNED 0x08
#define LBF_NOHZ_STATS 0x10
#define LBF_NOHZ_AGAIN 0x20
struct lb_env {
struct sched_domain *sd;
struct rq *src_rq;
int src_cpu;
int dst_cpu;
struct rq *dst_rq;
sched: Improve balance_cpu() to consider other cpus in its group as target of (pinned) task Current load balance scheme requires only one cpu in a sched_group (balance_cpu) to look at other peer sched_groups for imbalance and pull tasks towards itself from a busy cpu. Tasks thus pulled by balance_cpu could later get picked up by cpus that are in the same sched_group as that of balance_cpu. This scheme however fails to pull tasks that are not allowed to run on balance_cpu (but are allowed to run on other cpus in its sched_group). That can affect fairness and in some worst case scenarios cause starvation. Consider a two core (2 threads/core) system running tasks as below: Core0 Core1 / \ / \ C0 C1 C2 C3 | | | | v v v v F0 T1 F1 [idle] T2 F0 = SCHED_FIFO task (pinned to C0) F1 = SCHED_FIFO task (pinned to C2) T1 = SCHED_OTHER task (pinned to C1) T2 = SCHED_OTHER task (pinned to C1 and C2) F1 could become a cpu hog, which will starve T2 unless C1 pulls it. Between C0 and C1 however, C0 is required to look for imbalance between cores, which will fail to pull T2 towards Core0. T2 will starve eternally in this case. The same scenario can arise in presence of non-rt tasks as well (say we replace F1 with high irq load). We tackle this problem by having balance_cpu move pinned tasks to one of its sibling cpus (where they can run). We first check if load balance goal can be met by ignoring pinned tasks, failing which we retry move_tasks() with a new env->dst_cpu. This patch modifies load balance semantics on who can move load towards a given cpu in a given sched_domain. Before this patch, a given_cpu or a ilb_cpu acting on behalf of an idle given_cpu is responsible for moving load to given_cpu. With this patch applied, balance_cpu can in addition decide on moving some load to a given_cpu. There is a remote possibility that excess load could get moved as a result of this (balance_cpu and given_cpu/ilb_cpu deciding *independently* and at *same* time to move some load to a given_cpu). However we should see less of such conflicting decisions in practice and moreover subsequent load balance cycles should correct the excess load moved to given_cpu. Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Signed-off-by: Prashanth Nageshappa <prashanth@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/4FE06CDB.2060605@linux.vnet.ibm.com [ minor edits ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2012-06-19 12:13:15 +00:00
struct cpumask *dst_grpmask;
int new_dst_cpu;
enum cpu_idle_type idle;
long imbalance;
/* The set of CPUs under consideration for load-balancing */
struct cpumask *cpus;
unsigned int flags;
unsigned int loop;
unsigned int loop_break;
unsigned int loop_max;
enum fbq_type fbq_type;
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
enum migration_type migration_type;
struct list_head tasks;
};
/*
* Is this task likely cache-hot:
*/
static int task_hot(struct task_struct *p, struct lb_env *env)
{
s64 delta;
lockdep_assert_held(&env->src_rq->lock);
if (p->sched_class != &fair_sched_class)
return 0;
if (unlikely(task_has_idle_policy(p)))
return 0;
/*
* Buddy candidates are cache hot:
*/
if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
(&p->se == cfs_rq_of(&p->se)->next ||
&p->se == cfs_rq_of(&p->se)->last))
return 1;
if (sysctl_sched_migration_cost == -1)
return 1;
if (sysctl_sched_migration_cost == 0)
return 0;
delta = rq_clock_task(env->src_rq) - p->se.exec_start;
return delta < (s64)sysctl_sched_migration_cost;
}
#ifdef CONFIG_NUMA_BALANCING
/*
* Returns 1, if task migration degrades locality
* Returns 0, if task migration improves locality i.e migration preferred.
* Returns -1, if task migration is not affected by locality.
*/
static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
{
struct numa_group *numa_group = rcu_dereference(p->numa_group);
sched/numa: Use group_weights to identify if migration degrades locality On NUMA_BACKPLANE and NUMA_GLUELESS_MESH systems, tasks/memory should be consolidated to the closest group of nodes. In such a case, relying on group_fault metric may not always help to consolidate. There can always be a case where a node closer to the preferred node may have lesser faults than a node further away from the preferred node. In such a case, moving to node with more faults might avoid numa consolidation. Using group_weight would help to consolidate task/memory around the preferred_node. While here, to be on the conservative side, don't override migrate thread degrades locality logic for CPU_NEWLY_IDLE load balancing. Note: Similar problems exist with should_numa_migrate_memory and will be dealt separately. Running SPECjbb2005 on a 4 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 16 25645.4 25960 1.22 1 72142 73550 1.95 Running SPECjbb2005 on a 16 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 8 110199 120071 8.958 1 176303 176249 -0.03 (numbers from v1 based on v4.17-rc5) Testcase Time: Min Max Avg StdDev numa01.sh Real: 490.04 774.86 596.26 96.46 numa01.sh Sys: 151.52 242.88 184.82 31.71 numa01.sh User: 41418.41 60844.59 48776.09 6564.27 numa02.sh Real: 60.14 62.94 60.98 1.00 numa02.sh Sys: 16.11 30.77 21.20 5.28 numa02.sh User: 5184.33 5311.09 5228.50 44.24 numa03.sh Real: 790.95 856.35 826.41 24.11 numa03.sh Sys: 114.93 118.85 117.05 1.63 numa03.sh User: 60990.99 64959.28 63470.43 1415.44 numa04.sh Real: 434.37 597.92 504.87 59.70 numa04.sh Sys: 237.63 397.40 289.74 55.98 numa04.sh User: 34854.87 41121.83 38572.52 2615.84 numa05.sh Real: 386.77 448.90 417.22 22.79 numa05.sh Sys: 149.23 379.95 303.04 79.55 numa05.sh User: 32951.76 35959.58 34562.18 1034.05 Testcase Time: Min Max Avg StdDev %Change numa01.sh Real: 493.19 672.88 597.51 59.38 -0.20% numa01.sh Sys: 150.09 245.48 207.76 34.26 -11.0% numa01.sh User: 41928.51 53779.17 48747.06 3901.39 0.059% numa02.sh Real: 60.63 62.87 61.22 0.83 -0.39% numa02.sh Sys: 16.64 27.97 20.25 4.06 4.691% numa02.sh User: 5222.92 5309.60 5254.03 29.98 -0.48% numa03.sh Real: 821.52 902.15 863.60 32.41 -4.30% numa03.sh Sys: 112.04 130.66 118.35 7.08 -1.09% numa03.sh User: 62245.16 69165.14 66443.04 2450.32 -4.47% numa04.sh Real: 414.53 519.57 476.25 37.00 6.009% numa04.sh Sys: 181.84 335.67 280.41 54.07 3.327% numa04.sh User: 33924.50 39115.39 37343.78 1934.26 3.290% numa05.sh Real: 408.30 441.45 417.90 12.05 -0.16% numa05.sh Sys: 233.41 381.60 295.58 57.37 2.523% numa05.sh User: 33301.31 35972.50 34335.19 938.94 0.661% Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1529514181-9842-16-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-20 17:02:56 +00:00
unsigned long src_weight, dst_weight;
int src_nid, dst_nid, dist;
if (!static_branch_likely(&sched_numa_balancing))
return -1;
if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
return -1;
src_nid = cpu_to_node(env->src_cpu);
dst_nid = cpu_to_node(env->dst_cpu);
if (src_nid == dst_nid)
return -1;
/* Migrating away from the preferred node is always bad. */
if (src_nid == p->numa_preferred_nid) {
if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
return 1;
else
return -1;
}
/* Encourage migration to the preferred node. */
if (dst_nid == p->numa_preferred_nid)
return 0;
/* Leaving a core idle is often worse than degrading locality. */
sched/numa: Use group_weights to identify if migration degrades locality On NUMA_BACKPLANE and NUMA_GLUELESS_MESH systems, tasks/memory should be consolidated to the closest group of nodes. In such a case, relying on group_fault metric may not always help to consolidate. There can always be a case where a node closer to the preferred node may have lesser faults than a node further away from the preferred node. In such a case, moving to node with more faults might avoid numa consolidation. Using group_weight would help to consolidate task/memory around the preferred_node. While here, to be on the conservative side, don't override migrate thread degrades locality logic for CPU_NEWLY_IDLE load balancing. Note: Similar problems exist with should_numa_migrate_memory and will be dealt separately. Running SPECjbb2005 on a 4 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 16 25645.4 25960 1.22 1 72142 73550 1.95 Running SPECjbb2005 on a 16 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 8 110199 120071 8.958 1 176303 176249 -0.03 (numbers from v1 based on v4.17-rc5) Testcase Time: Min Max Avg StdDev numa01.sh Real: 490.04 774.86 596.26 96.46 numa01.sh Sys: 151.52 242.88 184.82 31.71 numa01.sh User: 41418.41 60844.59 48776.09 6564.27 numa02.sh Real: 60.14 62.94 60.98 1.00 numa02.sh Sys: 16.11 30.77 21.20 5.28 numa02.sh User: 5184.33 5311.09 5228.50 44.24 numa03.sh Real: 790.95 856.35 826.41 24.11 numa03.sh Sys: 114.93 118.85 117.05 1.63 numa03.sh User: 60990.99 64959.28 63470.43 1415.44 numa04.sh Real: 434.37 597.92 504.87 59.70 numa04.sh Sys: 237.63 397.40 289.74 55.98 numa04.sh User: 34854.87 41121.83 38572.52 2615.84 numa05.sh Real: 386.77 448.90 417.22 22.79 numa05.sh Sys: 149.23 379.95 303.04 79.55 numa05.sh User: 32951.76 35959.58 34562.18 1034.05 Testcase Time: Min Max Avg StdDev %Change numa01.sh Real: 493.19 672.88 597.51 59.38 -0.20% numa01.sh Sys: 150.09 245.48 207.76 34.26 -11.0% numa01.sh User: 41928.51 53779.17 48747.06 3901.39 0.059% numa02.sh Real: 60.63 62.87 61.22 0.83 -0.39% numa02.sh Sys: 16.64 27.97 20.25 4.06 4.691% numa02.sh User: 5222.92 5309.60 5254.03 29.98 -0.48% numa03.sh Real: 821.52 902.15 863.60 32.41 -4.30% numa03.sh Sys: 112.04 130.66 118.35 7.08 -1.09% numa03.sh User: 62245.16 69165.14 66443.04 2450.32 -4.47% numa04.sh Real: 414.53 519.57 476.25 37.00 6.009% numa04.sh Sys: 181.84 335.67 280.41 54.07 3.327% numa04.sh User: 33924.50 39115.39 37343.78 1934.26 3.290% numa05.sh Real: 408.30 441.45 417.90 12.05 -0.16% numa05.sh Sys: 233.41 381.60 295.58 57.37 2.523% numa05.sh User: 33301.31 35972.50 34335.19 938.94 0.661% Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1529514181-9842-16-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-20 17:02:56 +00:00
if (env->idle == CPU_IDLE)
return -1;
sched/numa: Use group_weights to identify if migration degrades locality On NUMA_BACKPLANE and NUMA_GLUELESS_MESH systems, tasks/memory should be consolidated to the closest group of nodes. In such a case, relying on group_fault metric may not always help to consolidate. There can always be a case where a node closer to the preferred node may have lesser faults than a node further away from the preferred node. In such a case, moving to node with more faults might avoid numa consolidation. Using group_weight would help to consolidate task/memory around the preferred_node. While here, to be on the conservative side, don't override migrate thread degrades locality logic for CPU_NEWLY_IDLE load balancing. Note: Similar problems exist with should_numa_migrate_memory and will be dealt separately. Running SPECjbb2005 on a 4 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 16 25645.4 25960 1.22 1 72142 73550 1.95 Running SPECjbb2005 on a 16 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 8 110199 120071 8.958 1 176303 176249 -0.03 (numbers from v1 based on v4.17-rc5) Testcase Time: Min Max Avg StdDev numa01.sh Real: 490.04 774.86 596.26 96.46 numa01.sh Sys: 151.52 242.88 184.82 31.71 numa01.sh User: 41418.41 60844.59 48776.09 6564.27 numa02.sh Real: 60.14 62.94 60.98 1.00 numa02.sh Sys: 16.11 30.77 21.20 5.28 numa02.sh User: 5184.33 5311.09 5228.50 44.24 numa03.sh Real: 790.95 856.35 826.41 24.11 numa03.sh Sys: 114.93 118.85 117.05 1.63 numa03.sh User: 60990.99 64959.28 63470.43 1415.44 numa04.sh Real: 434.37 597.92 504.87 59.70 numa04.sh Sys: 237.63 397.40 289.74 55.98 numa04.sh User: 34854.87 41121.83 38572.52 2615.84 numa05.sh Real: 386.77 448.90 417.22 22.79 numa05.sh Sys: 149.23 379.95 303.04 79.55 numa05.sh User: 32951.76 35959.58 34562.18 1034.05 Testcase Time: Min Max Avg StdDev %Change numa01.sh Real: 493.19 672.88 597.51 59.38 -0.20% numa01.sh Sys: 150.09 245.48 207.76 34.26 -11.0% numa01.sh User: 41928.51 53779.17 48747.06 3901.39 0.059% numa02.sh Real: 60.63 62.87 61.22 0.83 -0.39% numa02.sh Sys: 16.64 27.97 20.25 4.06 4.691% numa02.sh User: 5222.92 5309.60 5254.03 29.98 -0.48% numa03.sh Real: 821.52 902.15 863.60 32.41 -4.30% numa03.sh Sys: 112.04 130.66 118.35 7.08 -1.09% numa03.sh User: 62245.16 69165.14 66443.04 2450.32 -4.47% numa04.sh Real: 414.53 519.57 476.25 37.00 6.009% numa04.sh Sys: 181.84 335.67 280.41 54.07 3.327% numa04.sh User: 33924.50 39115.39 37343.78 1934.26 3.290% numa05.sh Real: 408.30 441.45 417.90 12.05 -0.16% numa05.sh Sys: 233.41 381.60 295.58 57.37 2.523% numa05.sh User: 33301.31 35972.50 34335.19 938.94 0.661% Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1529514181-9842-16-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-20 17:02:56 +00:00
dist = node_distance(src_nid, dst_nid);
if (numa_group) {
sched/numa: Use group_weights to identify if migration degrades locality On NUMA_BACKPLANE and NUMA_GLUELESS_MESH systems, tasks/memory should be consolidated to the closest group of nodes. In such a case, relying on group_fault metric may not always help to consolidate. There can always be a case where a node closer to the preferred node may have lesser faults than a node further away from the preferred node. In such a case, moving to node with more faults might avoid numa consolidation. Using group_weight would help to consolidate task/memory around the preferred_node. While here, to be on the conservative side, don't override migrate thread degrades locality logic for CPU_NEWLY_IDLE load balancing. Note: Similar problems exist with should_numa_migrate_memory and will be dealt separately. Running SPECjbb2005 on a 4 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 16 25645.4 25960 1.22 1 72142 73550 1.95 Running SPECjbb2005 on a 16 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 8 110199 120071 8.958 1 176303 176249 -0.03 (numbers from v1 based on v4.17-rc5) Testcase Time: Min Max Avg StdDev numa01.sh Real: 490.04 774.86 596.26 96.46 numa01.sh Sys: 151.52 242.88 184.82 31.71 numa01.sh User: 41418.41 60844.59 48776.09 6564.27 numa02.sh Real: 60.14 62.94 60.98 1.00 numa02.sh Sys: 16.11 30.77 21.20 5.28 numa02.sh User: 5184.33 5311.09 5228.50 44.24 numa03.sh Real: 790.95 856.35 826.41 24.11 numa03.sh Sys: 114.93 118.85 117.05 1.63 numa03.sh User: 60990.99 64959.28 63470.43 1415.44 numa04.sh Real: 434.37 597.92 504.87 59.70 numa04.sh Sys: 237.63 397.40 289.74 55.98 numa04.sh User: 34854.87 41121.83 38572.52 2615.84 numa05.sh Real: 386.77 448.90 417.22 22.79 numa05.sh Sys: 149.23 379.95 303.04 79.55 numa05.sh User: 32951.76 35959.58 34562.18 1034.05 Testcase Time: Min Max Avg StdDev %Change numa01.sh Real: 493.19 672.88 597.51 59.38 -0.20% numa01.sh Sys: 150.09 245.48 207.76 34.26 -11.0% numa01.sh User: 41928.51 53779.17 48747.06 3901.39 0.059% numa02.sh Real: 60.63 62.87 61.22 0.83 -0.39% numa02.sh Sys: 16.64 27.97 20.25 4.06 4.691% numa02.sh User: 5222.92 5309.60 5254.03 29.98 -0.48% numa03.sh Real: 821.52 902.15 863.60 32.41 -4.30% numa03.sh Sys: 112.04 130.66 118.35 7.08 -1.09% numa03.sh User: 62245.16 69165.14 66443.04 2450.32 -4.47% numa04.sh Real: 414.53 519.57 476.25 37.00 6.009% numa04.sh Sys: 181.84 335.67 280.41 54.07 3.327% numa04.sh User: 33924.50 39115.39 37343.78 1934.26 3.290% numa05.sh Real: 408.30 441.45 417.90 12.05 -0.16% numa05.sh Sys: 233.41 381.60 295.58 57.37 2.523% numa05.sh User: 33301.31 35972.50 34335.19 938.94 0.661% Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1529514181-9842-16-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-20 17:02:56 +00:00
src_weight = group_weight(p, src_nid, dist);
dst_weight = group_weight(p, dst_nid, dist);
} else {
sched/numa: Use group_weights to identify if migration degrades locality On NUMA_BACKPLANE and NUMA_GLUELESS_MESH systems, tasks/memory should be consolidated to the closest group of nodes. In such a case, relying on group_fault metric may not always help to consolidate. There can always be a case where a node closer to the preferred node may have lesser faults than a node further away from the preferred node. In such a case, moving to node with more faults might avoid numa consolidation. Using group_weight would help to consolidate task/memory around the preferred_node. While here, to be on the conservative side, don't override migrate thread degrades locality logic for CPU_NEWLY_IDLE load balancing. Note: Similar problems exist with should_numa_migrate_memory and will be dealt separately. Running SPECjbb2005 on a 4 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 16 25645.4 25960 1.22 1 72142 73550 1.95 Running SPECjbb2005 on a 16 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 8 110199 120071 8.958 1 176303 176249 -0.03 (numbers from v1 based on v4.17-rc5) Testcase Time: Min Max Avg StdDev numa01.sh Real: 490.04 774.86 596.26 96.46 numa01.sh Sys: 151.52 242.88 184.82 31.71 numa01.sh User: 41418.41 60844.59 48776.09 6564.27 numa02.sh Real: 60.14 62.94 60.98 1.00 numa02.sh Sys: 16.11 30.77 21.20 5.28 numa02.sh User: 5184.33 5311.09 5228.50 44.24 numa03.sh Real: 790.95 856.35 826.41 24.11 numa03.sh Sys: 114.93 118.85 117.05 1.63 numa03.sh User: 60990.99 64959.28 63470.43 1415.44 numa04.sh Real: 434.37 597.92 504.87 59.70 numa04.sh Sys: 237.63 397.40 289.74 55.98 numa04.sh User: 34854.87 41121.83 38572.52 2615.84 numa05.sh Real: 386.77 448.90 417.22 22.79 numa05.sh Sys: 149.23 379.95 303.04 79.55 numa05.sh User: 32951.76 35959.58 34562.18 1034.05 Testcase Time: Min Max Avg StdDev %Change numa01.sh Real: 493.19 672.88 597.51 59.38 -0.20% numa01.sh Sys: 150.09 245.48 207.76 34.26 -11.0% numa01.sh User: 41928.51 53779.17 48747.06 3901.39 0.059% numa02.sh Real: 60.63 62.87 61.22 0.83 -0.39% numa02.sh Sys: 16.64 27.97 20.25 4.06 4.691% numa02.sh User: 5222.92 5309.60 5254.03 29.98 -0.48% numa03.sh Real: 821.52 902.15 863.60 32.41 -4.30% numa03.sh Sys: 112.04 130.66 118.35 7.08 -1.09% numa03.sh User: 62245.16 69165.14 66443.04 2450.32 -4.47% numa04.sh Real: 414.53 519.57 476.25 37.00 6.009% numa04.sh Sys: 181.84 335.67 280.41 54.07 3.327% numa04.sh User: 33924.50 39115.39 37343.78 1934.26 3.290% numa05.sh Real: 408.30 441.45 417.90 12.05 -0.16% numa05.sh Sys: 233.41 381.60 295.58 57.37 2.523% numa05.sh User: 33301.31 35972.50 34335.19 938.94 0.661% Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1529514181-9842-16-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-20 17:02:56 +00:00
src_weight = task_weight(p, src_nid, dist);
dst_weight = task_weight(p, dst_nid, dist);
}
sched/numa: Use group_weights to identify if migration degrades locality On NUMA_BACKPLANE and NUMA_GLUELESS_MESH systems, tasks/memory should be consolidated to the closest group of nodes. In such a case, relying on group_fault metric may not always help to consolidate. There can always be a case where a node closer to the preferred node may have lesser faults than a node further away from the preferred node. In such a case, moving to node with more faults might avoid numa consolidation. Using group_weight would help to consolidate task/memory around the preferred_node. While here, to be on the conservative side, don't override migrate thread degrades locality logic for CPU_NEWLY_IDLE load balancing. Note: Similar problems exist with should_numa_migrate_memory and will be dealt separately. Running SPECjbb2005 on a 4 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 16 25645.4 25960 1.22 1 72142 73550 1.95 Running SPECjbb2005 on a 16 node machine and comparing bops/JVM JVMS LAST_PATCH WITH_PATCH %CHANGE 8 110199 120071 8.958 1 176303 176249 -0.03 (numbers from v1 based on v4.17-rc5) Testcase Time: Min Max Avg StdDev numa01.sh Real: 490.04 774.86 596.26 96.46 numa01.sh Sys: 151.52 242.88 184.82 31.71 numa01.sh User: 41418.41 60844.59 48776.09 6564.27 numa02.sh Real: 60.14 62.94 60.98 1.00 numa02.sh Sys: 16.11 30.77 21.20 5.28 numa02.sh User: 5184.33 5311.09 5228.50 44.24 numa03.sh Real: 790.95 856.35 826.41 24.11 numa03.sh Sys: 114.93 118.85 117.05 1.63 numa03.sh User: 60990.99 64959.28 63470.43 1415.44 numa04.sh Real: 434.37 597.92 504.87 59.70 numa04.sh Sys: 237.63 397.40 289.74 55.98 numa04.sh User: 34854.87 41121.83 38572.52 2615.84 numa05.sh Real: 386.77 448.90 417.22 22.79 numa05.sh Sys: 149.23 379.95 303.04 79.55 numa05.sh User: 32951.76 35959.58 34562.18 1034.05 Testcase Time: Min Max Avg StdDev %Change numa01.sh Real: 493.19 672.88 597.51 59.38 -0.20% numa01.sh Sys: 150.09 245.48 207.76 34.26 -11.0% numa01.sh User: 41928.51 53779.17 48747.06 3901.39 0.059% numa02.sh Real: 60.63 62.87 61.22 0.83 -0.39% numa02.sh Sys: 16.64 27.97 20.25 4.06 4.691% numa02.sh User: 5222.92 5309.60 5254.03 29.98 -0.48% numa03.sh Real: 821.52 902.15 863.60 32.41 -4.30% numa03.sh Sys: 112.04 130.66 118.35 7.08 -1.09% numa03.sh User: 62245.16 69165.14 66443.04 2450.32 -4.47% numa04.sh Real: 414.53 519.57 476.25 37.00 6.009% numa04.sh Sys: 181.84 335.67 280.41 54.07 3.327% numa04.sh User: 33924.50 39115.39 37343.78 1934.26 3.290% numa05.sh Real: 408.30 441.45 417.90 12.05 -0.16% numa05.sh Sys: 233.41 381.60 295.58 57.37 2.523% numa05.sh User: 33301.31 35972.50 34335.19 938.94 0.661% Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1529514181-9842-16-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-20 17:02:56 +00:00
return dst_weight < src_weight;
}
#else
static inline int migrate_degrades_locality(struct task_struct *p,
struct lb_env *env)
{
return -1;
}
#endif
/*
* can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
*/
static
int can_migrate_task(struct task_struct *p, struct lb_env *env)
{
int tsk_cache_hot;
lockdep_assert_held(&env->src_rq->lock);
/*
* We do not migrate tasks that are:
* 1) throttled_lb_pair, or
* 2) cannot be migrated to this CPU due to cpus_ptr, or
* 3) running (obviously), or
* 4) are cache-hot on their current CPU.
*/
if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
return 0;
if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) {
int cpu;
sched: Improve balance_cpu() to consider other cpus in its group as target of (pinned) task Current load balance scheme requires only one cpu in a sched_group (balance_cpu) to look at other peer sched_groups for imbalance and pull tasks towards itself from a busy cpu. Tasks thus pulled by balance_cpu could later get picked up by cpus that are in the same sched_group as that of balance_cpu. This scheme however fails to pull tasks that are not allowed to run on balance_cpu (but are allowed to run on other cpus in its sched_group). That can affect fairness and in some worst case scenarios cause starvation. Consider a two core (2 threads/core) system running tasks as below: Core0 Core1 / \ / \ C0 C1 C2 C3 | | | | v v v v F0 T1 F1 [idle] T2 F0 = SCHED_FIFO task (pinned to C0) F1 = SCHED_FIFO task (pinned to C2) T1 = SCHED_OTHER task (pinned to C1) T2 = SCHED_OTHER task (pinned to C1 and C2) F1 could become a cpu hog, which will starve T2 unless C1 pulls it. Between C0 and C1 however, C0 is required to look for imbalance between cores, which will fail to pull T2 towards Core0. T2 will starve eternally in this case. The same scenario can arise in presence of non-rt tasks as well (say we replace F1 with high irq load). We tackle this problem by having balance_cpu move pinned tasks to one of its sibling cpus (where they can run). We first check if load balance goal can be met by ignoring pinned tasks, failing which we retry move_tasks() with a new env->dst_cpu. This patch modifies load balance semantics on who can move load towards a given cpu in a given sched_domain. Before this patch, a given_cpu or a ilb_cpu acting on behalf of an idle given_cpu is responsible for moving load to given_cpu. With this patch applied, balance_cpu can in addition decide on moving some load to a given_cpu. There is a remote possibility that excess load could get moved as a result of this (balance_cpu and given_cpu/ilb_cpu deciding *independently* and at *same* time to move some load to a given_cpu). However we should see less of such conflicting decisions in practice and moreover subsequent load balance cycles should correct the excess load moved to given_cpu. Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Signed-off-by: Prashanth Nageshappa <prashanth@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/4FE06CDB.2060605@linux.vnet.ibm.com [ minor edits ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2012-06-19 12:13:15 +00:00
schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
sched: Improve balance_cpu() to consider other cpus in its group as target of (pinned) task Current load balance scheme requires only one cpu in a sched_group (balance_cpu) to look at other peer sched_groups for imbalance and pull tasks towards itself from a busy cpu. Tasks thus pulled by balance_cpu could later get picked up by cpus that are in the same sched_group as that of balance_cpu. This scheme however fails to pull tasks that are not allowed to run on balance_cpu (but are allowed to run on other cpus in its sched_group). That can affect fairness and in some worst case scenarios cause starvation. Consider a two core (2 threads/core) system running tasks as below: Core0 Core1 / \ / \ C0 C1 C2 C3 | | | | v v v v F0 T1 F1 [idle] T2 F0 = SCHED_FIFO task (pinned to C0) F1 = SCHED_FIFO task (pinned to C2) T1 = SCHED_OTHER task (pinned to C1) T2 = SCHED_OTHER task (pinned to C1 and C2) F1 could become a cpu hog, which will starve T2 unless C1 pulls it. Between C0 and C1 however, C0 is required to look for imbalance between cores, which will fail to pull T2 towards Core0. T2 will starve eternally in this case. The same scenario can arise in presence of non-rt tasks as well (say we replace F1 with high irq load). We tackle this problem by having balance_cpu move pinned tasks to one of its sibling cpus (where they can run). We first check if load balance goal can be met by ignoring pinned tasks, failing which we retry move_tasks() with a new env->dst_cpu. This patch modifies load balance semantics on who can move load towards a given cpu in a given sched_domain. Before this patch, a given_cpu or a ilb_cpu acting on behalf of an idle given_cpu is responsible for moving load to given_cpu. With this patch applied, balance_cpu can in addition decide on moving some load to a given_cpu. There is a remote possibility that excess load could get moved as a result of this (balance_cpu and given_cpu/ilb_cpu deciding *independently* and at *same* time to move some load to a given_cpu). However we should see less of such conflicting decisions in practice and moreover subsequent load balance cycles should correct the excess load moved to given_cpu. Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Signed-off-by: Prashanth Nageshappa <prashanth@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/4FE06CDB.2060605@linux.vnet.ibm.com [ minor edits ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2012-06-19 12:13:15 +00:00
env->flags |= LBF_SOME_PINNED;
sched: Improve balance_cpu() to consider other cpus in its group as target of (pinned) task Current load balance scheme requires only one cpu in a sched_group (balance_cpu) to look at other peer sched_groups for imbalance and pull tasks towards itself from a busy cpu. Tasks thus pulled by balance_cpu could later get picked up by cpus that are in the same sched_group as that of balance_cpu. This scheme however fails to pull tasks that are not allowed to run on balance_cpu (but are allowed to run on other cpus in its sched_group). That can affect fairness and in some worst case scenarios cause starvation. Consider a two core (2 threads/core) system running tasks as below: Core0 Core1 / \ / \ C0 C1 C2 C3 | | | | v v v v F0 T1 F1 [idle] T2 F0 = SCHED_FIFO task (pinned to C0) F1 = SCHED_FIFO task (pinned to C2) T1 = SCHED_OTHER task (pinned to C1) T2 = SCHED_OTHER task (pinned to C1 and C2) F1 could become a cpu hog, which will starve T2 unless C1 pulls it. Between C0 and C1 however, C0 is required to look for imbalance between cores, which will fail to pull T2 towards Core0. T2 will starve eternally in this case. The same scenario can arise in presence of non-rt tasks as well (say we replace F1 with high irq load). We tackle this problem by having balance_cpu move pinned tasks to one of its sibling cpus (where they can run). We first check if load balance goal can be met by ignoring pinned tasks, failing which we retry move_tasks() with a new env->dst_cpu. This patch modifies load balance semantics on who can move load towards a given cpu in a given sched_domain. Before this patch, a given_cpu or a ilb_cpu acting on behalf of an idle given_cpu is responsible for moving load to given_cpu. With this patch applied, balance_cpu can in addition decide on moving some load to a given_cpu. There is a remote possibility that excess load could get moved as a result of this (balance_cpu and given_cpu/ilb_cpu deciding *independently* and at *same* time to move some load to a given_cpu). However we should see less of such conflicting decisions in practice and moreover subsequent load balance cycles should correct the excess load moved to given_cpu. Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Signed-off-by: Prashanth Nageshappa <prashanth@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/4FE06CDB.2060605@linux.vnet.ibm.com [ minor edits ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2012-06-19 12:13:15 +00:00
/*
* Remember if this task can be migrated to any other CPU in
sched: Improve balance_cpu() to consider other cpus in its group as target of (pinned) task Current load balance scheme requires only one cpu in a sched_group (balance_cpu) to look at other peer sched_groups for imbalance and pull tasks towards itself from a busy cpu. Tasks thus pulled by balance_cpu could later get picked up by cpus that are in the same sched_group as that of balance_cpu. This scheme however fails to pull tasks that are not allowed to run on balance_cpu (but are allowed to run on other cpus in its sched_group). That can affect fairness and in some worst case scenarios cause starvation. Consider a two core (2 threads/core) system running tasks as below: Core0 Core1 / \ / \ C0 C1 C2 C3 | | | | v v v v F0 T1 F1 [idle] T2 F0 = SCHED_FIFO task (pinned to C0) F1 = SCHED_FIFO task (pinned to C2) T1 = SCHED_OTHER task (pinned to C1) T2 = SCHED_OTHER task (pinned to C1 and C2) F1 could become a cpu hog, which will starve T2 unless C1 pulls it. Between C0 and C1 however, C0 is required to look for imbalance between cores, which will fail to pull T2 towards Core0. T2 will starve eternally in this case. The same scenario can arise in presence of non-rt tasks as well (say we replace F1 with high irq load). We tackle this problem by having balance_cpu move pinned tasks to one of its sibling cpus (where they can run). We first check if load balance goal can be met by ignoring pinned tasks, failing which we retry move_tasks() with a new env->dst_cpu. This patch modifies load balance semantics on who can move load towards a given cpu in a given sched_domain. Before this patch, a given_cpu or a ilb_cpu acting on behalf of an idle given_cpu is responsible for moving load to given_cpu. With this patch applied, balance_cpu can in addition decide on moving some load to a given_cpu. There is a remote possibility that excess load could get moved as a result of this (balance_cpu and given_cpu/ilb_cpu deciding *independently* and at *same* time to move some load to a given_cpu). However we should see less of such conflicting decisions in practice and moreover subsequent load balance cycles should correct the excess load moved to given_cpu. Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Signed-off-by: Prashanth Nageshappa <prashanth@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/4FE06CDB.2060605@linux.vnet.ibm.com [ minor edits ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2012-06-19 12:13:15 +00:00
* our sched_group. We may want to revisit it if we couldn't
* meet load balance goals by pulling other tasks on src_cpu.
*
sched/fair: Fix load_balance() affinity redo path If load_balance() fails to migrate any tasks because all tasks were affined, load_balance() removes the source CPU from consideration and attempts to redo and balance among the new subset of CPUs. There is a bug in this code path where the algorithm considers all active CPUs in the system (minus the source that was just masked out). This is not valid for two reasons: some active CPUs may not be in the current scheduling domain and one of the active CPUs is dst_cpu. These CPUs should not be considered, as we cannot pull load from them. Instead of failing out of load_balance(), we may end up redoing the search with no valid CPUs and incorrectly concluding the domain is balanced. Additionally, if the group_imbalance flag was just set, it may also be incorrectly unset, thus the flag will not be seen by other CPUs in future load_balance() runs as that algorithm intends. Fix the check by removing CPUs not in the current domain and the dst_cpu from considertation, thus limiting the evaluation to valid remaining CPUs from which load might be migrated. Co-authored-by: Austin Christ <austinwc@codeaurora.org> Co-authored-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Tested-by: Tyler Baicar <tbaicar@codeaurora.org> Signed-off-by: Jeffrey Hugo <jhugo@codeaurora.org> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Austin Christ <austinwc@codeaurora.org> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Timur Tabi <timur@codeaurora.org> Link: http://lkml.kernel.org/r/1496863138-11322-2-git-send-email-jhugo@codeaurora.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-06-07 19:18:57 +00:00
* Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
* already computed one in current iteration.
sched: Improve balance_cpu() to consider other cpus in its group as target of (pinned) task Current load balance scheme requires only one cpu in a sched_group (balance_cpu) to look at other peer sched_groups for imbalance and pull tasks towards itself from a busy cpu. Tasks thus pulled by balance_cpu could later get picked up by cpus that are in the same sched_group as that of balance_cpu. This scheme however fails to pull tasks that are not allowed to run on balance_cpu (but are allowed to run on other cpus in its sched_group). That can affect fairness and in some worst case scenarios cause starvation. Consider a two core (2 threads/core) system running tasks as below: Core0 Core1 / \ / \ C0 C1 C2 C3 | | | | v v v v F0 T1 F1 [idle] T2 F0 = SCHED_FIFO task (pinned to C0) F1 = SCHED_FIFO task (pinned to C2) T1 = SCHED_OTHER task (pinned to C1) T2 = SCHED_OTHER task (pinned to C1 and C2) F1 could become a cpu hog, which will starve T2 unless C1 pulls it. Between C0 and C1 however, C0 is required to look for imbalance between cores, which will fail to pull T2 towards Core0. T2 will starve eternally in this case. The same scenario can arise in presence of non-rt tasks as well (say we replace F1 with high irq load). We tackle this problem by having balance_cpu move pinned tasks to one of its sibling cpus (where they can run). We first check if load balance goal can be met by ignoring pinned tasks, failing which we retry move_tasks() with a new env->dst_cpu. This patch modifies load balance semantics on who can move load towards a given cpu in a given sched_domain. Before this patch, a given_cpu or a ilb_cpu acting on behalf of an idle given_cpu is responsible for moving load to given_cpu. With this patch applied, balance_cpu can in addition decide on moving some load to a given_cpu. There is a remote possibility that excess load could get moved as a result of this (balance_cpu and given_cpu/ilb_cpu deciding *independently* and at *same* time to move some load to a given_cpu). However we should see less of such conflicting decisions in practice and moreover subsequent load balance cycles should correct the excess load moved to given_cpu. Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Signed-off-by: Prashanth Nageshappa <prashanth@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/4FE06CDB.2060605@linux.vnet.ibm.com [ minor edits ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2012-06-19 12:13:15 +00:00
*/
sched/fair: Fix load_balance() affinity redo path If load_balance() fails to migrate any tasks because all tasks were affined, load_balance() removes the source CPU from consideration and attempts to redo and balance among the new subset of CPUs. There is a bug in this code path where the algorithm considers all active CPUs in the system (minus the source that was just masked out). This is not valid for two reasons: some active CPUs may not be in the current scheduling domain and one of the active CPUs is dst_cpu. These CPUs should not be considered, as we cannot pull load from them. Instead of failing out of load_balance(), we may end up redoing the search with no valid CPUs and incorrectly concluding the domain is balanced. Additionally, if the group_imbalance flag was just set, it may also be incorrectly unset, thus the flag will not be seen by other CPUs in future load_balance() runs as that algorithm intends. Fix the check by removing CPUs not in the current domain and the dst_cpu from considertation, thus limiting the evaluation to valid remaining CPUs from which load might be migrated. Co-authored-by: Austin Christ <austinwc@codeaurora.org> Co-authored-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Tested-by: Tyler Baicar <tbaicar@codeaurora.org> Signed-off-by: Jeffrey Hugo <jhugo@codeaurora.org> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Austin Christ <austinwc@codeaurora.org> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Timur Tabi <timur@codeaurora.org> Link: http://lkml.kernel.org/r/1496863138-11322-2-git-send-email-jhugo@codeaurora.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-06-07 19:18:57 +00:00
if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED))
sched: Improve balance_cpu() to consider other cpus in its group as target of (pinned) task Current load balance scheme requires only one cpu in a sched_group (balance_cpu) to look at other peer sched_groups for imbalance and pull tasks towards itself from a busy cpu. Tasks thus pulled by balance_cpu could later get picked up by cpus that are in the same sched_group as that of balance_cpu. This scheme however fails to pull tasks that are not allowed to run on balance_cpu (but are allowed to run on other cpus in its sched_group). That can affect fairness and in some worst case scenarios cause starvation. Consider a two core (2 threads/core) system running tasks as below: Core0 Core1 / \ / \ C0 C1 C2 C3 | | | | v v v v F0 T1 F1 [idle] T2 F0 = SCHED_FIFO task (pinned to C0) F1 = SCHED_FIFO task (pinned to C2) T1 = SCHED_OTHER task (pinned to C1) T2 = SCHED_OTHER task (pinned to C1 and C2) F1 could become a cpu hog, which will starve T2 unless C1 pulls it. Between C0 and C1 however, C0 is required to look for imbalance between cores, which will fail to pull T2 towards Core0. T2 will starve eternally in this case. The same scenario can arise in presence of non-rt tasks as well (say we replace F1 with high irq load). We tackle this problem by having balance_cpu move pinned tasks to one of its sibling cpus (where they can run). We first check if load balance goal can be met by ignoring pinned tasks, failing which we retry move_tasks() with a new env->dst_cpu. This patch modifies load balance semantics on who can move load towards a given cpu in a given sched_domain. Before this patch, a given_cpu or a ilb_cpu acting on behalf of an idle given_cpu is responsible for moving load to given_cpu. With this patch applied, balance_cpu can in addition decide on moving some load to a given_cpu. There is a remote possibility that excess load could get moved as a result of this (balance_cpu and given_cpu/ilb_cpu deciding *independently* and at *same* time to move some load to a given_cpu). However we should see less of such conflicting decisions in practice and moreover subsequent load balance cycles should correct the excess load moved to given_cpu. Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Signed-off-by: Prashanth Nageshappa <prashanth@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/4FE06CDB.2060605@linux.vnet.ibm.com [ minor edits ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2012-06-19 12:13:15 +00:00
return 0;
/* Prevent to re-select dst_cpu via env's CPUs: */
for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
if (cpumask_test_cpu(cpu, p->cpus_ptr)) {
env->flags |= LBF_DST_PINNED;
env->new_dst_cpu = cpu;
break;
}
sched: Improve balance_cpu() to consider other cpus in its group as target of (pinned) task Current load balance scheme requires only one cpu in a sched_group (balance_cpu) to look at other peer sched_groups for imbalance and pull tasks towards itself from a busy cpu. Tasks thus pulled by balance_cpu could later get picked up by cpus that are in the same sched_group as that of balance_cpu. This scheme however fails to pull tasks that are not allowed to run on balance_cpu (but are allowed to run on other cpus in its sched_group). That can affect fairness and in some worst case scenarios cause starvation. Consider a two core (2 threads/core) system running tasks as below: Core0 Core1 / \ / \ C0 C1 C2 C3 | | | | v v v v F0 T1 F1 [idle] T2 F0 = SCHED_FIFO task (pinned to C0) F1 = SCHED_FIFO task (pinned to C2) T1 = SCHED_OTHER task (pinned to C1) T2 = SCHED_OTHER task (pinned to C1 and C2) F1 could become a cpu hog, which will starve T2 unless C1 pulls it. Between C0 and C1 however, C0 is required to look for imbalance between cores, which will fail to pull T2 towards Core0. T2 will starve eternally in this case. The same scenario can arise in presence of non-rt tasks as well (say we replace F1 with high irq load). We tackle this problem by having balance_cpu move pinned tasks to one of its sibling cpus (where they can run). We first check if load balance goal can be met by ignoring pinned tasks, failing which we retry move_tasks() with a new env->dst_cpu. This patch modifies load balance semantics on who can move load towards a given cpu in a given sched_domain. Before this patch, a given_cpu or a ilb_cpu acting on behalf of an idle given_cpu is responsible for moving load to given_cpu. With this patch applied, balance_cpu can in addition decide on moving some load to a given_cpu. There is a remote possibility that excess load could get moved as a result of this (balance_cpu and given_cpu/ilb_cpu deciding *independently* and at *same* time to move some load to a given_cpu). However we should see less of such conflicting decisions in practice and moreover subsequent load balance cycles should correct the excess load moved to given_cpu. Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Signed-off-by: Prashanth Nageshappa <prashanth@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/4FE06CDB.2060605@linux.vnet.ibm.com [ minor edits ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2012-06-19 12:13:15 +00:00
}
return 0;
}
sched: Improve balance_cpu() to consider other cpus in its group as target of (pinned) task Current load balance scheme requires only one cpu in a sched_group (balance_cpu) to look at other peer sched_groups for imbalance and pull tasks towards itself from a busy cpu. Tasks thus pulled by balance_cpu could later get picked up by cpus that are in the same sched_group as that of balance_cpu. This scheme however fails to pull tasks that are not allowed to run on balance_cpu (but are allowed to run on other cpus in its sched_group). That can affect fairness and in some worst case scenarios cause starvation. Consider a two core (2 threads/core) system running tasks as below: Core0 Core1 / \ / \ C0 C1 C2 C3 | | | | v v v v F0 T1 F1 [idle] T2 F0 = SCHED_FIFO task (pinned to C0) F1 = SCHED_FIFO task (pinned to C2) T1 = SCHED_OTHER task (pinned to C1) T2 = SCHED_OTHER task (pinned to C1 and C2) F1 could become a cpu hog, which will starve T2 unless C1 pulls it. Between C0 and C1 however, C0 is required to look for imbalance between cores, which will fail to pull T2 towards Core0. T2 will starve eternally in this case. The same scenario can arise in presence of non-rt tasks as well (say we replace F1 with high irq load). We tackle this problem by having balance_cpu move pinned tasks to one of its sibling cpus (where they can run). We first check if load balance goal can be met by ignoring pinned tasks, failing which we retry move_tasks() with a new env->dst_cpu. This patch modifies load balance semantics on who can move load towards a given cpu in a given sched_domain. Before this patch, a given_cpu or a ilb_cpu acting on behalf of an idle given_cpu is responsible for moving load to given_cpu. With this patch applied, balance_cpu can in addition decide on moving some load to a given_cpu. There is a remote possibility that excess load could get moved as a result of this (balance_cpu and given_cpu/ilb_cpu deciding *independently* and at *same* time to move some load to a given_cpu). However we should see less of such conflicting decisions in practice and moreover subsequent load balance cycles should correct the excess load moved to given_cpu. Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Signed-off-by: Prashanth Nageshappa <prashanth@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/4FE06CDB.2060605@linux.vnet.ibm.com [ minor edits ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2012-06-19 12:13:15 +00:00
/* Record that we found atleast one task that could run on dst_cpu */
env->flags &= ~LBF_ALL_PINNED;
if (task_running(env->src_rq, p)) {
schedstat_inc(p->se.statistics.nr_failed_migrations_running);
return 0;
}
/*
* Aggressive migration if:
* 1) destination numa is preferred
* 2) task is cache cold, or
* 3) too many balance attempts have failed.
*/
tsk_cache_hot = migrate_degrades_locality(p, env);
if (tsk_cache_hot == -1)
tsk_cache_hot = task_hot(p, env);
if (tsk_cache_hot <= 0 ||
env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
if (tsk_cache_hot == 1) {
schedstat_inc(env->sd->lb_hot_gained[env->idle]);
schedstat_inc(p->se.statistics.nr_forced_migrations);
}
return 1;
}
schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
return 0;
}
/*
* detach_task() -- detach the task for the migration specified in env
*/
static void detach_task(struct task_struct *p, struct lb_env *env)
{
lockdep_assert_held(&env->src_rq->lock);
deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
set_task_cpu(p, env->dst_cpu);
}
/*
* detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
* part of active balancing operations within "domain".
*
* Returns a task if successful and NULL otherwise.
*/
static struct task_struct *detach_one_task(struct lb_env *env)
{
sched/fair: Search a task from the tail of the queue As a first step this patch makes cfs_tasks list as MRU one. It means, that when a next task is picked to run on physical CPU it is moved to the front of the list. Therefore, the cfs_tasks list is more or less sorted (except woken tasks) starting from recently given CPU time tasks toward tasks with max wait time in a run-queue, i.e. MRU list. Second, as part of the load balance operation, this approach starts detach_tasks()/detach_one_task() from the tail of the queue instead of the head, giving some advantages: - tends to pick a task with highest wait time; - tasks located in the tail are less likely cache-hot, therefore the can_migrate_task() decision is higher. hackbench illustrates slightly better performance. For example doing 1000 samples and 40 groups on i5-3320M CPU, it shows below figures: default: 0.657 avg patched: 0.646 avg Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Kirill Tkhai <tkhai@yandex.ru> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Mike Galbraith <umgwanakikbuti@gmail.com> Cc: Nicolas Pitre <nicolas.pitre@linaro.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Paul Turner <pjt@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Link: http://lkml.kernel.org/r/20170913102430.8985-2-urezki@gmail.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-09-13 10:24:30 +00:00
struct task_struct *p;
lockdep_assert_held(&env->src_rq->lock);
sched/fair: Search a task from the tail of the queue As a first step this patch makes cfs_tasks list as MRU one. It means, that when a next task is picked to run on physical CPU it is moved to the front of the list. Therefore, the cfs_tasks list is more or less sorted (except woken tasks) starting from recently given CPU time tasks toward tasks with max wait time in a run-queue, i.e. MRU list. Second, as part of the load balance operation, this approach starts detach_tasks()/detach_one_task() from the tail of the queue instead of the head, giving some advantages: - tends to pick a task with highest wait time; - tasks located in the tail are less likely cache-hot, therefore the can_migrate_task() decision is higher. hackbench illustrates slightly better performance. For example doing 1000 samples and 40 groups on i5-3320M CPU, it shows below figures: default: 0.657 avg patched: 0.646 avg Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Kirill Tkhai <tkhai@yandex.ru> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Mike Galbraith <umgwanakikbuti@gmail.com> Cc: Nicolas Pitre <nicolas.pitre@linaro.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Paul Turner <pjt@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Link: http://lkml.kernel.org/r/20170913102430.8985-2-urezki@gmail.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-09-13 10:24:30 +00:00
list_for_each_entry_reverse(p,
&env->src_rq->cfs_tasks, se.group_node) {
if (!can_migrate_task(p, env))
continue;
detach_task(p, env);
/*
* Right now, this is only the second place where
* lb_gained[env->idle] is updated (other is detach_tasks)
* so we can safely collect stats here rather than
* inside detach_tasks().
*/
schedstat_inc(env->sd->lb_gained[env->idle]);
return p;
}
return NULL;
}
static const unsigned int sched_nr_migrate_break = 32;
/*
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
* detach_tasks() -- tries to detach up to imbalance load/util/tasks from
* busiest_rq, as part of a balancing operation within domain "sd".
*
* Returns number of detached tasks if successful and 0 otherwise.
*/
static int detach_tasks(struct lb_env *env)
{
struct list_head *tasks = &env->src_rq->cfs_tasks;
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
unsigned long util, load;
struct task_struct *p;
int detached = 0;
lockdep_assert_held(&env->src_rq->lock);
if (env->imbalance <= 0)
return 0;
while (!list_empty(tasks)) {
/*
* We don't want to steal all, otherwise we may be treated likewise,
* which could at worst lead to a livelock crash.
*/
if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
break;
sched/fair: Search a task from the tail of the queue As a first step this patch makes cfs_tasks list as MRU one. It means, that when a next task is picked to run on physical CPU it is moved to the front of the list. Therefore, the cfs_tasks list is more or less sorted (except woken tasks) starting from recently given CPU time tasks toward tasks with max wait time in a run-queue, i.e. MRU list. Second, as part of the load balance operation, this approach starts detach_tasks()/detach_one_task() from the tail of the queue instead of the head, giving some advantages: - tends to pick a task with highest wait time; - tasks located in the tail are less likely cache-hot, therefore the can_migrate_task() decision is higher. hackbench illustrates slightly better performance. For example doing 1000 samples and 40 groups on i5-3320M CPU, it shows below figures: default: 0.657 avg patched: 0.646 avg Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Kirill Tkhai <tkhai@yandex.ru> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Mike Galbraith <umgwanakikbuti@gmail.com> Cc: Nicolas Pitre <nicolas.pitre@linaro.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Paul Turner <pjt@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Link: http://lkml.kernel.org/r/20170913102430.8985-2-urezki@gmail.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-09-13 10:24:30 +00:00
p = list_last_entry(tasks, struct task_struct, se.group_node);
env->loop++;
/* We've more or less seen every task there is, call it quits */
if (env->loop > env->loop_max)
break;
/* take a breather every nr_migrate tasks */
if (env->loop > env->loop_break) {
env->loop_break += sched_nr_migrate_break;
env->flags |= LBF_NEED_BREAK;
break;
}
if (!can_migrate_task(p, env))
goto next;
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
switch (env->migration_type) {
case migrate_load:
load = task_h_load(p);
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
if (sched_feat(LB_MIN) &&
load < 16 && !env->sd->nr_balance_failed)
goto next;
/*
* Make sure that we don't migrate too much load.
* Nevertheless, let relax the constraint if
* scheduler fails to find a good waiting task to
* migrate.
*/
if (load/2 > env->imbalance &&
env->sd->nr_balance_failed <= env->sd->cache_nice_tries)
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
goto next;
env->imbalance -= load;
break;
case migrate_util:
util = task_util_est(p);
if (util > env->imbalance)
goto next;
env->imbalance -= util;
break;
case migrate_task:
env->imbalance--;
break;
case migrate_misfit:
/* This is not a misfit task */
if (task_fits_capacity(p, capacity_of(env->src_cpu)))
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
goto next;
env->imbalance = 0;
break;
}
detach_task(p, env);
list_add(&p->se.group_node, &env->tasks);
detached++;
#ifdef CONFIG_PREEMPTION
/*
* NEWIDLE balancing is a source of latency, so preemptible
* kernels will stop after the first task is detached to minimize
* the critical section.
*/
if (env->idle == CPU_NEWLY_IDLE)
break;
#endif
/*
* We only want to steal up to the prescribed amount of
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
* load/util/tasks.
*/
if (env->imbalance <= 0)
break;
continue;
next:
sched/fair: Search a task from the tail of the queue As a first step this patch makes cfs_tasks list as MRU one. It means, that when a next task is picked to run on physical CPU it is moved to the front of the list. Therefore, the cfs_tasks list is more or less sorted (except woken tasks) starting from recently given CPU time tasks toward tasks with max wait time in a run-queue, i.e. MRU list. Second, as part of the load balance operation, this approach starts detach_tasks()/detach_one_task() from the tail of the queue instead of the head, giving some advantages: - tends to pick a task with highest wait time; - tasks located in the tail are less likely cache-hot, therefore the can_migrate_task() decision is higher. hackbench illustrates slightly better performance. For example doing 1000 samples and 40 groups on i5-3320M CPU, it shows below figures: default: 0.657 avg patched: 0.646 avg Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Kirill Tkhai <tkhai@yandex.ru> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Mike Galbraith <umgwanakikbuti@gmail.com> Cc: Nicolas Pitre <nicolas.pitre@linaro.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Paul Turner <pjt@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Link: http://lkml.kernel.org/r/20170913102430.8985-2-urezki@gmail.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-09-13 10:24:30 +00:00
list_move(&p->se.group_node, tasks);
}
/*
* Right now, this is one of only two places we collect this stat
* so we can safely collect detach_one_task() stats here rather
* than inside detach_one_task().
*/
schedstat_add(env->sd->lb_gained[env->idle], detached);
return detached;
}
/*
* attach_task() -- attach the task detached by detach_task() to its new rq.
*/
static void attach_task(struct rq *rq, struct task_struct *p)
{
lockdep_assert_held(&rq->lock);
BUG_ON(task_rq(p) != rq);
activate_task(rq, p, ENQUEUE_NOCLOCK);
check_preempt_curr(rq, p, 0);
}
/*
* attach_one_task() -- attaches the task returned from detach_one_task() to
* its new rq.
*/
static void attach_one_task(struct rq *rq, struct task_struct *p)
{
struct rq_flags rf;
rq_lock(rq, &rf);
update_rq_clock(rq);
attach_task(rq, p);
rq_unlock(rq, &rf);
}
/*
* attach_tasks() -- attaches all tasks detached by detach_tasks() to their
* new rq.
*/
static void attach_tasks(struct lb_env *env)
{
struct list_head *tasks = &env->tasks;
struct task_struct *p;
struct rq_flags rf;
rq_lock(env->dst_rq, &rf);
update_rq_clock(env->dst_rq);
while (!list_empty(tasks)) {
p = list_first_entry(tasks, struct task_struct, se.group_node);
list_del_init(&p->se.group_node);
attach_task(env->dst_rq, p);
}
rq_unlock(env->dst_rq, &rf);
}
sched/fair: Clean up definition of NOHZ blocked load functions cfs_rq_has_blocked() and others_have_blocked() are only used within update_blocked_averages(). The !CONFIG_FAIR_GROUP_SCHED version of the latter calls them within a #define CONFIG_NO_HZ_COMMON block, whereas the CONFIG_FAIR_GROUP_SCHED one calls them unconditionnally. As reported by Qian, the above leads to this warning in !CONFIG_NO_HZ_COMMON configs: kernel/sched/fair.c: In function 'update_blocked_averages': kernel/sched/fair.c:7750:7: warning: variable 'done' set but not used [-Wunused-but-set-variable] It wouldn't be wrong to keep cfs_rq_has_blocked() and others_have_blocked() as they are, but since their only current use is to figure out when we can stop calling update_blocked_averages() on fully decayed NOHZ idle CPUs, we can give them a new definition for !CONFIG_NO_HZ_COMMON. Change the definition of cfs_rq_has_blocked() and others_have_blocked() for !CONFIG_NO_HZ_COMMON so that the NOHZ-specific blocks of update_blocked_averages() become no-ops and the 'done' variable gets optimised out. While at it, remove the CONFIG_NO_HZ_COMMON block from the !CONFIG_FAIR_GROUP_SCHED definition of update_blocked_averages() by using the newly-introduced update_blocked_load_status() helper. No change in functionality intended. [ Additions by Peter Zijlstra. ] Reported-by: Qian Cai <cai@lca.pw> Signed-off-by: Valentin Schneider <valentin.schneider@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: https://lkml.kernel.org/r/20190603115424.7951-1-valentin.schneider@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-06-03 11:54:24 +00:00
#ifdef CONFIG_NO_HZ_COMMON
static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
{
if (cfs_rq->avg.load_avg)
return true;
if (cfs_rq->avg.util_avg)
return true;
return false;
}
sched/irq: Add IRQ utilization tracking interrupt and steal time are the only remaining activities tracked by rt_avg. Like for sched classes, we can use PELT to track their average utilization of the CPU. But unlike sched class, we don't track when entering/leaving interrupt; Instead, we take into account the time spent under interrupt context when we update rqs' clock (rq_clock_task). This also means that we have to decay the normal context time and account for interrupt time during the update. That's also important to note that because: rq_clock == rq_clock_task + interrupt time and rq_clock_task is used by a sched class to compute its utilization, the util_avg of a sched class only reflects the utilization of the time spent in normal context and not of the whole time of the CPU. The utilization of interrupt gives an more accurate level of utilization of CPU. The CPU utilization is: avg_irq + (1 - avg_irq / max capacity) * /Sum avg_rq Most of the time, avg_irq is small and neglictible so the use of the approximation CPU utilization = /Sum avg_rq was enough. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: claudio@evidence.eu.com Cc: daniel.lezcano@linaro.org Cc: dietmar.eggemann@arm.com Cc: joel@joelfernandes.org Cc: juri.lelli@redhat.com Cc: luca.abeni@santannapisa.it Cc: patrick.bellasi@arm.com Cc: quentin.perret@arm.com Cc: rjw@rjwysocki.net Cc: valentin.schneider@arm.com Cc: viresh.kumar@linaro.org Link: http://lkml.kernel.org/r/1530200714-4504-7-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-28 15:45:09 +00:00
static inline bool others_have_blocked(struct rq *rq)
2018-06-28 15:45:05 +00:00
{
if (READ_ONCE(rq->avg_rt.util_avg))
return true;
if (READ_ONCE(rq->avg_dl.util_avg))
return true;
if (thermal_load_avg(rq))
return true;
#ifdef CONFIG_HAVE_SCHED_AVG_IRQ
sched/irq: Add IRQ utilization tracking interrupt and steal time are the only remaining activities tracked by rt_avg. Like for sched classes, we can use PELT to track their average utilization of the CPU. But unlike sched class, we don't track when entering/leaving interrupt; Instead, we take into account the time spent under interrupt context when we update rqs' clock (rq_clock_task). This also means that we have to decay the normal context time and account for interrupt time during the update. That's also important to note that because: rq_clock == rq_clock_task + interrupt time and rq_clock_task is used by a sched class to compute its utilization, the util_avg of a sched class only reflects the utilization of the time spent in normal context and not of the whole time of the CPU. The utilization of interrupt gives an more accurate level of utilization of CPU. The CPU utilization is: avg_irq + (1 - avg_irq / max capacity) * /Sum avg_rq Most of the time, avg_irq is small and neglictible so the use of the approximation CPU utilization = /Sum avg_rq was enough. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: claudio@evidence.eu.com Cc: daniel.lezcano@linaro.org Cc: dietmar.eggemann@arm.com Cc: joel@joelfernandes.org Cc: juri.lelli@redhat.com Cc: luca.abeni@santannapisa.it Cc: patrick.bellasi@arm.com Cc: quentin.perret@arm.com Cc: rjw@rjwysocki.net Cc: valentin.schneider@arm.com Cc: viresh.kumar@linaro.org Link: http://lkml.kernel.org/r/1530200714-4504-7-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-28 15:45:09 +00:00
if (READ_ONCE(rq->avg_irq.util_avg))
return true;
#endif
2018-06-28 15:45:05 +00:00
return false;
}
sched/fair: Clean up definition of NOHZ blocked load functions cfs_rq_has_blocked() and others_have_blocked() are only used within update_blocked_averages(). The !CONFIG_FAIR_GROUP_SCHED version of the latter calls them within a #define CONFIG_NO_HZ_COMMON block, whereas the CONFIG_FAIR_GROUP_SCHED one calls them unconditionnally. As reported by Qian, the above leads to this warning in !CONFIG_NO_HZ_COMMON configs: kernel/sched/fair.c: In function 'update_blocked_averages': kernel/sched/fair.c:7750:7: warning: variable 'done' set but not used [-Wunused-but-set-variable] It wouldn't be wrong to keep cfs_rq_has_blocked() and others_have_blocked() as they are, but since their only current use is to figure out when we can stop calling update_blocked_averages() on fully decayed NOHZ idle CPUs, we can give them a new definition for !CONFIG_NO_HZ_COMMON. Change the definition of cfs_rq_has_blocked() and others_have_blocked() for !CONFIG_NO_HZ_COMMON so that the NOHZ-specific blocks of update_blocked_averages() become no-ops and the 'done' variable gets optimised out. While at it, remove the CONFIG_NO_HZ_COMMON block from the !CONFIG_FAIR_GROUP_SCHED definition of update_blocked_averages() by using the newly-introduced update_blocked_load_status() helper. No change in functionality intended. [ Additions by Peter Zijlstra. ] Reported-by: Qian Cai <cai@lca.pw> Signed-off-by: Valentin Schneider <valentin.schneider@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: https://lkml.kernel.org/r/20190603115424.7951-1-valentin.schneider@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-06-03 11:54:24 +00:00
static inline void update_blocked_load_status(struct rq *rq, bool has_blocked)
{
rq->last_blocked_load_update_tick = jiffies;
if (!has_blocked)
rq->has_blocked_load = 0;
}
#else
static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; }
static inline bool others_have_blocked(struct rq *rq) { return false; }
static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {}
#endif
sched/cpufreq: Move the cfs_rq_util_change() call to cpufreq_update_util() update_cfs_rq_load_avg() calls cfs_rq_util_change() every time PELT decays, which might be inefficient when the cpufreq driver has rate limitation. When a task is attached on a CPU, we have this call path: update_load_avg() update_cfs_rq_load_avg() cfs_rq_util_change -- > trig frequency update attach_entity_load_avg() cfs_rq_util_change -- > trig frequency update The 1st frequency update will not take into account the utilization of the newly attached task and the 2nd one might be discarded because of rate limitation of the cpufreq driver. update_cfs_rq_load_avg() is only called by update_blocked_averages() and update_load_avg() so we can move the call to cfs_rq_util_change/cpufreq_update_util() into these two functions. It's also interesting to note that update_load_avg() already calls cfs_rq_util_change() directly for the !SMP case. This change will also ensure that cpufreq_update_util() is called even when there is no more CFS rq in the leaf_cfs_rq_list to update, but only IRQ, RT or DL PELT signals. [ mingo: Minor updates. ] Reported-by: Doug Smythies <dsmythies@telus.net> Tested-by: Doug Smythies <dsmythies@telus.net> Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Acked-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: juri.lelli@redhat.com Cc: linux-pm@vger.kernel.org Cc: mgorman@suse.de Cc: rostedt@goodmis.org Cc: sargun@sargun.me Cc: srinivas.pandruvada@linux.intel.com Cc: tj@kernel.org Cc: xiexiuqi@huawei.com Cc: xiezhipeng1@huawei.com Fixes: 039ae8bcf7a5 ("sched/fair: Fix O(nr_cgroups) in the load balancing path") Link: https://lkml.kernel.org/r/1574083279-799-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-11-18 13:21:19 +00:00
static bool __update_blocked_others(struct rq *rq, bool *done)
{
const struct sched_class *curr_class;
u64 now = rq_clock_pelt(rq);
unsigned long thermal_pressure;
sched/cpufreq: Move the cfs_rq_util_change() call to cpufreq_update_util() update_cfs_rq_load_avg() calls cfs_rq_util_change() every time PELT decays, which might be inefficient when the cpufreq driver has rate limitation. When a task is attached on a CPU, we have this call path: update_load_avg() update_cfs_rq_load_avg() cfs_rq_util_change -- > trig frequency update attach_entity_load_avg() cfs_rq_util_change -- > trig frequency update The 1st frequency update will not take into account the utilization of the newly attached task and the 2nd one might be discarded because of rate limitation of the cpufreq driver. update_cfs_rq_load_avg() is only called by update_blocked_averages() and update_load_avg() so we can move the call to cfs_rq_util_change/cpufreq_update_util() into these two functions. It's also interesting to note that update_load_avg() already calls cfs_rq_util_change() directly for the !SMP case. This change will also ensure that cpufreq_update_util() is called even when there is no more CFS rq in the leaf_cfs_rq_list to update, but only IRQ, RT or DL PELT signals. [ mingo: Minor updates. ] Reported-by: Doug Smythies <dsmythies@telus.net> Tested-by: Doug Smythies <dsmythies@telus.net> Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Acked-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: juri.lelli@redhat.com Cc: linux-pm@vger.kernel.org Cc: mgorman@suse.de Cc: rostedt@goodmis.org Cc: sargun@sargun.me Cc: srinivas.pandruvada@linux.intel.com Cc: tj@kernel.org Cc: xiexiuqi@huawei.com Cc: xiezhipeng1@huawei.com Fixes: 039ae8bcf7a5 ("sched/fair: Fix O(nr_cgroups) in the load balancing path") Link: https://lkml.kernel.org/r/1574083279-799-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-11-18 13:21:19 +00:00
bool decayed;
/*
* update_load_avg() can call cpufreq_update_util(). Make sure that RT,
* DL and IRQ signals have been updated before updating CFS.
*/
curr_class = rq->curr->sched_class;
thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
sched/cpufreq: Move the cfs_rq_util_change() call to cpufreq_update_util() update_cfs_rq_load_avg() calls cfs_rq_util_change() every time PELT decays, which might be inefficient when the cpufreq driver has rate limitation. When a task is attached on a CPU, we have this call path: update_load_avg() update_cfs_rq_load_avg() cfs_rq_util_change -- > trig frequency update attach_entity_load_avg() cfs_rq_util_change -- > trig frequency update The 1st frequency update will not take into account the utilization of the newly attached task and the 2nd one might be discarded because of rate limitation of the cpufreq driver. update_cfs_rq_load_avg() is only called by update_blocked_averages() and update_load_avg() so we can move the call to cfs_rq_util_change/cpufreq_update_util() into these two functions. It's also interesting to note that update_load_avg() already calls cfs_rq_util_change() directly for the !SMP case. This change will also ensure that cpufreq_update_util() is called even when there is no more CFS rq in the leaf_cfs_rq_list to update, but only IRQ, RT or DL PELT signals. [ mingo: Minor updates. ] Reported-by: Doug Smythies <dsmythies@telus.net> Tested-by: Doug Smythies <dsmythies@telus.net> Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Acked-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: juri.lelli@redhat.com Cc: linux-pm@vger.kernel.org Cc: mgorman@suse.de Cc: rostedt@goodmis.org Cc: sargun@sargun.me Cc: srinivas.pandruvada@linux.intel.com Cc: tj@kernel.org Cc: xiexiuqi@huawei.com Cc: xiezhipeng1@huawei.com Fixes: 039ae8bcf7a5 ("sched/fair: Fix O(nr_cgroups) in the load balancing path") Link: https://lkml.kernel.org/r/1574083279-799-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-11-18 13:21:19 +00:00
decayed = update_rt_rq_load_avg(now, rq, curr_class == &rt_sched_class) |
update_dl_rq_load_avg(now, rq, curr_class == &dl_sched_class) |
update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure) |
sched/cpufreq: Move the cfs_rq_util_change() call to cpufreq_update_util() update_cfs_rq_load_avg() calls cfs_rq_util_change() every time PELT decays, which might be inefficient when the cpufreq driver has rate limitation. When a task is attached on a CPU, we have this call path: update_load_avg() update_cfs_rq_load_avg() cfs_rq_util_change -- > trig frequency update attach_entity_load_avg() cfs_rq_util_change -- > trig frequency update The 1st frequency update will not take into account the utilization of the newly attached task and the 2nd one might be discarded because of rate limitation of the cpufreq driver. update_cfs_rq_load_avg() is only called by update_blocked_averages() and update_load_avg() so we can move the call to cfs_rq_util_change/cpufreq_update_util() into these two functions. It's also interesting to note that update_load_avg() already calls cfs_rq_util_change() directly for the !SMP case. This change will also ensure that cpufreq_update_util() is called even when there is no more CFS rq in the leaf_cfs_rq_list to update, but only IRQ, RT or DL PELT signals. [ mingo: Minor updates. ] Reported-by: Doug Smythies <dsmythies@telus.net> Tested-by: Doug Smythies <dsmythies@telus.net> Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Acked-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: juri.lelli@redhat.com Cc: linux-pm@vger.kernel.org Cc: mgorman@suse.de Cc: rostedt@goodmis.org Cc: sargun@sargun.me Cc: srinivas.pandruvada@linux.intel.com Cc: tj@kernel.org Cc: xiexiuqi@huawei.com Cc: xiezhipeng1@huawei.com Fixes: 039ae8bcf7a5 ("sched/fair: Fix O(nr_cgroups) in the load balancing path") Link: https://lkml.kernel.org/r/1574083279-799-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-11-18 13:21:19 +00:00
update_irq_load_avg(rq, 0);
if (others_have_blocked(rq))
*done = false;
return decayed;
}
#ifdef CONFIG_FAIR_GROUP_SCHED
static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
{
if (cfs_rq->load.weight)
return false;
if (cfs_rq->avg.load_sum)
return false;
if (cfs_rq->avg.util_sum)
return false;
sched/pelt: Add a new runnable average signal Now that runnable_load_avg has been removed, we can replace it by a new signal that will highlight the runnable pressure on a cfs_rq. This signal track the waiting time of tasks on rq and can help to better define the state of rqs. At now, only util_avg is used to define the state of a rq: A rq with more that around 80% of utilization and more than 1 tasks is considered as overloaded. But the util_avg signal of a rq can become temporaly low after that a task migrated onto another rq which can bias the classification of the rq. When tasks compete for the same rq, their runnable average signal will be higher than util_avg as it will include the waiting time and we can use this signal to better classify cfs_rqs. The new runnable_avg will track the runnable time of a task which simply adds the waiting time to the running time. The runnable _avg of cfs_rq will be the /Sum of se's runnable_avg and the runnable_avg of group entity will follow the one of the rq similarly to util_avg. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: "Dietmar Eggemann <dietmar.eggemann@arm.com>" Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Valentin Schneider <valentin.schneider@arm.com> Cc: Phil Auld <pauld@redhat.com> Cc: Hillf Danton <hdanton@sina.com> Link: https://lore.kernel.org/r/20200224095223.13361-9-mgorman@techsingularity.net
2020-02-24 09:52:18 +00:00
if (cfs_rq->avg.runnable_sum)
return false;
return true;
}
sched/cpufreq: Move the cfs_rq_util_change() call to cpufreq_update_util() update_cfs_rq_load_avg() calls cfs_rq_util_change() every time PELT decays, which might be inefficient when the cpufreq driver has rate limitation. When a task is attached on a CPU, we have this call path: update_load_avg() update_cfs_rq_load_avg() cfs_rq_util_change -- > trig frequency update attach_entity_load_avg() cfs_rq_util_change -- > trig frequency update The 1st frequency update will not take into account the utilization of the newly attached task and the 2nd one might be discarded because of rate limitation of the cpufreq driver. update_cfs_rq_load_avg() is only called by update_blocked_averages() and update_load_avg() so we can move the call to cfs_rq_util_change/cpufreq_update_util() into these two functions. It's also interesting to note that update_load_avg() already calls cfs_rq_util_change() directly for the !SMP case. This change will also ensure that cpufreq_update_util() is called even when there is no more CFS rq in the leaf_cfs_rq_list to update, but only IRQ, RT or DL PELT signals. [ mingo: Minor updates. ] Reported-by: Doug Smythies <dsmythies@telus.net> Tested-by: Doug Smythies <dsmythies@telus.net> Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Acked-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: juri.lelli@redhat.com Cc: linux-pm@vger.kernel.org Cc: mgorman@suse.de Cc: rostedt@goodmis.org Cc: sargun@sargun.me Cc: srinivas.pandruvada@linux.intel.com Cc: tj@kernel.org Cc: xiexiuqi@huawei.com Cc: xiezhipeng1@huawei.com Fixes: 039ae8bcf7a5 ("sched/fair: Fix O(nr_cgroups) in the load balancing path") Link: https://lkml.kernel.org/r/1574083279-799-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-11-18 13:21:19 +00:00
static bool __update_blocked_fair(struct rq *rq, bool *done)
{
struct cfs_rq *cfs_rq, *pos;
sched/cpufreq: Move the cfs_rq_util_change() call to cpufreq_update_util() update_cfs_rq_load_avg() calls cfs_rq_util_change() every time PELT decays, which might be inefficient when the cpufreq driver has rate limitation. When a task is attached on a CPU, we have this call path: update_load_avg() update_cfs_rq_load_avg() cfs_rq_util_change -- > trig frequency update attach_entity_load_avg() cfs_rq_util_change -- > trig frequency update The 1st frequency update will not take into account the utilization of the newly attached task and the 2nd one might be discarded because of rate limitation of the cpufreq driver. update_cfs_rq_load_avg() is only called by update_blocked_averages() and update_load_avg() so we can move the call to cfs_rq_util_change/cpufreq_update_util() into these two functions. It's also interesting to note that update_load_avg() already calls cfs_rq_util_change() directly for the !SMP case. This change will also ensure that cpufreq_update_util() is called even when there is no more CFS rq in the leaf_cfs_rq_list to update, but only IRQ, RT or DL PELT signals. [ mingo: Minor updates. ] Reported-by: Doug Smythies <dsmythies@telus.net> Tested-by: Doug Smythies <dsmythies@telus.net> Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Acked-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: juri.lelli@redhat.com Cc: linux-pm@vger.kernel.org Cc: mgorman@suse.de Cc: rostedt@goodmis.org Cc: sargun@sargun.me Cc: srinivas.pandruvada@linux.intel.com Cc: tj@kernel.org Cc: xiexiuqi@huawei.com Cc: xiezhipeng1@huawei.com Fixes: 039ae8bcf7a5 ("sched/fair: Fix O(nr_cgroups) in the load balancing path") Link: https://lkml.kernel.org/r/1574083279-799-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-11-18 13:21:19 +00:00
bool decayed = false;
int cpu = cpu_of(rq);
/*
* Iterates the task_group tree in a bottom up fashion, see
* list_add_leaf_cfs_rq() for details.
*/
for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
struct sched_entity *se;
sched/cpufreq: Move the cfs_rq_util_change() call to cpufreq_update_util() update_cfs_rq_load_avg() calls cfs_rq_util_change() every time PELT decays, which might be inefficient when the cpufreq driver has rate limitation. When a task is attached on a CPU, we have this call path: update_load_avg() update_cfs_rq_load_avg() cfs_rq_util_change -- > trig frequency update attach_entity_load_avg() cfs_rq_util_change -- > trig frequency update The 1st frequency update will not take into account the utilization of the newly attached task and the 2nd one might be discarded because of rate limitation of the cpufreq driver. update_cfs_rq_load_avg() is only called by update_blocked_averages() and update_load_avg() so we can move the call to cfs_rq_util_change/cpufreq_update_util() into these two functions. It's also interesting to note that update_load_avg() already calls cfs_rq_util_change() directly for the !SMP case. This change will also ensure that cpufreq_update_util() is called even when there is no more CFS rq in the leaf_cfs_rq_list to update, but only IRQ, RT or DL PELT signals. [ mingo: Minor updates. ] Reported-by: Doug Smythies <dsmythies@telus.net> Tested-by: Doug Smythies <dsmythies@telus.net> Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Acked-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: juri.lelli@redhat.com Cc: linux-pm@vger.kernel.org Cc: mgorman@suse.de Cc: rostedt@goodmis.org Cc: sargun@sargun.me Cc: srinivas.pandruvada@linux.intel.com Cc: tj@kernel.org Cc: xiexiuqi@huawei.com Cc: xiezhipeng1@huawei.com Fixes: 039ae8bcf7a5 ("sched/fair: Fix O(nr_cgroups) in the load balancing path") Link: https://lkml.kernel.org/r/1574083279-799-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-11-18 13:21:19 +00:00
if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq)) {
sched/fair: Rewrite runnable load and utilization average tracking The idea of runnable load average (let runnable time contribute to weight) was proposed by Paul Turner and Ben Segall, and it is still followed by this rewrite. This rewrite aims to solve the following issues: 1. cfs_rq's load average (namely runnable_load_avg and blocked_load_avg) is updated at the granularity of an entity at a time, which results in the cfs_rq's load average is stale or partially updated: at any time, only one entity is up to date, all other entities are effectively lagging behind. This is undesirable. To illustrate, if we have n runnable entities in the cfs_rq, as time elapses, they certainly become outdated: t0: cfs_rq { e1_old, e2_old, ..., en_old } and when we update: t1: update e1, then we have cfs_rq { e1_new, e2_old, ..., en_old } t2: update e2, then we have cfs_rq { e1_old, e2_new, ..., en_old } ... We solve this by combining all runnable entities' load averages together in cfs_rq's avg, and update the cfs_rq's avg as a whole. This is based on the fact that if we regard the update as a function, then: w * update(e) = update(w * e) and update(e1) + update(e2) = update(e1 + e2), then w1 * update(e1) + w2 * update(e2) = update(w1 * e1 + w2 * e2) therefore, by this rewrite, we have an entirely updated cfs_rq at the time we update it: t1: update cfs_rq { e1_new, e2_new, ..., en_new } t2: update cfs_rq { e1_new, e2_new, ..., en_new } ... 2. cfs_rq's load average is different between top rq->cfs_rq and other task_group's per CPU cfs_rqs in whether or not blocked_load_average contributes to the load. The basic idea behind runnable load average (the same for utilization) is that the blocked state is taken into account as opposed to only accounting for the currently runnable state. Therefore, the average should include both the runnable/running and blocked load averages. This rewrite does that. In addition, we also combine runnable/running and blocked averages of all entities into the cfs_rq's average, and update it together at once. This is based on the fact that: update(runnable) + update(blocked) = update(runnable + blocked) This significantly reduces the code as we don't need to separately maintain/update runnable/running load and blocked load. 3. How task_group entities' share is calculated is complex and imprecise. We reduce the complexity in this rewrite to allow a very simple rule: the task_group's load_avg is aggregated from its per CPU cfs_rqs's load_avgs. Then group entity's weight is simply proportional to its own cfs_rq's load_avg / task_group's load_avg. To illustrate, if a task_group has { cfs_rq1, cfs_rq2, ..., cfs_rqn }, then, task_group_avg = cfs_rq1_avg + cfs_rq2_avg + ... + cfs_rqn_avg, then cfs_rqx's entity's share = cfs_rqx_avg / task_group_avg * task_group's share To sum up, this rewrite in principle is equivalent to the current one, but fixes the issues described above. Turns out, it significantly reduces the code complexity and hence increases clarity and efficiency. In addition, the new averages are more smooth/continuous (no spurious spikes and valleys) and updated more consistently and quickly to reflect the load dynamics. As a result, we have less load tracking overhead, better performance, and especially better power efficiency due to more balanced load. Signed-off-by: Yuyang Du <yuyang.du@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: arjan@linux.intel.com Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: fengguang.wu@intel.com Cc: len.brown@intel.com Cc: morten.rasmussen@arm.com Cc: pjt@google.com Cc: rafael.j.wysocki@intel.com Cc: umgwanakikbuti@gmail.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1436918682-4971-3-git-send-email-yuyang.du@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-07-15 00:04:37 +00:00
update_tg_load_avg(cfs_rq, 0);
sched/cpufreq: Move the cfs_rq_util_change() call to cpufreq_update_util() update_cfs_rq_load_avg() calls cfs_rq_util_change() every time PELT decays, which might be inefficient when the cpufreq driver has rate limitation. When a task is attached on a CPU, we have this call path: update_load_avg() update_cfs_rq_load_avg() cfs_rq_util_change -- > trig frequency update attach_entity_load_avg() cfs_rq_util_change -- > trig frequency update The 1st frequency update will not take into account the utilization of the newly attached task and the 2nd one might be discarded because of rate limitation of the cpufreq driver. update_cfs_rq_load_avg() is only called by update_blocked_averages() and update_load_avg() so we can move the call to cfs_rq_util_change/cpufreq_update_util() into these two functions. It's also interesting to note that update_load_avg() already calls cfs_rq_util_change() directly for the !SMP case. This change will also ensure that cpufreq_update_util() is called even when there is no more CFS rq in the leaf_cfs_rq_list to update, but only IRQ, RT or DL PELT signals. [ mingo: Minor updates. ] Reported-by: Doug Smythies <dsmythies@telus.net> Tested-by: Doug Smythies <dsmythies@telus.net> Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Acked-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: juri.lelli@redhat.com Cc: linux-pm@vger.kernel.org Cc: mgorman@suse.de Cc: rostedt@goodmis.org Cc: sargun@sargun.me Cc: srinivas.pandruvada@linux.intel.com Cc: tj@kernel.org Cc: xiexiuqi@huawei.com Cc: xiezhipeng1@huawei.com Fixes: 039ae8bcf7a5 ("sched/fair: Fix O(nr_cgroups) in the load balancing path") Link: https://lkml.kernel.org/r/1574083279-799-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-11-18 13:21:19 +00:00
if (cfs_rq == &rq->cfs)
decayed = true;
}
/* Propagate pending load changes to the parent, if any: */
se = cfs_rq->tg->se[cpu];
if (se && !skip_blocked_update(se))
update_load_avg(cfs_rq_of(se), se, 0);
/*
* There can be a lot of idle CPU cgroups. Don't let fully
* decayed cfs_rqs linger on the list.
*/
if (cfs_rq_is_decayed(cfs_rq))
list_del_leaf_cfs_rq(cfs_rq);
/* Don't need periodic decay once load/util_avg are null */
if (cfs_rq_has_blocked(cfs_rq))
sched/cpufreq: Move the cfs_rq_util_change() call to cpufreq_update_util() update_cfs_rq_load_avg() calls cfs_rq_util_change() every time PELT decays, which might be inefficient when the cpufreq driver has rate limitation. When a task is attached on a CPU, we have this call path: update_load_avg() update_cfs_rq_load_avg() cfs_rq_util_change -- > trig frequency update attach_entity_load_avg() cfs_rq_util_change -- > trig frequency update The 1st frequency update will not take into account the utilization of the newly attached task and the 2nd one might be discarded because of rate limitation of the cpufreq driver. update_cfs_rq_load_avg() is only called by update_blocked_averages() and update_load_avg() so we can move the call to cfs_rq_util_change/cpufreq_update_util() into these two functions. It's also interesting to note that update_load_avg() already calls cfs_rq_util_change() directly for the !SMP case. This change will also ensure that cpufreq_update_util() is called even when there is no more CFS rq in the leaf_cfs_rq_list to update, but only IRQ, RT or DL PELT signals. [ mingo: Minor updates. ] Reported-by: Doug Smythies <dsmythies@telus.net> Tested-by: Doug Smythies <dsmythies@telus.net> Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Acked-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: juri.lelli@redhat.com Cc: linux-pm@vger.kernel.org Cc: mgorman@suse.de Cc: rostedt@goodmis.org Cc: sargun@sargun.me Cc: srinivas.pandruvada@linux.intel.com Cc: tj@kernel.org Cc: xiexiuqi@huawei.com Cc: xiezhipeng1@huawei.com Fixes: 039ae8bcf7a5 ("sched/fair: Fix O(nr_cgroups) in the load balancing path") Link: https://lkml.kernel.org/r/1574083279-799-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-11-18 13:21:19 +00:00
*done = false;
sched/fair: Rewrite runnable load and utilization average tracking The idea of runnable load average (let runnable time contribute to weight) was proposed by Paul Turner and Ben Segall, and it is still followed by this rewrite. This rewrite aims to solve the following issues: 1. cfs_rq's load average (namely runnable_load_avg and blocked_load_avg) is updated at the granularity of an entity at a time, which results in the cfs_rq's load average is stale or partially updated: at any time, only one entity is up to date, all other entities are effectively lagging behind. This is undesirable. To illustrate, if we have n runnable entities in the cfs_rq, as time elapses, they certainly become outdated: t0: cfs_rq { e1_old, e2_old, ..., en_old } and when we update: t1: update e1, then we have cfs_rq { e1_new, e2_old, ..., en_old } t2: update e2, then we have cfs_rq { e1_old, e2_new, ..., en_old } ... We solve this by combining all runnable entities' load averages together in cfs_rq's avg, and update the cfs_rq's avg as a whole. This is based on the fact that if we regard the update as a function, then: w * update(e) = update(w * e) and update(e1) + update(e2) = update(e1 + e2), then w1 * update(e1) + w2 * update(e2) = update(w1 * e1 + w2 * e2) therefore, by this rewrite, we have an entirely updated cfs_rq at the time we update it: t1: update cfs_rq { e1_new, e2_new, ..., en_new } t2: update cfs_rq { e1_new, e2_new, ..., en_new } ... 2. cfs_rq's load average is different between top rq->cfs_rq and other task_group's per CPU cfs_rqs in whether or not blocked_load_average contributes to the load. The basic idea behind runnable load average (the same for utilization) is that the blocked state is taken into account as opposed to only accounting for the currently runnable state. Therefore, the average should include both the runnable/running and blocked load averages. This rewrite does that. In addition, we also combine runnable/running and blocked averages of all entities into the cfs_rq's average, and update it together at once. This is based on the fact that: update(runnable) + update(blocked) = update(runnable + blocked) This significantly reduces the code as we don't need to separately maintain/update runnable/running load and blocked load. 3. How task_group entities' share is calculated is complex and imprecise. We reduce the complexity in this rewrite to allow a very simple rule: the task_group's load_avg is aggregated from its per CPU cfs_rqs's load_avgs. Then group entity's weight is simply proportional to its own cfs_rq's load_avg / task_group's load_avg. To illustrate, if a task_group has { cfs_rq1, cfs_rq2, ..., cfs_rqn }, then, task_group_avg = cfs_rq1_avg + cfs_rq2_avg + ... + cfs_rqn_avg, then cfs_rqx's entity's share = cfs_rqx_avg / task_group_avg * task_group's share To sum up, this rewrite in principle is equivalent to the current one, but fixes the issues described above. Turns out, it significantly reduces the code complexity and hence increases clarity and efficiency. In addition, the new averages are more smooth/continuous (no spurious spikes and valleys) and updated more consistently and quickly to reflect the load dynamics. As a result, we have less load tracking overhead, better performance, and especially better power efficiency due to more balanced load. Signed-off-by: Yuyang Du <yuyang.du@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: arjan@linux.intel.com Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: fengguang.wu@intel.com Cc: len.brown@intel.com Cc: morten.rasmussen@arm.com Cc: pjt@google.com Cc: rafael.j.wysocki@intel.com Cc: umgwanakikbuti@gmail.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1436918682-4971-3-git-send-email-yuyang.du@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-07-15 00:04:37 +00:00
}
sched/cpufreq: Move the cfs_rq_util_change() call to cpufreq_update_util() update_cfs_rq_load_avg() calls cfs_rq_util_change() every time PELT decays, which might be inefficient when the cpufreq driver has rate limitation. When a task is attached on a CPU, we have this call path: update_load_avg() update_cfs_rq_load_avg() cfs_rq_util_change -- > trig frequency update attach_entity_load_avg() cfs_rq_util_change -- > trig frequency update The 1st frequency update will not take into account the utilization of the newly attached task and the 2nd one might be discarded because of rate limitation of the cpufreq driver. update_cfs_rq_load_avg() is only called by update_blocked_averages() and update_load_avg() so we can move the call to cfs_rq_util_change/cpufreq_update_util() into these two functions. It's also interesting to note that update_load_avg() already calls cfs_rq_util_change() directly for the !SMP case. This change will also ensure that cpufreq_update_util() is called even when there is no more CFS rq in the leaf_cfs_rq_list to update, but only IRQ, RT or DL PELT signals. [ mingo: Minor updates. ] Reported-by: Doug Smythies <dsmythies@telus.net> Tested-by: Doug Smythies <dsmythies@telus.net> Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Acked-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: juri.lelli@redhat.com Cc: linux-pm@vger.kernel.org Cc: mgorman@suse.de Cc: rostedt@goodmis.org Cc: sargun@sargun.me Cc: srinivas.pandruvada@linux.intel.com Cc: tj@kernel.org Cc: xiexiuqi@huawei.com Cc: xiezhipeng1@huawei.com Fixes: 039ae8bcf7a5 ("sched/fair: Fix O(nr_cgroups) in the load balancing path") Link: https://lkml.kernel.org/r/1574083279-799-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-11-18 13:21:19 +00:00
return decayed;
}
/*
sched: Move h_load calculation to task_h_load() The bad thing about update_h_load(), which computes hierarchical load factor for task groups, is that it is called for each task group in the system before every load balancer run, and since rebalance can be triggered very often, this function can eat really a lot of cpu time if there are many cpu cgroups in the system. Although the situation was improved significantly by commit a35b646 ('sched, cgroup: Reduce rq->lock hold times for large cgroup hierarchies'), the problem still can arise under some kinds of loads, e.g. when cpus are switching from idle to busy and back very frequently. For instance, when I start 1000 of processes that wake up every millisecond on my 8 cpus host, 'top' and 'perf top' show: Cpu(s): 17.8%us, 24.3%sy, 0.0%ni, 57.9%id, 0.0%wa, 0.0%hi, 0.0%si Events: 243K cycles 7.57% [kernel] [k] __schedule 7.08% [kernel] [k] timerqueue_add 6.13% libc-2.12.so [.] usleep Then if I create 10000 *idle* cpu cgroups (no processes in them), cpu usage increases significantly although the 'wakers' are still executing in the root cpu cgroup: Cpu(s): 19.1%us, 48.7%sy, 0.0%ni, 31.6%id, 0.0%wa, 0.0%hi, 0.7%si Events: 230K cycles 24.56% [kernel] [k] tg_load_down 5.76% [kernel] [k] __schedule This happens because this particular kind of load triggers 'new idle' rebalance very frequently, which requires calling update_h_load(), which, in turn, calls tg_load_down() for every *idle* cpu cgroup even though it is absolutely useless, because idle cpu cgroups have no tasks to pull. This patch tries to improve the situation by making h_load calculation proceed only when h_load is really necessary. To achieve this, it substitutes update_h_load() with update_cfs_rq_h_load(), which computes h_load only for a given cfs_rq and all its ascendants, and makes the load balancer call this function whenever it considers if a task should be pulled, i.e. it moves h_load calculations directly to task_h_load(). For h_load of the same cfs_rq not to be updated multiple times (in case several tasks in the same cgroup are considered during the same balance run), the patch keeps the time of the last h_load update for each cfs_rq and breaks calculation when it finds h_load to be uptodate. The benefit of it is that h_load is computed only for those cfs_rq's, which really need it, in particular all idle task groups are skipped. Although this, in fact, moves h_load calculation under rq lock, it should not affect latency much, because the amount of work done under rq lock while trying to pull tasks is limited by sched_nr_migrate. After the patch applied with the setup described above (1000 wakers in the root cgroup and 10000 idle cgroups), I get: Cpu(s): 16.9%us, 24.8%sy, 0.0%ni, 58.4%id, 0.0%wa, 0.0%hi, 0.0%si Events: 242K cycles 7.57% [kernel] [k] __schedule 6.70% [kernel] [k] timerqueue_add 5.93% libc-2.12.so [.] usleep Signed-off-by: Vladimir Davydov <vdavydov@parallels.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/1373896159-1278-1-git-send-email-vdavydov@parallels.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-07-15 13:49:19 +00:00
* Compute the hierarchical load factor for cfs_rq and all its ascendants.
* This needs to be done in a top-down fashion because the load of a child
* group is a fraction of its parents load.
*/
sched: Move h_load calculation to task_h_load() The bad thing about update_h_load(), which computes hierarchical load factor for task groups, is that it is called for each task group in the system before every load balancer run, and since rebalance can be triggered very often, this function can eat really a lot of cpu time if there are many cpu cgroups in the system. Although the situation was improved significantly by commit a35b646 ('sched, cgroup: Reduce rq->lock hold times for large cgroup hierarchies'), the problem still can arise under some kinds of loads, e.g. when cpus are switching from idle to busy and back very frequently. For instance, when I start 1000 of processes that wake up every millisecond on my 8 cpus host, 'top' and 'perf top' show: Cpu(s): 17.8%us, 24.3%sy, 0.0%ni, 57.9%id, 0.0%wa, 0.0%hi, 0.0%si Events: 243K cycles 7.57% [kernel] [k] __schedule 7.08% [kernel] [k] timerqueue_add 6.13% libc-2.12.so [.] usleep Then if I create 10000 *idle* cpu cgroups (no processes in them), cpu usage increases significantly although the 'wakers' are still executing in the root cpu cgroup: Cpu(s): 19.1%us, 48.7%sy, 0.0%ni, 31.6%id, 0.0%wa, 0.0%hi, 0.7%si Events: 230K cycles 24.56% [kernel] [k] tg_load_down 5.76% [kernel] [k] __schedule This happens because this particular kind of load triggers 'new idle' rebalance very frequently, which requires calling update_h_load(), which, in turn, calls tg_load_down() for every *idle* cpu cgroup even though it is absolutely useless, because idle cpu cgroups have no tasks to pull. This patch tries to improve the situation by making h_load calculation proceed only when h_load is really necessary. To achieve this, it substitutes update_h_load() with update_cfs_rq_h_load(), which computes h_load only for a given cfs_rq and all its ascendants, and makes the load balancer call this function whenever it considers if a task should be pulled, i.e. it moves h_load calculations directly to task_h_load(). For h_load of the same cfs_rq not to be updated multiple times (in case several tasks in the same cgroup are considered during the same balance run), the patch keeps the time of the last h_load update for each cfs_rq and breaks calculation when it finds h_load to be uptodate. The benefit of it is that h_load is computed only for those cfs_rq's, which really need it, in particular all idle task groups are skipped. Although this, in fact, moves h_load calculation under rq lock, it should not affect latency much, because the amount of work done under rq lock while trying to pull tasks is limited by sched_nr_migrate. After the patch applied with the setup described above (1000 wakers in the root cgroup and 10000 idle cgroups), I get: Cpu(s): 16.9%us, 24.8%sy, 0.0%ni, 58.4%id, 0.0%wa, 0.0%hi, 0.0%si Events: 242K cycles 7.57% [kernel] [k] __schedule 6.70% [kernel] [k] timerqueue_add 5.93% libc-2.12.so [.] usleep Signed-off-by: Vladimir Davydov <vdavydov@parallels.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/1373896159-1278-1-git-send-email-vdavydov@parallels.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-07-15 13:49:19 +00:00
static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
{
sched: Move h_load calculation to task_h_load() The bad thing about update_h_load(), which computes hierarchical load factor for task groups, is that it is called for each task group in the system before every load balancer run, and since rebalance can be triggered very often, this function can eat really a lot of cpu time if there are many cpu cgroups in the system. Although the situation was improved significantly by commit a35b646 ('sched, cgroup: Reduce rq->lock hold times for large cgroup hierarchies'), the problem still can arise under some kinds of loads, e.g. when cpus are switching from idle to busy and back very frequently. For instance, when I start 1000 of processes that wake up every millisecond on my 8 cpus host, 'top' and 'perf top' show: Cpu(s): 17.8%us, 24.3%sy, 0.0%ni, 57.9%id, 0.0%wa, 0.0%hi, 0.0%si Events: 243K cycles 7.57% [kernel] [k] __schedule 7.08% [kernel] [k] timerqueue_add 6.13% libc-2.12.so [.] usleep Then if I create 10000 *idle* cpu cgroups (no processes in them), cpu usage increases significantly although the 'wakers' are still executing in the root cpu cgroup: Cpu(s): 19.1%us, 48.7%sy, 0.0%ni, 31.6%id, 0.0%wa, 0.0%hi, 0.7%si Events: 230K cycles 24.56% [kernel] [k] tg_load_down 5.76% [kernel] [k] __schedule This happens because this particular kind of load triggers 'new idle' rebalance very frequently, which requires calling update_h_load(), which, in turn, calls tg_load_down() for every *idle* cpu cgroup even though it is absolutely useless, because idle cpu cgroups have no tasks to pull. This patch tries to improve the situation by making h_load calculation proceed only when h_load is really necessary. To achieve this, it substitutes update_h_load() with update_cfs_rq_h_load(), which computes h_load only for a given cfs_rq and all its ascendants, and makes the load balancer call this function whenever it considers if a task should be pulled, i.e. it moves h_load calculations directly to task_h_load(). For h_load of the same cfs_rq not to be updated multiple times (in case several tasks in the same cgroup are considered during the same balance run), the patch keeps the time of the last h_load update for each cfs_rq and breaks calculation when it finds h_load to be uptodate. The benefit of it is that h_load is computed only for those cfs_rq's, which really need it, in particular all idle task groups are skipped. Although this, in fact, moves h_load calculation under rq lock, it should not affect latency much, because the amount of work done under rq lock while trying to pull tasks is limited by sched_nr_migrate. After the patch applied with the setup described above (1000 wakers in the root cgroup and 10000 idle cgroups), I get: Cpu(s): 16.9%us, 24.8%sy, 0.0%ni, 58.4%id, 0.0%wa, 0.0%hi, 0.0%si Events: 242K cycles 7.57% [kernel] [k] __schedule 6.70% [kernel] [k] timerqueue_add 5.93% libc-2.12.so [.] usleep Signed-off-by: Vladimir Davydov <vdavydov@parallels.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/1373896159-1278-1-git-send-email-vdavydov@parallels.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-07-15 13:49:19 +00:00
struct rq *rq = rq_of(cfs_rq);
struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
sched, cgroup: Reduce rq->lock hold times for large cgroup hierarchies Peter Portante reported that for large cgroup hierarchies (and or on large CPU counts) we get immense lock contention on rq->lock and stuff stops working properly. His workload was a ton of processes, each in their own cgroup, everybody idling except for a sporadic wakeup once every so often. It was found that: schedule() idle_balance() load_balance() local_irq_save() double_rq_lock() update_h_load() walk_tg_tree(tg_load_down) tg_load_down() Results in an entire cgroup hierarchy walk under rq->lock for every new-idle balance and since new-idle balance isn't throttled this results in a lot of work while holding the rq->lock. This patch does two things, it removes the work from under rq->lock based on the good principle of race and pray which is widely employed in the load-balancer as a whole. And secondly it throttles the update_h_load() calculation to max once per jiffy. I considered excluding update_h_load() for new-idle balance all-together, but purely relying on regular balance passes to update this data might not work out under some rare circumstances where the new-idle busiest isn't the regular busiest for a while (unlikely, but a nightmare to debug if someone hits it and suffers). Cc: pjt@google.com Cc: Larry Woodman <lwoodman@redhat.com> Cc: Mike Galbraith <efault@gmx.de> Reported-by: Peter Portante <pportant@redhat.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/n/tip-aaarrzfpnaam7pqrekofu8a6@git.kernel.org Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
2012-08-08 19:46:40 +00:00
unsigned long now = jiffies;
sched: Move h_load calculation to task_h_load() The bad thing about update_h_load(), which computes hierarchical load factor for task groups, is that it is called for each task group in the system before every load balancer run, and since rebalance can be triggered very often, this function can eat really a lot of cpu time if there are many cpu cgroups in the system. Although the situation was improved significantly by commit a35b646 ('sched, cgroup: Reduce rq->lock hold times for large cgroup hierarchies'), the problem still can arise under some kinds of loads, e.g. when cpus are switching from idle to busy and back very frequently. For instance, when I start 1000 of processes that wake up every millisecond on my 8 cpus host, 'top' and 'perf top' show: Cpu(s): 17.8%us, 24.3%sy, 0.0%ni, 57.9%id, 0.0%wa, 0.0%hi, 0.0%si Events: 243K cycles 7.57% [kernel] [k] __schedule 7.08% [kernel] [k] timerqueue_add 6.13% libc-2.12.so [.] usleep Then if I create 10000 *idle* cpu cgroups (no processes in them), cpu usage increases significantly although the 'wakers' are still executing in the root cpu cgroup: Cpu(s): 19.1%us, 48.7%sy, 0.0%ni, 31.6%id, 0.0%wa, 0.0%hi, 0.7%si Events: 230K cycles 24.56% [kernel] [k] tg_load_down 5.76% [kernel] [k] __schedule This happens because this particular kind of load triggers 'new idle' rebalance very frequently, which requires calling update_h_load(), which, in turn, calls tg_load_down() for every *idle* cpu cgroup even though it is absolutely useless, because idle cpu cgroups have no tasks to pull. This patch tries to improve the situation by making h_load calculation proceed only when h_load is really necessary. To achieve this, it substitutes update_h_load() with update_cfs_rq_h_load(), which computes h_load only for a given cfs_rq and all its ascendants, and makes the load balancer call this function whenever it considers if a task should be pulled, i.e. it moves h_load calculations directly to task_h_load(). For h_load of the same cfs_rq not to be updated multiple times (in case several tasks in the same cgroup are considered during the same balance run), the patch keeps the time of the last h_load update for each cfs_rq and breaks calculation when it finds h_load to be uptodate. The benefit of it is that h_load is computed only for those cfs_rq's, which really need it, in particular all idle task groups are skipped. Although this, in fact, moves h_load calculation under rq lock, it should not affect latency much, because the amount of work done under rq lock while trying to pull tasks is limited by sched_nr_migrate. After the patch applied with the setup described above (1000 wakers in the root cgroup and 10000 idle cgroups), I get: Cpu(s): 16.9%us, 24.8%sy, 0.0%ni, 58.4%id, 0.0%wa, 0.0%hi, 0.0%si Events: 242K cycles 7.57% [kernel] [k] __schedule 6.70% [kernel] [k] timerqueue_add 5.93% libc-2.12.so [.] usleep Signed-off-by: Vladimir Davydov <vdavydov@parallels.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/1373896159-1278-1-git-send-email-vdavydov@parallels.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-07-15 13:49:19 +00:00
unsigned long load;
sched, cgroup: Reduce rq->lock hold times for large cgroup hierarchies Peter Portante reported that for large cgroup hierarchies (and or on large CPU counts) we get immense lock contention on rq->lock and stuff stops working properly. His workload was a ton of processes, each in their own cgroup, everybody idling except for a sporadic wakeup once every so often. It was found that: schedule() idle_balance() load_balance() local_irq_save() double_rq_lock() update_h_load() walk_tg_tree(tg_load_down) tg_load_down() Results in an entire cgroup hierarchy walk under rq->lock for every new-idle balance and since new-idle balance isn't throttled this results in a lot of work while holding the rq->lock. This patch does two things, it removes the work from under rq->lock based on the good principle of race and pray which is widely employed in the load-balancer as a whole. And secondly it throttles the update_h_load() calculation to max once per jiffy. I considered excluding update_h_load() for new-idle balance all-together, but purely relying on regular balance passes to update this data might not work out under some rare circumstances where the new-idle busiest isn't the regular busiest for a while (unlikely, but a nightmare to debug if someone hits it and suffers). Cc: pjt@google.com Cc: Larry Woodman <lwoodman@redhat.com> Cc: Mike Galbraith <efault@gmx.de> Reported-by: Peter Portante <pportant@redhat.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/n/tip-aaarrzfpnaam7pqrekofu8a6@git.kernel.org Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
2012-08-08 19:46:40 +00:00
sched: Move h_load calculation to task_h_load() The bad thing about update_h_load(), which computes hierarchical load factor for task groups, is that it is called for each task group in the system before every load balancer run, and since rebalance can be triggered very often, this function can eat really a lot of cpu time if there are many cpu cgroups in the system. Although the situation was improved significantly by commit a35b646 ('sched, cgroup: Reduce rq->lock hold times for large cgroup hierarchies'), the problem still can arise under some kinds of loads, e.g. when cpus are switching from idle to busy and back very frequently. For instance, when I start 1000 of processes that wake up every millisecond on my 8 cpus host, 'top' and 'perf top' show: Cpu(s): 17.8%us, 24.3%sy, 0.0%ni, 57.9%id, 0.0%wa, 0.0%hi, 0.0%si Events: 243K cycles 7.57% [kernel] [k] __schedule 7.08% [kernel] [k] timerqueue_add 6.13% libc-2.12.so [.] usleep Then if I create 10000 *idle* cpu cgroups (no processes in them), cpu usage increases significantly although the 'wakers' are still executing in the root cpu cgroup: Cpu(s): 19.1%us, 48.7%sy, 0.0%ni, 31.6%id, 0.0%wa, 0.0%hi, 0.7%si Events: 230K cycles 24.56% [kernel] [k] tg_load_down 5.76% [kernel] [k] __schedule This happens because this particular kind of load triggers 'new idle' rebalance very frequently, which requires calling update_h_load(), which, in turn, calls tg_load_down() for every *idle* cpu cgroup even though it is absolutely useless, because idle cpu cgroups have no tasks to pull. This patch tries to improve the situation by making h_load calculation proceed only when h_load is really necessary. To achieve this, it substitutes update_h_load() with update_cfs_rq_h_load(), which computes h_load only for a given cfs_rq and all its ascendants, and makes the load balancer call this function whenever it considers if a task should be pulled, i.e. it moves h_load calculations directly to task_h_load(). For h_load of the same cfs_rq not to be updated multiple times (in case several tasks in the same cgroup are considered during the same balance run), the patch keeps the time of the last h_load update for each cfs_rq and breaks calculation when it finds h_load to be uptodate. The benefit of it is that h_load is computed only for those cfs_rq's, which really need it, in particular all idle task groups are skipped. Although this, in fact, moves h_load calculation under rq lock, it should not affect latency much, because the amount of work done under rq lock while trying to pull tasks is limited by sched_nr_migrate. After the patch applied with the setup described above (1000 wakers in the root cgroup and 10000 idle cgroups), I get: Cpu(s): 16.9%us, 24.8%sy, 0.0%ni, 58.4%id, 0.0%wa, 0.0%hi, 0.0%si Events: 242K cycles 7.57% [kernel] [k] __schedule 6.70% [kernel] [k] timerqueue_add 5.93% libc-2.12.so [.] usleep Signed-off-by: Vladimir Davydov <vdavydov@parallels.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/1373896159-1278-1-git-send-email-vdavydov@parallels.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-07-15 13:49:19 +00:00
if (cfs_rq->last_h_load_update == now)
sched, cgroup: Reduce rq->lock hold times for large cgroup hierarchies Peter Portante reported that for large cgroup hierarchies (and or on large CPU counts) we get immense lock contention on rq->lock and stuff stops working properly. His workload was a ton of processes, each in their own cgroup, everybody idling except for a sporadic wakeup once every so often. It was found that: schedule() idle_balance() load_balance() local_irq_save() double_rq_lock() update_h_load() walk_tg_tree(tg_load_down) tg_load_down() Results in an entire cgroup hierarchy walk under rq->lock for every new-idle balance and since new-idle balance isn't throttled this results in a lot of work while holding the rq->lock. This patch does two things, it removes the work from under rq->lock based on the good principle of race and pray which is widely employed in the load-balancer as a whole. And secondly it throttles the update_h_load() calculation to max once per jiffy. I considered excluding update_h_load() for new-idle balance all-together, but purely relying on regular balance passes to update this data might not work out under some rare circumstances where the new-idle busiest isn't the regular busiest for a while (unlikely, but a nightmare to debug if someone hits it and suffers). Cc: pjt@google.com Cc: Larry Woodman <lwoodman@redhat.com> Cc: Mike Galbraith <efault@gmx.de> Reported-by: Peter Portante <pportant@redhat.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/n/tip-aaarrzfpnaam7pqrekofu8a6@git.kernel.org Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
2012-08-08 19:46:40 +00:00
return;
sched/fair: Do not re-read ->h_load_next during hierarchical load calculation A NULL pointer dereference bug was reported on a distribution kernel but the same issue should be present on mainline kernel. It occured on s390 but should not be arch-specific. A partial oops looks like: Unable to handle kernel pointer dereference in virtual kernel address space ... Call Trace: ... try_to_wake_up+0xfc/0x450 vhost_poll_wakeup+0x3a/0x50 [vhost] __wake_up_common+0xbc/0x178 __wake_up_common_lock+0x9e/0x160 __wake_up_sync_key+0x4e/0x60 sock_def_readable+0x5e/0x98 The bug hits any time between 1 hour to 3 days. The dereference occurs in update_cfs_rq_h_load when accumulating h_load. The problem is that cfq_rq->h_load_next is not protected by any locking and can be updated by parallel calls to task_h_load. Depending on the compiler, code may be generated that re-reads cfq_rq->h_load_next after the check for NULL and then oops when reading se->avg.load_avg. The dissassembly showed that it was possible to reread h_load_next after the check for NULL. While this does not appear to be an issue for later compilers, it's still an accident if the correct code is generated. Full locking in this path would have high overhead so this patch uses READ_ONCE to read h_load_next only once and check for NULL before dereferencing. It was confirmed that there were no further oops after 10 days of testing. As Peter pointed out, it is also necessary to use WRITE_ONCE() to avoid any potential problems with store tearing. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Valentin Schneider <valentin.schneider@arm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: <stable@vger.kernel.org> Fixes: 685207963be9 ("sched: Move h_load calculation to task_h_load()") Link: https://lkml.kernel.org/r/20190319123610.nsivgf3mjbjjesxb@techsingularity.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-03-19 12:36:10 +00:00
WRITE_ONCE(cfs_rq->h_load_next, NULL);
sched: Move h_load calculation to task_h_load() The bad thing about update_h_load(), which computes hierarchical load factor for task groups, is that it is called for each task group in the system before every load balancer run, and since rebalance can be triggered very often, this function can eat really a lot of cpu time if there are many cpu cgroups in the system. Although the situation was improved significantly by commit a35b646 ('sched, cgroup: Reduce rq->lock hold times for large cgroup hierarchies'), the problem still can arise under some kinds of loads, e.g. when cpus are switching from idle to busy and back very frequently. For instance, when I start 1000 of processes that wake up every millisecond on my 8 cpus host, 'top' and 'perf top' show: Cpu(s): 17.8%us, 24.3%sy, 0.0%ni, 57.9%id, 0.0%wa, 0.0%hi, 0.0%si Events: 243K cycles 7.57% [kernel] [k] __schedule 7.08% [kernel] [k] timerqueue_add 6.13% libc-2.12.so [.] usleep Then if I create 10000 *idle* cpu cgroups (no processes in them), cpu usage increases significantly although the 'wakers' are still executing in the root cpu cgroup: Cpu(s): 19.1%us, 48.7%sy, 0.0%ni, 31.6%id, 0.0%wa, 0.0%hi, 0.7%si Events: 230K cycles 24.56% [kernel] [k] tg_load_down 5.76% [kernel] [k] __schedule This happens because this particular kind of load triggers 'new idle' rebalance very frequently, which requires calling update_h_load(), which, in turn, calls tg_load_down() for every *idle* cpu cgroup even though it is absolutely useless, because idle cpu cgroups have no tasks to pull. This patch tries to improve the situation by making h_load calculation proceed only when h_load is really necessary. To achieve this, it substitutes update_h_load() with update_cfs_rq_h_load(), which computes h_load only for a given cfs_rq and all its ascendants, and makes the load balancer call this function whenever it considers if a task should be pulled, i.e. it moves h_load calculations directly to task_h_load(). For h_load of the same cfs_rq not to be updated multiple times (in case several tasks in the same cgroup are considered during the same balance run), the patch keeps the time of the last h_load update for each cfs_rq and breaks calculation when it finds h_load to be uptodate. The benefit of it is that h_load is computed only for those cfs_rq's, which really need it, in particular all idle task groups are skipped. Although this, in fact, moves h_load calculation under rq lock, it should not affect latency much, because the amount of work done under rq lock while trying to pull tasks is limited by sched_nr_migrate. After the patch applied with the setup described above (1000 wakers in the root cgroup and 10000 idle cgroups), I get: Cpu(s): 16.9%us, 24.8%sy, 0.0%ni, 58.4%id, 0.0%wa, 0.0%hi, 0.0%si Events: 242K cycles 7.57% [kernel] [k] __schedule 6.70% [kernel] [k] timerqueue_add 5.93% libc-2.12.so [.] usleep Signed-off-by: Vladimir Davydov <vdavydov@parallels.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/1373896159-1278-1-git-send-email-vdavydov@parallels.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-07-15 13:49:19 +00:00
for_each_sched_entity(se) {
cfs_rq = cfs_rq_of(se);
sched/fair: Do not re-read ->h_load_next during hierarchical load calculation A NULL pointer dereference bug was reported on a distribution kernel but the same issue should be present on mainline kernel. It occured on s390 but should not be arch-specific. A partial oops looks like: Unable to handle kernel pointer dereference in virtual kernel address space ... Call Trace: ... try_to_wake_up+0xfc/0x450 vhost_poll_wakeup+0x3a/0x50 [vhost] __wake_up_common+0xbc/0x178 __wake_up_common_lock+0x9e/0x160 __wake_up_sync_key+0x4e/0x60 sock_def_readable+0x5e/0x98 The bug hits any time between 1 hour to 3 days. The dereference occurs in update_cfs_rq_h_load when accumulating h_load. The problem is that cfq_rq->h_load_next is not protected by any locking and can be updated by parallel calls to task_h_load. Depending on the compiler, code may be generated that re-reads cfq_rq->h_load_next after the check for NULL and then oops when reading se->avg.load_avg. The dissassembly showed that it was possible to reread h_load_next after the check for NULL. While this does not appear to be an issue for later compilers, it's still an accident if the correct code is generated. Full locking in this path would have high overhead so this patch uses READ_ONCE to read h_load_next only once and check for NULL before dereferencing. It was confirmed that there were no further oops after 10 days of testing. As Peter pointed out, it is also necessary to use WRITE_ONCE() to avoid any potential problems with store tearing. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Valentin Schneider <valentin.schneider@arm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: <stable@vger.kernel.org> Fixes: 685207963be9 ("sched: Move h_load calculation to task_h_load()") Link: https://lkml.kernel.org/r/20190319123610.nsivgf3mjbjjesxb@techsingularity.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-03-19 12:36:10 +00:00
WRITE_ONCE(cfs_rq->h_load_next, se);
sched: Move h_load calculation to task_h_load() The bad thing about update_h_load(), which computes hierarchical load factor for task groups, is that it is called for each task group in the system before every load balancer run, and since rebalance can be triggered very often, this function can eat really a lot of cpu time if there are many cpu cgroups in the system. Although the situation was improved significantly by commit a35b646 ('sched, cgroup: Reduce rq->lock hold times for large cgroup hierarchies'), the problem still can arise under some kinds of loads, e.g. when cpus are switching from idle to busy and back very frequently. For instance, when I start 1000 of processes that wake up every millisecond on my 8 cpus host, 'top' and 'perf top' show: Cpu(s): 17.8%us, 24.3%sy, 0.0%ni, 57.9%id, 0.0%wa, 0.0%hi, 0.0%si Events: 243K cycles 7.57% [kernel] [k] __schedule 7.08% [kernel] [k] timerqueue_add 6.13% libc-2.12.so [.] usleep Then if I create 10000 *idle* cpu cgroups (no processes in them), cpu usage increases significantly although the 'wakers' are still executing in the root cpu cgroup: Cpu(s): 19.1%us, 48.7%sy, 0.0%ni, 31.6%id, 0.0%wa, 0.0%hi, 0.7%si Events: 230K cycles 24.56% [kernel] [k] tg_load_down 5.76% [kernel] [k] __schedule This happens because this particular kind of load triggers 'new idle' rebalance very frequently, which requires calling update_h_load(), which, in turn, calls tg_load_down() for every *idle* cpu cgroup even though it is absolutely useless, because idle cpu cgroups have no tasks to pull. This patch tries to improve the situation by making h_load calculation proceed only when h_load is really necessary. To achieve this, it substitutes update_h_load() with update_cfs_rq_h_load(), which computes h_load only for a given cfs_rq and all its ascendants, and makes the load balancer call this function whenever it considers if a task should be pulled, i.e. it moves h_load calculations directly to task_h_load(). For h_load of the same cfs_rq not to be updated multiple times (in case several tasks in the same cgroup are considered during the same balance run), the patch keeps the time of the last h_load update for each cfs_rq and breaks calculation when it finds h_load to be uptodate. The benefit of it is that h_load is computed only for those cfs_rq's, which really need it, in particular all idle task groups are skipped. Although this, in fact, moves h_load calculation under rq lock, it should not affect latency much, because the amount of work done under rq lock while trying to pull tasks is limited by sched_nr_migrate. After the patch applied with the setup described above (1000 wakers in the root cgroup and 10000 idle cgroups), I get: Cpu(s): 16.9%us, 24.8%sy, 0.0%ni, 58.4%id, 0.0%wa, 0.0%hi, 0.0%si Events: 242K cycles 7.57% [kernel] [k] __schedule 6.70% [kernel] [k] timerqueue_add 5.93% libc-2.12.so [.] usleep Signed-off-by: Vladimir Davydov <vdavydov@parallels.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/1373896159-1278-1-git-send-email-vdavydov@parallels.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-07-15 13:49:19 +00:00
if (cfs_rq->last_h_load_update == now)
break;
}
sched, cgroup: Reduce rq->lock hold times for large cgroup hierarchies Peter Portante reported that for large cgroup hierarchies (and or on large CPU counts) we get immense lock contention on rq->lock and stuff stops working properly. His workload was a ton of processes, each in their own cgroup, everybody idling except for a sporadic wakeup once every so often. It was found that: schedule() idle_balance() load_balance() local_irq_save() double_rq_lock() update_h_load() walk_tg_tree(tg_load_down) tg_load_down() Results in an entire cgroup hierarchy walk under rq->lock for every new-idle balance and since new-idle balance isn't throttled this results in a lot of work while holding the rq->lock. This patch does two things, it removes the work from under rq->lock based on the good principle of race and pray which is widely employed in the load-balancer as a whole. And secondly it throttles the update_h_load() calculation to max once per jiffy. I considered excluding update_h_load() for new-idle balance all-together, but purely relying on regular balance passes to update this data might not work out under some rare circumstances where the new-idle busiest isn't the regular busiest for a while (unlikely, but a nightmare to debug if someone hits it and suffers). Cc: pjt@google.com Cc: Larry Woodman <lwoodman@redhat.com> Cc: Mike Galbraith <efault@gmx.de> Reported-by: Peter Portante <pportant@redhat.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/n/tip-aaarrzfpnaam7pqrekofu8a6@git.kernel.org Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
2012-08-08 19:46:40 +00:00
sched: Move h_load calculation to task_h_load() The bad thing about update_h_load(), which computes hierarchical load factor for task groups, is that it is called for each task group in the system before every load balancer run, and since rebalance can be triggered very often, this function can eat really a lot of cpu time if there are many cpu cgroups in the system. Although the situation was improved significantly by commit a35b646 ('sched, cgroup: Reduce rq->lock hold times for large cgroup hierarchies'), the problem still can arise under some kinds of loads, e.g. when cpus are switching from idle to busy and back very frequently. For instance, when I start 1000 of processes that wake up every millisecond on my 8 cpus host, 'top' and 'perf top' show: Cpu(s): 17.8%us, 24.3%sy, 0.0%ni, 57.9%id, 0.0%wa, 0.0%hi, 0.0%si Events: 243K cycles 7.57% [kernel] [k] __schedule 7.08% [kernel] [k] timerqueue_add 6.13% libc-2.12.so [.] usleep Then if I create 10000 *idle* cpu cgroups (no processes in them), cpu usage increases significantly although the 'wakers' are still executing in the root cpu cgroup: Cpu(s): 19.1%us, 48.7%sy, 0.0%ni, 31.6%id, 0.0%wa, 0.0%hi, 0.7%si Events: 230K cycles 24.56% [kernel] [k] tg_load_down 5.76% [kernel] [k] __schedule This happens because this particular kind of load triggers 'new idle' rebalance very frequently, which requires calling update_h_load(), which, in turn, calls tg_load_down() for every *idle* cpu cgroup even though it is absolutely useless, because idle cpu cgroups have no tasks to pull. This patch tries to improve the situation by making h_load calculation proceed only when h_load is really necessary. To achieve this, it substitutes update_h_load() with update_cfs_rq_h_load(), which computes h_load only for a given cfs_rq and all its ascendants, and makes the load balancer call this function whenever it considers if a task should be pulled, i.e. it moves h_load calculations directly to task_h_load(). For h_load of the same cfs_rq not to be updated multiple times (in case several tasks in the same cgroup are considered during the same balance run), the patch keeps the time of the last h_load update for each cfs_rq and breaks calculation when it finds h_load to be uptodate. The benefit of it is that h_load is computed only for those cfs_rq's, which really need it, in particular all idle task groups are skipped. Although this, in fact, moves h_load calculation under rq lock, it should not affect latency much, because the amount of work done under rq lock while trying to pull tasks is limited by sched_nr_migrate. After the patch applied with the setup described above (1000 wakers in the root cgroup and 10000 idle cgroups), I get: Cpu(s): 16.9%us, 24.8%sy, 0.0%ni, 58.4%id, 0.0%wa, 0.0%hi, 0.0%si Events: 242K cycles 7.57% [kernel] [k] __schedule 6.70% [kernel] [k] timerqueue_add 5.93% libc-2.12.so [.] usleep Signed-off-by: Vladimir Davydov <vdavydov@parallels.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/1373896159-1278-1-git-send-email-vdavydov@parallels.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-07-15 13:49:19 +00:00
if (!se) {
cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
sched: Move h_load calculation to task_h_load() The bad thing about update_h_load(), which computes hierarchical load factor for task groups, is that it is called for each task group in the system before every load balancer run, and since rebalance can be triggered very often, this function can eat really a lot of cpu time if there are many cpu cgroups in the system. Although the situation was improved significantly by commit a35b646 ('sched, cgroup: Reduce rq->lock hold times for large cgroup hierarchies'), the problem still can arise under some kinds of loads, e.g. when cpus are switching from idle to busy and back very frequently. For instance, when I start 1000 of processes that wake up every millisecond on my 8 cpus host, 'top' and 'perf top' show: Cpu(s): 17.8%us, 24.3%sy, 0.0%ni, 57.9%id, 0.0%wa, 0.0%hi, 0.0%si Events: 243K cycles 7.57% [kernel] [k] __schedule 7.08% [kernel] [k] timerqueue_add 6.13% libc-2.12.so [.] usleep Then if I create 10000 *idle* cpu cgroups (no processes in them), cpu usage increases significantly although the 'wakers' are still executing in the root cpu cgroup: Cpu(s): 19.1%us, 48.7%sy, 0.0%ni, 31.6%id, 0.0%wa, 0.0%hi, 0.7%si Events: 230K cycles 24.56% [kernel] [k] tg_load_down 5.76% [kernel] [k] __schedule This happens because this particular kind of load triggers 'new idle' rebalance very frequently, which requires calling update_h_load(), which, in turn, calls tg_load_down() for every *idle* cpu cgroup even though it is absolutely useless, because idle cpu cgroups have no tasks to pull. This patch tries to improve the situation by making h_load calculation proceed only when h_load is really necessary. To achieve this, it substitutes update_h_load() with update_cfs_rq_h_load(), which computes h_load only for a given cfs_rq and all its ascendants, and makes the load balancer call this function whenever it considers if a task should be pulled, i.e. it moves h_load calculations directly to task_h_load(). For h_load of the same cfs_rq not to be updated multiple times (in case several tasks in the same cgroup are considered during the same balance run), the patch keeps the time of the last h_load update for each cfs_rq and breaks calculation when it finds h_load to be uptodate. The benefit of it is that h_load is computed only for those cfs_rq's, which really need it, in particular all idle task groups are skipped. Although this, in fact, moves h_load calculation under rq lock, it should not affect latency much, because the amount of work done under rq lock while trying to pull tasks is limited by sched_nr_migrate. After the patch applied with the setup described above (1000 wakers in the root cgroup and 10000 idle cgroups), I get: Cpu(s): 16.9%us, 24.8%sy, 0.0%ni, 58.4%id, 0.0%wa, 0.0%hi, 0.0%si Events: 242K cycles 7.57% [kernel] [k] __schedule 6.70% [kernel] [k] timerqueue_add 5.93% libc-2.12.so [.] usleep Signed-off-by: Vladimir Davydov <vdavydov@parallels.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/1373896159-1278-1-git-send-email-vdavydov@parallels.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-07-15 13:49:19 +00:00
cfs_rq->last_h_load_update = now;
}
sched/fair: Do not re-read ->h_load_next during hierarchical load calculation A NULL pointer dereference bug was reported on a distribution kernel but the same issue should be present on mainline kernel. It occured on s390 but should not be arch-specific. A partial oops looks like: Unable to handle kernel pointer dereference in virtual kernel address space ... Call Trace: ... try_to_wake_up+0xfc/0x450 vhost_poll_wakeup+0x3a/0x50 [vhost] __wake_up_common+0xbc/0x178 __wake_up_common_lock+0x9e/0x160 __wake_up_sync_key+0x4e/0x60 sock_def_readable+0x5e/0x98 The bug hits any time between 1 hour to 3 days. The dereference occurs in update_cfs_rq_h_load when accumulating h_load. The problem is that cfq_rq->h_load_next is not protected by any locking and can be updated by parallel calls to task_h_load. Depending on the compiler, code may be generated that re-reads cfq_rq->h_load_next after the check for NULL and then oops when reading se->avg.load_avg. The dissassembly showed that it was possible to reread h_load_next after the check for NULL. While this does not appear to be an issue for later compilers, it's still an accident if the correct code is generated. Full locking in this path would have high overhead so this patch uses READ_ONCE to read h_load_next only once and check for NULL before dereferencing. It was confirmed that there were no further oops after 10 days of testing. As Peter pointed out, it is also necessary to use WRITE_ONCE() to avoid any potential problems with store tearing. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Valentin Schneider <valentin.schneider@arm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: <stable@vger.kernel.org> Fixes: 685207963be9 ("sched: Move h_load calculation to task_h_load()") Link: https://lkml.kernel.org/r/20190319123610.nsivgf3mjbjjesxb@techsingularity.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-03-19 12:36:10 +00:00
while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
sched: Move h_load calculation to task_h_load() The bad thing about update_h_load(), which computes hierarchical load factor for task groups, is that it is called for each task group in the system before every load balancer run, and since rebalance can be triggered very often, this function can eat really a lot of cpu time if there are many cpu cgroups in the system. Although the situation was improved significantly by commit a35b646 ('sched, cgroup: Reduce rq->lock hold times for large cgroup hierarchies'), the problem still can arise under some kinds of loads, e.g. when cpus are switching from idle to busy and back very frequently. For instance, when I start 1000 of processes that wake up every millisecond on my 8 cpus host, 'top' and 'perf top' show: Cpu(s): 17.8%us, 24.3%sy, 0.0%ni, 57.9%id, 0.0%wa, 0.0%hi, 0.0%si Events: 243K cycles 7.57% [kernel] [k] __schedule 7.08% [kernel] [k] timerqueue_add 6.13% libc-2.12.so [.] usleep Then if I create 10000 *idle* cpu cgroups (no processes in them), cpu usage increases significantly although the 'wakers' are still executing in the root cpu cgroup: Cpu(s): 19.1%us, 48.7%sy, 0.0%ni, 31.6%id, 0.0%wa, 0.0%hi, 0.7%si Events: 230K cycles 24.56% [kernel] [k] tg_load_down 5.76% [kernel] [k] __schedule This happens because this particular kind of load triggers 'new idle' rebalance very frequently, which requires calling update_h_load(), which, in turn, calls tg_load_down() for every *idle* cpu cgroup even though it is absolutely useless, because idle cpu cgroups have no tasks to pull. This patch tries to improve the situation by making h_load calculation proceed only when h_load is really necessary. To achieve this, it substitutes update_h_load() with update_cfs_rq_h_load(), which computes h_load only for a given cfs_rq and all its ascendants, and makes the load balancer call this function whenever it considers if a task should be pulled, i.e. it moves h_load calculations directly to task_h_load(). For h_load of the same cfs_rq not to be updated multiple times (in case several tasks in the same cgroup are considered during the same balance run), the patch keeps the time of the last h_load update for each cfs_rq and breaks calculation when it finds h_load to be uptodate. The benefit of it is that h_load is computed only for those cfs_rq's, which really need it, in particular all idle task groups are skipped. Although this, in fact, moves h_load calculation under rq lock, it should not affect latency much, because the amount of work done under rq lock while trying to pull tasks is limited by sched_nr_migrate. After the patch applied with the setup described above (1000 wakers in the root cgroup and 10000 idle cgroups), I get: Cpu(s): 16.9%us, 24.8%sy, 0.0%ni, 58.4%id, 0.0%wa, 0.0%hi, 0.0%si Events: 242K cycles 7.57% [kernel] [k] __schedule 6.70% [kernel] [k] timerqueue_add 5.93% libc-2.12.so [.] usleep Signed-off-by: Vladimir Davydov <vdavydov@parallels.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/1373896159-1278-1-git-send-email-vdavydov@parallels.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-07-15 13:49:19 +00:00
load = cfs_rq->h_load;
load = div64_ul(load * se->avg.load_avg,
cfs_rq_load_avg(cfs_rq) + 1);
sched: Move h_load calculation to task_h_load() The bad thing about update_h_load(), which computes hierarchical load factor for task groups, is that it is called for each task group in the system before every load balancer run, and since rebalance can be triggered very often, this function can eat really a lot of cpu time if there are many cpu cgroups in the system. Although the situation was improved significantly by commit a35b646 ('sched, cgroup: Reduce rq->lock hold times for large cgroup hierarchies'), the problem still can arise under some kinds of loads, e.g. when cpus are switching from idle to busy and back very frequently. For instance, when I start 1000 of processes that wake up every millisecond on my 8 cpus host, 'top' and 'perf top' show: Cpu(s): 17.8%us, 24.3%sy, 0.0%ni, 57.9%id, 0.0%wa, 0.0%hi, 0.0%si Events: 243K cycles 7.57% [kernel] [k] __schedule 7.08% [kernel] [k] timerqueue_add 6.13% libc-2.12.so [.] usleep Then if I create 10000 *idle* cpu cgroups (no processes in them), cpu usage increases significantly although the 'wakers' are still executing in the root cpu cgroup: Cpu(s): 19.1%us, 48.7%sy, 0.0%ni, 31.6%id, 0.0%wa, 0.0%hi, 0.7%si Events: 230K cycles 24.56% [kernel] [k] tg_load_down 5.76% [kernel] [k] __schedule This happens because this particular kind of load triggers 'new idle' rebalance very frequently, which requires calling update_h_load(), which, in turn, calls tg_load_down() for every *idle* cpu cgroup even though it is absolutely useless, because idle cpu cgroups have no tasks to pull. This patch tries to improve the situation by making h_load calculation proceed only when h_load is really necessary. To achieve this, it substitutes update_h_load() with update_cfs_rq_h_load(), which computes h_load only for a given cfs_rq and all its ascendants, and makes the load balancer call this function whenever it considers if a task should be pulled, i.e. it moves h_load calculations directly to task_h_load(). For h_load of the same cfs_rq not to be updated multiple times (in case several tasks in the same cgroup are considered during the same balance run), the patch keeps the time of the last h_load update for each cfs_rq and breaks calculation when it finds h_load to be uptodate. The benefit of it is that h_load is computed only for those cfs_rq's, which really need it, in particular all idle task groups are skipped. Although this, in fact, moves h_load calculation under rq lock, it should not affect latency much, because the amount of work done under rq lock while trying to pull tasks is limited by sched_nr_migrate. After the patch applied with the setup described above (1000 wakers in the root cgroup and 10000 idle cgroups), I get: Cpu(s): 16.9%us, 24.8%sy, 0.0%ni, 58.4%id, 0.0%wa, 0.0%hi, 0.0%si Events: 242K cycles 7.57% [kernel] [k] __schedule 6.70% [kernel] [k] timerqueue_add 5.93% libc-2.12.so [.] usleep Signed-off-by: Vladimir Davydov <vdavydov@parallels.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/1373896159-1278-1-git-send-email-vdavydov@parallels.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-07-15 13:49:19 +00:00
cfs_rq = group_cfs_rq(se);
cfs_rq->h_load = load;
cfs_rq->last_h_load_update = now;
}
}
static unsigned long task_h_load(struct task_struct *p)
{
struct cfs_rq *cfs_rq = task_cfs_rq(p);
sched: Move h_load calculation to task_h_load() The bad thing about update_h_load(), which computes hierarchical load factor for task groups, is that it is called for each task group in the system before every load balancer run, and since rebalance can be triggered very often, this function can eat really a lot of cpu time if there are many cpu cgroups in the system. Although the situation was improved significantly by commit a35b646 ('sched, cgroup: Reduce rq->lock hold times for large cgroup hierarchies'), the problem still can arise under some kinds of loads, e.g. when cpus are switching from idle to busy and back very frequently. For instance, when I start 1000 of processes that wake up every millisecond on my 8 cpus host, 'top' and 'perf top' show: Cpu(s): 17.8%us, 24.3%sy, 0.0%ni, 57.9%id, 0.0%wa, 0.0%hi, 0.0%si Events: 243K cycles 7.57% [kernel] [k] __schedule 7.08% [kernel] [k] timerqueue_add 6.13% libc-2.12.so [.] usleep Then if I create 10000 *idle* cpu cgroups (no processes in them), cpu usage increases significantly although the 'wakers' are still executing in the root cpu cgroup: Cpu(s): 19.1%us, 48.7%sy, 0.0%ni, 31.6%id, 0.0%wa, 0.0%hi, 0.7%si Events: 230K cycles 24.56% [kernel] [k] tg_load_down 5.76% [kernel] [k] __schedule This happens because this particular kind of load triggers 'new idle' rebalance very frequently, which requires calling update_h_load(), which, in turn, calls tg_load_down() for every *idle* cpu cgroup even though it is absolutely useless, because idle cpu cgroups have no tasks to pull. This patch tries to improve the situation by making h_load calculation proceed only when h_load is really necessary. To achieve this, it substitutes update_h_load() with update_cfs_rq_h_load(), which computes h_load only for a given cfs_rq and all its ascendants, and makes the load balancer call this function whenever it considers if a task should be pulled, i.e. it moves h_load calculations directly to task_h_load(). For h_load of the same cfs_rq not to be updated multiple times (in case several tasks in the same cgroup are considered during the same balance run), the patch keeps the time of the last h_load update for each cfs_rq and breaks calculation when it finds h_load to be uptodate. The benefit of it is that h_load is computed only for those cfs_rq's, which really need it, in particular all idle task groups are skipped. Although this, in fact, moves h_load calculation under rq lock, it should not affect latency much, because the amount of work done under rq lock while trying to pull tasks is limited by sched_nr_migrate. After the patch applied with the setup described above (1000 wakers in the root cgroup and 10000 idle cgroups), I get: Cpu(s): 16.9%us, 24.8%sy, 0.0%ni, 58.4%id, 0.0%wa, 0.0%hi, 0.0%si Events: 242K cycles 7.57% [kernel] [k] __schedule 6.70% [kernel] [k] timerqueue_add 5.93% libc-2.12.so [.] usleep Signed-off-by: Vladimir Davydov <vdavydov@parallels.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/1373896159-1278-1-git-send-email-vdavydov@parallels.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-07-15 13:49:19 +00:00
update_cfs_rq_h_load(cfs_rq);
sched/fair: Rewrite runnable load and utilization average tracking The idea of runnable load average (let runnable time contribute to weight) was proposed by Paul Turner and Ben Segall, and it is still followed by this rewrite. This rewrite aims to solve the following issues: 1. cfs_rq's load average (namely runnable_load_avg and blocked_load_avg) is updated at the granularity of an entity at a time, which results in the cfs_rq's load average is stale or partially updated: at any time, only one entity is up to date, all other entities are effectively lagging behind. This is undesirable. To illustrate, if we have n runnable entities in the cfs_rq, as time elapses, they certainly become outdated: t0: cfs_rq { e1_old, e2_old, ..., en_old } and when we update: t1: update e1, then we have cfs_rq { e1_new, e2_old, ..., en_old } t2: update e2, then we have cfs_rq { e1_old, e2_new, ..., en_old } ... We solve this by combining all runnable entities' load averages together in cfs_rq's avg, and update the cfs_rq's avg as a whole. This is based on the fact that if we regard the update as a function, then: w * update(e) = update(w * e) and update(e1) + update(e2) = update(e1 + e2), then w1 * update(e1) + w2 * update(e2) = update(w1 * e1 + w2 * e2) therefore, by this rewrite, we have an entirely updated cfs_rq at the time we update it: t1: update cfs_rq { e1_new, e2_new, ..., en_new } t2: update cfs_rq { e1_new, e2_new, ..., en_new } ... 2. cfs_rq's load average is different between top rq->cfs_rq and other task_group's per CPU cfs_rqs in whether or not blocked_load_average contributes to the load. The basic idea behind runnable load average (the same for utilization) is that the blocked state is taken into account as opposed to only accounting for the currently runnable state. Therefore, the average should include both the runnable/running and blocked load averages. This rewrite does that. In addition, we also combine runnable/running and blocked averages of all entities into the cfs_rq's average, and update it together at once. This is based on the fact that: update(runnable) + update(blocked) = update(runnable + blocked) This significantly reduces the code as we don't need to separately maintain/update runnable/running load and blocked load. 3. How task_group entities' share is calculated is complex and imprecise. We reduce the complexity in this rewrite to allow a very simple rule: the task_group's load_avg is aggregated from its per CPU cfs_rqs's load_avgs. Then group entity's weight is simply proportional to its own cfs_rq's load_avg / task_group's load_avg. To illustrate, if a task_group has { cfs_rq1, cfs_rq2, ..., cfs_rqn }, then, task_group_avg = cfs_rq1_avg + cfs_rq2_avg + ... + cfs_rqn_avg, then cfs_rqx's entity's share = cfs_rqx_avg / task_group_avg * task_group's share To sum up, this rewrite in principle is equivalent to the current one, but fixes the issues described above. Turns out, it significantly reduces the code complexity and hence increases clarity and efficiency. In addition, the new averages are more smooth/continuous (no spurious spikes and valleys) and updated more consistently and quickly to reflect the load dynamics. As a result, we have less load tracking overhead, better performance, and especially better power efficiency due to more balanced load. Signed-off-by: Yuyang Du <yuyang.du@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: arjan@linux.intel.com Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: fengguang.wu@intel.com Cc: len.brown@intel.com Cc: morten.rasmussen@arm.com Cc: pjt@google.com Cc: rafael.j.wysocki@intel.com Cc: umgwanakikbuti@gmail.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1436918682-4971-3-git-send-email-yuyang.du@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-07-15 00:04:37 +00:00
return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
cfs_rq_load_avg(cfs_rq) + 1);
}
#else
sched/cpufreq: Move the cfs_rq_util_change() call to cpufreq_update_util() update_cfs_rq_load_avg() calls cfs_rq_util_change() every time PELT decays, which might be inefficient when the cpufreq driver has rate limitation. When a task is attached on a CPU, we have this call path: update_load_avg() update_cfs_rq_load_avg() cfs_rq_util_change -- > trig frequency update attach_entity_load_avg() cfs_rq_util_change -- > trig frequency update The 1st frequency update will not take into account the utilization of the newly attached task and the 2nd one might be discarded because of rate limitation of the cpufreq driver. update_cfs_rq_load_avg() is only called by update_blocked_averages() and update_load_avg() so we can move the call to cfs_rq_util_change/cpufreq_update_util() into these two functions. It's also interesting to note that update_load_avg() already calls cfs_rq_util_change() directly for the !SMP case. This change will also ensure that cpufreq_update_util() is called even when there is no more CFS rq in the leaf_cfs_rq_list to update, but only IRQ, RT or DL PELT signals. [ mingo: Minor updates. ] Reported-by: Doug Smythies <dsmythies@telus.net> Tested-by: Doug Smythies <dsmythies@telus.net> Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Acked-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: juri.lelli@redhat.com Cc: linux-pm@vger.kernel.org Cc: mgorman@suse.de Cc: rostedt@goodmis.org Cc: sargun@sargun.me Cc: srinivas.pandruvada@linux.intel.com Cc: tj@kernel.org Cc: xiexiuqi@huawei.com Cc: xiezhipeng1@huawei.com Fixes: 039ae8bcf7a5 ("sched/fair: Fix O(nr_cgroups) in the load balancing path") Link: https://lkml.kernel.org/r/1574083279-799-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-11-18 13:21:19 +00:00
static bool __update_blocked_fair(struct rq *rq, bool *done)
{
struct cfs_rq *cfs_rq = &rq->cfs;
sched/cpufreq: Move the cfs_rq_util_change() call to cpufreq_update_util() update_cfs_rq_load_avg() calls cfs_rq_util_change() every time PELT decays, which might be inefficient when the cpufreq driver has rate limitation. When a task is attached on a CPU, we have this call path: update_load_avg() update_cfs_rq_load_avg() cfs_rq_util_change -- > trig frequency update attach_entity_load_avg() cfs_rq_util_change -- > trig frequency update The 1st frequency update will not take into account the utilization of the newly attached task and the 2nd one might be discarded because of rate limitation of the cpufreq driver. update_cfs_rq_load_avg() is only called by update_blocked_averages() and update_load_avg() so we can move the call to cfs_rq_util_change/cpufreq_update_util() into these two functions. It's also interesting to note that update_load_avg() already calls cfs_rq_util_change() directly for the !SMP case. This change will also ensure that cpufreq_update_util() is called even when there is no more CFS rq in the leaf_cfs_rq_list to update, but only IRQ, RT or DL PELT signals. [ mingo: Minor updates. ] Reported-by: Doug Smythies <dsmythies@telus.net> Tested-by: Doug Smythies <dsmythies@telus.net> Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Acked-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: juri.lelli@redhat.com Cc: linux-pm@vger.kernel.org Cc: mgorman@suse.de Cc: rostedt@goodmis.org Cc: sargun@sargun.me Cc: srinivas.pandruvada@linux.intel.com Cc: tj@kernel.org Cc: xiexiuqi@huawei.com Cc: xiezhipeng1@huawei.com Fixes: 039ae8bcf7a5 ("sched/fair: Fix O(nr_cgroups) in the load balancing path") Link: https://lkml.kernel.org/r/1574083279-799-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-11-18 13:21:19 +00:00
bool decayed;
sched/cpufreq: Move the cfs_rq_util_change() call to cpufreq_update_util() update_cfs_rq_load_avg() calls cfs_rq_util_change() every time PELT decays, which might be inefficient when the cpufreq driver has rate limitation. When a task is attached on a CPU, we have this call path: update_load_avg() update_cfs_rq_load_avg() cfs_rq_util_change -- > trig frequency update attach_entity_load_avg() cfs_rq_util_change -- > trig frequency update The 1st frequency update will not take into account the utilization of the newly attached task and the 2nd one might be discarded because of rate limitation of the cpufreq driver. update_cfs_rq_load_avg() is only called by update_blocked_averages() and update_load_avg() so we can move the call to cfs_rq_util_change/cpufreq_update_util() into these two functions. It's also interesting to note that update_load_avg() already calls cfs_rq_util_change() directly for the !SMP case. This change will also ensure that cpufreq_update_util() is called even when there is no more CFS rq in the leaf_cfs_rq_list to update, but only IRQ, RT or DL PELT signals. [ mingo: Minor updates. ] Reported-by: Doug Smythies <dsmythies@telus.net> Tested-by: Doug Smythies <dsmythies@telus.net> Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Acked-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: juri.lelli@redhat.com Cc: linux-pm@vger.kernel.org Cc: mgorman@suse.de Cc: rostedt@goodmis.org Cc: sargun@sargun.me Cc: srinivas.pandruvada@linux.intel.com Cc: tj@kernel.org Cc: xiexiuqi@huawei.com Cc: xiezhipeng1@huawei.com Fixes: 039ae8bcf7a5 ("sched/fair: Fix O(nr_cgroups) in the load balancing path") Link: https://lkml.kernel.org/r/1574083279-799-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-11-18 13:21:19 +00:00
decayed = update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq);
if (cfs_rq_has_blocked(cfs_rq))
*done = false;
sched/cpufreq: Move the cfs_rq_util_change() call to cpufreq_update_util() update_cfs_rq_load_avg() calls cfs_rq_util_change() every time PELT decays, which might be inefficient when the cpufreq driver has rate limitation. When a task is attached on a CPU, we have this call path: update_load_avg() update_cfs_rq_load_avg() cfs_rq_util_change -- > trig frequency update attach_entity_load_avg() cfs_rq_util_change -- > trig frequency update The 1st frequency update will not take into account the utilization of the newly attached task and the 2nd one might be discarded because of rate limitation of the cpufreq driver. update_cfs_rq_load_avg() is only called by update_blocked_averages() and update_load_avg() so we can move the call to cfs_rq_util_change/cpufreq_update_util() into these two functions. It's also interesting to note that update_load_avg() already calls cfs_rq_util_change() directly for the !SMP case. This change will also ensure that cpufreq_update_util() is called even when there is no more CFS rq in the leaf_cfs_rq_list to update, but only IRQ, RT or DL PELT signals. [ mingo: Minor updates. ] Reported-by: Doug Smythies <dsmythies@telus.net> Tested-by: Doug Smythies <dsmythies@telus.net> Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Acked-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: juri.lelli@redhat.com Cc: linux-pm@vger.kernel.org Cc: mgorman@suse.de Cc: rostedt@goodmis.org Cc: sargun@sargun.me Cc: srinivas.pandruvada@linux.intel.com Cc: tj@kernel.org Cc: xiexiuqi@huawei.com Cc: xiezhipeng1@huawei.com Fixes: 039ae8bcf7a5 ("sched/fair: Fix O(nr_cgroups) in the load balancing path") Link: https://lkml.kernel.org/r/1574083279-799-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-11-18 13:21:19 +00:00
return decayed;
}
static unsigned long task_h_load(struct task_struct *p)
{
sched/fair: Rewrite runnable load and utilization average tracking The idea of runnable load average (let runnable time contribute to weight) was proposed by Paul Turner and Ben Segall, and it is still followed by this rewrite. This rewrite aims to solve the following issues: 1. cfs_rq's load average (namely runnable_load_avg and blocked_load_avg) is updated at the granularity of an entity at a time, which results in the cfs_rq's load average is stale or partially updated: at any time, only one entity is up to date, all other entities are effectively lagging behind. This is undesirable. To illustrate, if we have n runnable entities in the cfs_rq, as time elapses, they certainly become outdated: t0: cfs_rq { e1_old, e2_old, ..., en_old } and when we update: t1: update e1, then we have cfs_rq { e1_new, e2_old, ..., en_old } t2: update e2, then we have cfs_rq { e1_old, e2_new, ..., en_old } ... We solve this by combining all runnable entities' load averages together in cfs_rq's avg, and update the cfs_rq's avg as a whole. This is based on the fact that if we regard the update as a function, then: w * update(e) = update(w * e) and update(e1) + update(e2) = update(e1 + e2), then w1 * update(e1) + w2 * update(e2) = update(w1 * e1 + w2 * e2) therefore, by this rewrite, we have an entirely updated cfs_rq at the time we update it: t1: update cfs_rq { e1_new, e2_new, ..., en_new } t2: update cfs_rq { e1_new, e2_new, ..., en_new } ... 2. cfs_rq's load average is different between top rq->cfs_rq and other task_group's per CPU cfs_rqs in whether or not blocked_load_average contributes to the load. The basic idea behind runnable load average (the same for utilization) is that the blocked state is taken into account as opposed to only accounting for the currently runnable state. Therefore, the average should include both the runnable/running and blocked load averages. This rewrite does that. In addition, we also combine runnable/running and blocked averages of all entities into the cfs_rq's average, and update it together at once. This is based on the fact that: update(runnable) + update(blocked) = update(runnable + blocked) This significantly reduces the code as we don't need to separately maintain/update runnable/running load and blocked load. 3. How task_group entities' share is calculated is complex and imprecise. We reduce the complexity in this rewrite to allow a very simple rule: the task_group's load_avg is aggregated from its per CPU cfs_rqs's load_avgs. Then group entity's weight is simply proportional to its own cfs_rq's load_avg / task_group's load_avg. To illustrate, if a task_group has { cfs_rq1, cfs_rq2, ..., cfs_rqn }, then, task_group_avg = cfs_rq1_avg + cfs_rq2_avg + ... + cfs_rqn_avg, then cfs_rqx's entity's share = cfs_rqx_avg / task_group_avg * task_group's share To sum up, this rewrite in principle is equivalent to the current one, but fixes the issues described above. Turns out, it significantly reduces the code complexity and hence increases clarity and efficiency. In addition, the new averages are more smooth/continuous (no spurious spikes and valleys) and updated more consistently and quickly to reflect the load dynamics. As a result, we have less load tracking overhead, better performance, and especially better power efficiency due to more balanced load. Signed-off-by: Yuyang Du <yuyang.du@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: arjan@linux.intel.com Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: fengguang.wu@intel.com Cc: len.brown@intel.com Cc: morten.rasmussen@arm.com Cc: pjt@google.com Cc: rafael.j.wysocki@intel.com Cc: umgwanakikbuti@gmail.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1436918682-4971-3-git-send-email-yuyang.du@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-07-15 00:04:37 +00:00
return p->se.avg.load_avg;
}
#endif
sched/cpufreq: Move the cfs_rq_util_change() call to cpufreq_update_util() update_cfs_rq_load_avg() calls cfs_rq_util_change() every time PELT decays, which might be inefficient when the cpufreq driver has rate limitation. When a task is attached on a CPU, we have this call path: update_load_avg() update_cfs_rq_load_avg() cfs_rq_util_change -- > trig frequency update attach_entity_load_avg() cfs_rq_util_change -- > trig frequency update The 1st frequency update will not take into account the utilization of the newly attached task and the 2nd one might be discarded because of rate limitation of the cpufreq driver. update_cfs_rq_load_avg() is only called by update_blocked_averages() and update_load_avg() so we can move the call to cfs_rq_util_change/cpufreq_update_util() into these two functions. It's also interesting to note that update_load_avg() already calls cfs_rq_util_change() directly for the !SMP case. This change will also ensure that cpufreq_update_util() is called even when there is no more CFS rq in the leaf_cfs_rq_list to update, but only IRQ, RT or DL PELT signals. [ mingo: Minor updates. ] Reported-by: Doug Smythies <dsmythies@telus.net> Tested-by: Doug Smythies <dsmythies@telus.net> Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Acked-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: juri.lelli@redhat.com Cc: linux-pm@vger.kernel.org Cc: mgorman@suse.de Cc: rostedt@goodmis.org Cc: sargun@sargun.me Cc: srinivas.pandruvada@linux.intel.com Cc: tj@kernel.org Cc: xiexiuqi@huawei.com Cc: xiezhipeng1@huawei.com Fixes: 039ae8bcf7a5 ("sched/fair: Fix O(nr_cgroups) in the load balancing path") Link: https://lkml.kernel.org/r/1574083279-799-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-11-18 13:21:19 +00:00
static void update_blocked_averages(int cpu)
{
bool decayed = false, done = true;
struct rq *rq = cpu_rq(cpu);
struct rq_flags rf;
rq_lock_irqsave(rq, &rf);
update_rq_clock(rq);
decayed |= __update_blocked_others(rq, &done);
decayed |= __update_blocked_fair(rq, &done);
update_blocked_load_status(rq, !done);
if (decayed)
cpufreq_update_util(rq, 0);
rq_unlock_irqrestore(rq, &rf);
}
/********** Helpers for find_busiest_group ************************/
sched/fair: Make update_sd_pick_busiest() return 'true' on a busier sd Currently update_sd_pick_busiest only identifies the busiest sd that is either overloaded, or has a group imbalance. When no sd is imbalanced or overloaded, the load balancer fails to find the busiest domain. This breaks load balancing between domains that are not overloaded, in the !SD_ASYM_PACKING case. This patch makes update_sd_pick_busiest return true when the busiest sd yet is encountered. Groups are ranked in the order overloaded > imbalanced > other, with higher ranked groups getting priority even when their load is lower. This is necessary due to the possibility of unequal capacities and cpumasks between domains within a sched group. Behaviour for SD_ASYM_PACKING does not seem to match the comment, but I have no hardware to test that so I have left the behaviour of that code unchanged. Enum for group classification suggested by Peter Zijlstra. Signed-off-by: Rik van Riel <riel@redhat.com> [peterz: replaced sg_lb_stats::group_imb with the new enum group_type in an attempt to avoid endless recalculation] Signed-off-by: Peter Zijlstra <peterz@infradead.org> Acked-by: Vincent Guittot <vincent.guittot@linaro.org> Acked-by: Michael Neuling <mikey@neuling.org> Cc: ktkhai@parallels.com Cc: tim.c.chen@linux.intel.com Cc: nicolas.pitre@linaro.org Cc: jhladky@redhat.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/20140729152743.GI3935@laptop Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-07-28 18:16:28 +00:00
/*
* sg_lb_stats - stats of a sched_group required for load_balancing
*/
struct sg_lb_stats {
unsigned long avg_load; /*Avg load across the CPUs of the group */
unsigned long group_load; /* Total load over the CPUs of the group */
unsigned long group_capacity;
unsigned long group_util; /* Total utilization over the CPUs of the group */
unsigned long group_runnable; /* Total runnable time over the CPUs of the group */
unsigned int sum_nr_running; /* Nr of tasks running in the group */
unsigned int sum_h_nr_running; /* Nr of CFS tasks running in the group */
unsigned int idle_cpus;
unsigned int group_weight;
sched/fair: Make update_sd_pick_busiest() return 'true' on a busier sd Currently update_sd_pick_busiest only identifies the busiest sd that is either overloaded, or has a group imbalance. When no sd is imbalanced or overloaded, the load balancer fails to find the busiest domain. This breaks load balancing between domains that are not overloaded, in the !SD_ASYM_PACKING case. This patch makes update_sd_pick_busiest return true when the busiest sd yet is encountered. Groups are ranked in the order overloaded > imbalanced > other, with higher ranked groups getting priority even when their load is lower. This is necessary due to the possibility of unequal capacities and cpumasks between domains within a sched group. Behaviour for SD_ASYM_PACKING does not seem to match the comment, but I have no hardware to test that so I have left the behaviour of that code unchanged. Enum for group classification suggested by Peter Zijlstra. Signed-off-by: Rik van Riel <riel@redhat.com> [peterz: replaced sg_lb_stats::group_imb with the new enum group_type in an attempt to avoid endless recalculation] Signed-off-by: Peter Zijlstra <peterz@infradead.org> Acked-by: Vincent Guittot <vincent.guittot@linaro.org> Acked-by: Michael Neuling <mikey@neuling.org> Cc: ktkhai@parallels.com Cc: tim.c.chen@linux.intel.com Cc: nicolas.pitre@linaro.org Cc: jhladky@redhat.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/20140729152743.GI3935@laptop Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-07-28 18:16:28 +00:00
enum group_type group_type;
unsigned int group_asym_packing; /* Tasks should be moved to preferred CPU */
sched/fair: Add 'group_misfit_task' load-balance type To maximize throughput in systems with asymmetric CPU capacities (e.g. ARM big.LITTLE) load-balancing has to consider task and CPU utilization as well as per-CPU compute capacity when load-balancing in addition to the current average load based load-balancing policy. Tasks with high utilization that are scheduled on a lower capacity CPU need to be identified and migrated to a higher capacity CPU if possible to maximize throughput. To implement this additional policy an additional group_type (load-balance scenario) is added: 'group_misfit_task'. This represents scenarios where a sched_group has one or more tasks that are not suitable for its per-CPU capacity. 'group_misfit_task' is only considered if the system is not overloaded or imbalanced ('group_imbalanced' or 'group_overloaded'). Identifying misfit tasks requires the rq lock to be held. To avoid taking remote rq locks to examine source sched_groups for misfit tasks, each CPU is responsible for tracking misfit tasks themselves and update the rq->misfit_task flag. This means checking task utilization when tasks are scheduled and on sched_tick. Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: gaku.inami.xh@renesas.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1530699470-29808-3-git-send-email-morten.rasmussen@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-07-04 10:17:40 +00:00
unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */
#ifdef CONFIG_NUMA_BALANCING
unsigned int nr_numa_running;
unsigned int nr_preferred_running;
#endif
};
/*
* sd_lb_stats - Structure to store the statistics of a sched_domain
* during load balancing.
*/
struct sd_lb_stats {
struct sched_group *busiest; /* Busiest group in this sd */
struct sched_group *local; /* Local group in this sd */
unsigned long total_load; /* Total load of all groups in sd */
unsigned long total_capacity; /* Total capacity of all groups in sd */
unsigned long avg_load; /* Average load across all groups in sd */
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
unsigned int prefer_sibling; /* tasks should go to sibling first */
struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
struct sg_lb_stats local_stat; /* Statistics of the local group */
};
static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
{
/*
* Skimp on the clearing to avoid duplicate work. We can avoid clearing
* local_stat because update_sg_lb_stats() does a full clear/assignment.
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
* We must however set busiest_stat::group_type and
* busiest_stat::idle_cpus to the worst busiest group because
* update_sd_pick_busiest() reads these before assignment.
*/
*sds = (struct sd_lb_stats){
.busiest = NULL,
.local = NULL,
.total_load = 0UL,
.total_capacity = 0UL,
.busiest_stat = {
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
.idle_cpus = UINT_MAX,
.group_type = group_has_spare,
},
};
}
static unsigned long scale_rt_capacity(struct sched_domain *sd, int cpu)
{
struct rq *rq = cpu_rq(cpu);
unsigned long max = arch_scale_cpu_capacity(cpu);
unsigned long used, free;
unsigned long irq;
irq = cpu_util_irq(rq);
if (unlikely(irq >= max))
return 1;
/*
* avg_rt.util_avg and avg_dl.util_avg track binary signals
* (running and not running) with weights 0 and 1024 respectively.
* avg_thermal.load_avg tracks thermal pressure and the weighted
* average uses the actual delta max capacity(load).
*/
used = READ_ONCE(rq->avg_rt.util_avg);
used += READ_ONCE(rq->avg_dl.util_avg);
used += thermal_load_avg(rq);
if (unlikely(used >= max))
return 1;
free = max - used;
return scale_irq_capacity(free, irq, max);
}
static void update_cpu_capacity(struct sched_domain *sd, int cpu)
{
unsigned long capacity = scale_rt_capacity(sd, cpu);
struct sched_group *sdg = sd->groups;
cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(cpu);
if (!capacity)
capacity = 1;
cpu_rq(cpu)->cpu_capacity = capacity;
sdg->sgc->capacity = capacity;
sdg->sgc->min_capacity = capacity;
sdg->sgc->max_capacity = capacity;
}
void update_group_capacity(struct sched_domain *sd, int cpu)
{
struct sched_domain *child = sd->child;
struct sched_group *group, *sdg = sd->groups;
unsigned long capacity, min_capacity, max_capacity;
unsigned long interval;
interval = msecs_to_jiffies(sd->balance_interval);
interval = clamp(interval, 1UL, max_load_balance_interval);
sdg->sgc->next_update = jiffies + interval;
if (!child) {
update_cpu_capacity(sd, cpu);
return;
}
capacity = 0;
min_capacity = ULONG_MAX;
max_capacity = 0;
if (child->flags & SD_OVERLAP) {
/*
* SD_OVERLAP domains cannot assume that child groups
* span the current group.
*/
for_each_cpu(cpu, sched_group_span(sdg)) {
unsigned long cpu_cap = capacity_of(cpu);
capacity += cpu_cap;
min_capacity = min(cpu_cap, min_capacity);
max_capacity = max(cpu_cap, max_capacity);
}
} else {
/*
* !SD_OVERLAP domains can assume that child groups
* span the current group.
*/
group = child->groups;
do {
struct sched_group_capacity *sgc = group->sgc;
capacity += sgc->capacity;
min_capacity = min(sgc->min_capacity, min_capacity);
max_capacity = max(sgc->max_capacity, max_capacity);
group = group->next;
} while (group != child->groups);
}
sdg->sgc->capacity = capacity;
sdg->sgc->min_capacity = min_capacity;
sdg->sgc->max_capacity = max_capacity;
}
/*
sched: Replace capacity_factor by usage The scheduler tries to compute how many tasks a group of CPUs can handle by assuming that a task's load is SCHED_LOAD_SCALE and a CPU's capacity is SCHED_CAPACITY_SCALE. 'struct sg_lb_stats:group_capacity_factor' divides the capacity of the group by SCHED_LOAD_SCALE to estimate how many task can run in the group. Then, it compares this value with the sum of nr_running to decide if the group is overloaded or not. But the 'group_capacity_factor' concept is hardly working for SMT systems, it sometimes works for big cores but fails to do the right thing for little cores. Below are two examples to illustrate the problem that this patch solves: 1- If the original capacity of a CPU is less than SCHED_CAPACITY_SCALE (640 as an example), a group of 3 CPUS will have a max capacity_factor of 2 (div_round_closest(3x640/1024) = 2) which means that it will be seen as overloaded even if we have only one task per CPU. 2 - If the original capacity of a CPU is greater than SCHED_CAPACITY_SCALE (1512 as an example), a group of 4 CPUs will have a capacity_factor of 4 (at max and thanks to the fix [0] for SMT system that prevent the apparition of ghost CPUs) but if one CPU is fully used by rt tasks (and its capacity is reduced to nearly nothing), the capacity factor of the group will still be 4 (div_round_closest(3*1512/1024) = 5 which is cap to 4 with [0]). So, this patch tries to solve this issue by removing capacity_factor and replacing it with the 2 following metrics: - The available CPU's capacity for CFS tasks which is already used by load_balance(). - The usage of the CPU by the CFS tasks. For the latter, utilization_avg_contrib has been re-introduced to compute the usage of a CPU by CFS tasks. 'group_capacity_factor' and 'group_has_free_capacity' has been removed and replaced by 'group_no_capacity'. We compare the number of task with the number of CPUs and we evaluate the level of utilization of the CPUs to define if a group is overloaded or if a group has capacity to handle more tasks. For SD_PREFER_SIBLING, a group is tagged overloaded if it has more than 1 task so it will be selected in priority (among the overloaded groups). Since [1], SD_PREFER_SIBLING is no more concerned by the computation of 'load_above_capacity' because local is not overloaded. [1] 9a5d9ba6a363 ("sched/fair: Allow calculate_imbalance() to move idle cpus") Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Morten.Rasmussen@arm.com Cc: dietmar.eggemann@arm.com Cc: efault@gmx.de Cc: kamalesh@linux.vnet.ibm.com Cc: linaro-kernel@lists.linaro.org Cc: nicolas.pitre@linaro.org Cc: preeti@linux.vnet.ibm.com Cc: riel@redhat.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/1425052454-25797-9-git-send-email-vincent.guittot@linaro.org [ Tidied up the changelog. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-02-27 15:54:11 +00:00
* Check whether the capacity of the rq has been noticeably reduced by side
* activity. The imbalance_pct is used for the threshold.
* Return true is the capacity is reduced
*/
static inline int
sched: Replace capacity_factor by usage The scheduler tries to compute how many tasks a group of CPUs can handle by assuming that a task's load is SCHED_LOAD_SCALE and a CPU's capacity is SCHED_CAPACITY_SCALE. 'struct sg_lb_stats:group_capacity_factor' divides the capacity of the group by SCHED_LOAD_SCALE to estimate how many task can run in the group. Then, it compares this value with the sum of nr_running to decide if the group is overloaded or not. But the 'group_capacity_factor' concept is hardly working for SMT systems, it sometimes works for big cores but fails to do the right thing for little cores. Below are two examples to illustrate the problem that this patch solves: 1- If the original capacity of a CPU is less than SCHED_CAPACITY_SCALE (640 as an example), a group of 3 CPUS will have a max capacity_factor of 2 (div_round_closest(3x640/1024) = 2) which means that it will be seen as overloaded even if we have only one task per CPU. 2 - If the original capacity of a CPU is greater than SCHED_CAPACITY_SCALE (1512 as an example), a group of 4 CPUs will have a capacity_factor of 4 (at max and thanks to the fix [0] for SMT system that prevent the apparition of ghost CPUs) but if one CPU is fully used by rt tasks (and its capacity is reduced to nearly nothing), the capacity factor of the group will still be 4 (div_round_closest(3*1512/1024) = 5 which is cap to 4 with [0]). So, this patch tries to solve this issue by removing capacity_factor and replacing it with the 2 following metrics: - The available CPU's capacity for CFS tasks which is already used by load_balance(). - The usage of the CPU by the CFS tasks. For the latter, utilization_avg_contrib has been re-introduced to compute the usage of a CPU by CFS tasks. 'group_capacity_factor' and 'group_has_free_capacity' has been removed and replaced by 'group_no_capacity'. We compare the number of task with the number of CPUs and we evaluate the level of utilization of the CPUs to define if a group is overloaded or if a group has capacity to handle more tasks. For SD_PREFER_SIBLING, a group is tagged overloaded if it has more than 1 task so it will be selected in priority (among the overloaded groups). Since [1], SD_PREFER_SIBLING is no more concerned by the computation of 'load_above_capacity' because local is not overloaded. [1] 9a5d9ba6a363 ("sched/fair: Allow calculate_imbalance() to move idle cpus") Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Morten.Rasmussen@arm.com Cc: dietmar.eggemann@arm.com Cc: efault@gmx.de Cc: kamalesh@linux.vnet.ibm.com Cc: linaro-kernel@lists.linaro.org Cc: nicolas.pitre@linaro.org Cc: preeti@linux.vnet.ibm.com Cc: riel@redhat.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/1425052454-25797-9-git-send-email-vincent.guittot@linaro.org [ Tidied up the changelog. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-02-27 15:54:11 +00:00
check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
{
sched: Replace capacity_factor by usage The scheduler tries to compute how many tasks a group of CPUs can handle by assuming that a task's load is SCHED_LOAD_SCALE and a CPU's capacity is SCHED_CAPACITY_SCALE. 'struct sg_lb_stats:group_capacity_factor' divides the capacity of the group by SCHED_LOAD_SCALE to estimate how many task can run in the group. Then, it compares this value with the sum of nr_running to decide if the group is overloaded or not. But the 'group_capacity_factor' concept is hardly working for SMT systems, it sometimes works for big cores but fails to do the right thing for little cores. Below are two examples to illustrate the problem that this patch solves: 1- If the original capacity of a CPU is less than SCHED_CAPACITY_SCALE (640 as an example), a group of 3 CPUS will have a max capacity_factor of 2 (div_round_closest(3x640/1024) = 2) which means that it will be seen as overloaded even if we have only one task per CPU. 2 - If the original capacity of a CPU is greater than SCHED_CAPACITY_SCALE (1512 as an example), a group of 4 CPUs will have a capacity_factor of 4 (at max and thanks to the fix [0] for SMT system that prevent the apparition of ghost CPUs) but if one CPU is fully used by rt tasks (and its capacity is reduced to nearly nothing), the capacity factor of the group will still be 4 (div_round_closest(3*1512/1024) = 5 which is cap to 4 with [0]). So, this patch tries to solve this issue by removing capacity_factor and replacing it with the 2 following metrics: - The available CPU's capacity for CFS tasks which is already used by load_balance(). - The usage of the CPU by the CFS tasks. For the latter, utilization_avg_contrib has been re-introduced to compute the usage of a CPU by CFS tasks. 'group_capacity_factor' and 'group_has_free_capacity' has been removed and replaced by 'group_no_capacity'. We compare the number of task with the number of CPUs and we evaluate the level of utilization of the CPUs to define if a group is overloaded or if a group has capacity to handle more tasks. For SD_PREFER_SIBLING, a group is tagged overloaded if it has more than 1 task so it will be selected in priority (among the overloaded groups). Since [1], SD_PREFER_SIBLING is no more concerned by the computation of 'load_above_capacity' because local is not overloaded. [1] 9a5d9ba6a363 ("sched/fair: Allow calculate_imbalance() to move idle cpus") Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Morten.Rasmussen@arm.com Cc: dietmar.eggemann@arm.com Cc: efault@gmx.de Cc: kamalesh@linux.vnet.ibm.com Cc: linaro-kernel@lists.linaro.org Cc: nicolas.pitre@linaro.org Cc: preeti@linux.vnet.ibm.com Cc: riel@redhat.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/1425052454-25797-9-git-send-email-vincent.guittot@linaro.org [ Tidied up the changelog. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-02-27 15:54:11 +00:00
return ((rq->cpu_capacity * sd->imbalance_pct) <
(rq->cpu_capacity_orig * 100));
}
sched/fair: Tune down misfit NOHZ kicks In this commit: 3b1baa6496e6 ("sched/fair: Add 'group_misfit_task' load-balance type") we set rq->misfit_task_load whenever the current running task has a utilization greater than 80% of rq->cpu_capacity. A non-zero value in this field enables misfit load balancing. However, if the task being looked at is already running on a CPU of highest capacity, there's nothing more we can do for it. We can currently spot this in update_sd_pick_busiest(), which prevents us from selecting a sched_group of group_type == group_misfit_task as the busiest group, but we don't do any of that in nohz_balancer_kick(). This means that we could repeatedly kick NOHZ CPUs when there's no improvements in terms of load balance to be done. Introduce a check_misfit_status() helper that returns true iff there is a CPU in the system that could give more CPU capacity to a rq's misfit task - IOW, there exists a CPU of higher capacity_orig or the rq's CPU is severely pressured by rt/IRQ. Signed-off-by: Valentin Schneider <valentin.schneider@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andy Lutomirski <luto@kernel.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Dietmar.Eggemann@arm.com Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: morten.rasmussen@arm.com Cc: vincent.guittot@linaro.org Link: https://lkml.kernel.org/r/20190211175946.4961-3-valentin.schneider@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-02-11 17:59:45 +00:00
/*
* Check whether a rq has a misfit task and if it looks like we can actually
* help that task: we can migrate the task to a CPU of higher capacity, or
* the task's current CPU is heavily pressured.
*/
static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd)
{
return rq->misfit_task_load &&
(rq->cpu_capacity_orig < rq->rd->max_cpu_capacity ||
check_cpu_capacity(rq, sd));
}
/*
* Group imbalance indicates (and tries to solve) the problem where balancing
* groups is inadequate due to ->cpus_ptr constraints.
*
* Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
* cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
* Something like:
*
* { 0 1 2 3 } { 4 5 6 7 }
* * * * *
*
* If we were to balance group-wise we'd place two tasks in the first group and
* two tasks in the second group. Clearly this is undesired as it will overload
* cpu 3 and leave one of the CPUs in the second group unused.
*
* The current solution to this issue is detecting the skew in the first group
* by noticing the lower domain failed to reach balance and had difficulty
* moving tasks due to affinity constraints.
*
* When this is so detected; this group becomes a candidate for busiest; see
* update_sd_pick_busiest(). And calculate_imbalance() and
* find_busiest_group() avoid some of the usual balance conditions to allow it
* to create an effective group imbalance.
*
* This is a somewhat tricky proposition since the next run might not find the
* group imbalance and decide the groups need to be balanced again. A most
* subtle and fragile situation.
*/
static inline int sg_imbalanced(struct sched_group *group)
{
return group->sgc->imbalance;
}
/*
sched: Replace capacity_factor by usage The scheduler tries to compute how many tasks a group of CPUs can handle by assuming that a task's load is SCHED_LOAD_SCALE and a CPU's capacity is SCHED_CAPACITY_SCALE. 'struct sg_lb_stats:group_capacity_factor' divides the capacity of the group by SCHED_LOAD_SCALE to estimate how many task can run in the group. Then, it compares this value with the sum of nr_running to decide if the group is overloaded or not. But the 'group_capacity_factor' concept is hardly working for SMT systems, it sometimes works for big cores but fails to do the right thing for little cores. Below are two examples to illustrate the problem that this patch solves: 1- If the original capacity of a CPU is less than SCHED_CAPACITY_SCALE (640 as an example), a group of 3 CPUS will have a max capacity_factor of 2 (div_round_closest(3x640/1024) = 2) which means that it will be seen as overloaded even if we have only one task per CPU. 2 - If the original capacity of a CPU is greater than SCHED_CAPACITY_SCALE (1512 as an example), a group of 4 CPUs will have a capacity_factor of 4 (at max and thanks to the fix [0] for SMT system that prevent the apparition of ghost CPUs) but if one CPU is fully used by rt tasks (and its capacity is reduced to nearly nothing), the capacity factor of the group will still be 4 (div_round_closest(3*1512/1024) = 5 which is cap to 4 with [0]). So, this patch tries to solve this issue by removing capacity_factor and replacing it with the 2 following metrics: - The available CPU's capacity for CFS tasks which is already used by load_balance(). - The usage of the CPU by the CFS tasks. For the latter, utilization_avg_contrib has been re-introduced to compute the usage of a CPU by CFS tasks. 'group_capacity_factor' and 'group_has_free_capacity' has been removed and replaced by 'group_no_capacity'. We compare the number of task with the number of CPUs and we evaluate the level of utilization of the CPUs to define if a group is overloaded or if a group has capacity to handle more tasks. For SD_PREFER_SIBLING, a group is tagged overloaded if it has more than 1 task so it will be selected in priority (among the overloaded groups). Since [1], SD_PREFER_SIBLING is no more concerned by the computation of 'load_above_capacity' because local is not overloaded. [1] 9a5d9ba6a363 ("sched/fair: Allow calculate_imbalance() to move idle cpus") Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Morten.Rasmussen@arm.com Cc: dietmar.eggemann@arm.com Cc: efault@gmx.de Cc: kamalesh@linux.vnet.ibm.com Cc: linaro-kernel@lists.linaro.org Cc: nicolas.pitre@linaro.org Cc: preeti@linux.vnet.ibm.com Cc: riel@redhat.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/1425052454-25797-9-git-send-email-vincent.guittot@linaro.org [ Tidied up the changelog. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-02-27 15:54:11 +00:00
* group_has_capacity returns true if the group has spare capacity that could
* be used by some tasks.
* We consider that a group has spare capacity if the * number of task is
* smaller than the number of CPUs or if the utilization is lower than the
* available capacity for CFS tasks.
sched: Replace capacity_factor by usage The scheduler tries to compute how many tasks a group of CPUs can handle by assuming that a task's load is SCHED_LOAD_SCALE and a CPU's capacity is SCHED_CAPACITY_SCALE. 'struct sg_lb_stats:group_capacity_factor' divides the capacity of the group by SCHED_LOAD_SCALE to estimate how many task can run in the group. Then, it compares this value with the sum of nr_running to decide if the group is overloaded or not. But the 'group_capacity_factor' concept is hardly working for SMT systems, it sometimes works for big cores but fails to do the right thing for little cores. Below are two examples to illustrate the problem that this patch solves: 1- If the original capacity of a CPU is less than SCHED_CAPACITY_SCALE (640 as an example), a group of 3 CPUS will have a max capacity_factor of 2 (div_round_closest(3x640/1024) = 2) which means that it will be seen as overloaded even if we have only one task per CPU. 2 - If the original capacity of a CPU is greater than SCHED_CAPACITY_SCALE (1512 as an example), a group of 4 CPUs will have a capacity_factor of 4 (at max and thanks to the fix [0] for SMT system that prevent the apparition of ghost CPUs) but if one CPU is fully used by rt tasks (and its capacity is reduced to nearly nothing), the capacity factor of the group will still be 4 (div_round_closest(3*1512/1024) = 5 which is cap to 4 with [0]). So, this patch tries to solve this issue by removing capacity_factor and replacing it with the 2 following metrics: - The available CPU's capacity for CFS tasks which is already used by load_balance(). - The usage of the CPU by the CFS tasks. For the latter, utilization_avg_contrib has been re-introduced to compute the usage of a CPU by CFS tasks. 'group_capacity_factor' and 'group_has_free_capacity' has been removed and replaced by 'group_no_capacity'. We compare the number of task with the number of CPUs and we evaluate the level of utilization of the CPUs to define if a group is overloaded or if a group has capacity to handle more tasks. For SD_PREFER_SIBLING, a group is tagged overloaded if it has more than 1 task so it will be selected in priority (among the overloaded groups). Since [1], SD_PREFER_SIBLING is no more concerned by the computation of 'load_above_capacity' because local is not overloaded. [1] 9a5d9ba6a363 ("sched/fair: Allow calculate_imbalance() to move idle cpus") Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Morten.Rasmussen@arm.com Cc: dietmar.eggemann@arm.com Cc: efault@gmx.de Cc: kamalesh@linux.vnet.ibm.com Cc: linaro-kernel@lists.linaro.org Cc: nicolas.pitre@linaro.org Cc: preeti@linux.vnet.ibm.com Cc: riel@redhat.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/1425052454-25797-9-git-send-email-vincent.guittot@linaro.org [ Tidied up the changelog. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-02-27 15:54:11 +00:00
* For the latter, we use a threshold to stabilize the state, to take into
* account the variance of the tasks' load and to return true if the available
* capacity in meaningful for the load balancer.
* As an example, an available capacity of 1% can appear but it doesn't make
* any benefit for the load balance.
*/
sched: Replace capacity_factor by usage The scheduler tries to compute how many tasks a group of CPUs can handle by assuming that a task's load is SCHED_LOAD_SCALE and a CPU's capacity is SCHED_CAPACITY_SCALE. 'struct sg_lb_stats:group_capacity_factor' divides the capacity of the group by SCHED_LOAD_SCALE to estimate how many task can run in the group. Then, it compares this value with the sum of nr_running to decide if the group is overloaded or not. But the 'group_capacity_factor' concept is hardly working for SMT systems, it sometimes works for big cores but fails to do the right thing for little cores. Below are two examples to illustrate the problem that this patch solves: 1- If the original capacity of a CPU is less than SCHED_CAPACITY_SCALE (640 as an example), a group of 3 CPUS will have a max capacity_factor of 2 (div_round_closest(3x640/1024) = 2) which means that it will be seen as overloaded even if we have only one task per CPU. 2 - If the original capacity of a CPU is greater than SCHED_CAPACITY_SCALE (1512 as an example), a group of 4 CPUs will have a capacity_factor of 4 (at max and thanks to the fix [0] for SMT system that prevent the apparition of ghost CPUs) but if one CPU is fully used by rt tasks (and its capacity is reduced to nearly nothing), the capacity factor of the group will still be 4 (div_round_closest(3*1512/1024) = 5 which is cap to 4 with [0]). So, this patch tries to solve this issue by removing capacity_factor and replacing it with the 2 following metrics: - The available CPU's capacity for CFS tasks which is already used by load_balance(). - The usage of the CPU by the CFS tasks. For the latter, utilization_avg_contrib has been re-introduced to compute the usage of a CPU by CFS tasks. 'group_capacity_factor' and 'group_has_free_capacity' has been removed and replaced by 'group_no_capacity'. We compare the number of task with the number of CPUs and we evaluate the level of utilization of the CPUs to define if a group is overloaded or if a group has capacity to handle more tasks. For SD_PREFER_SIBLING, a group is tagged overloaded if it has more than 1 task so it will be selected in priority (among the overloaded groups). Since [1], SD_PREFER_SIBLING is no more concerned by the computation of 'load_above_capacity' because local is not overloaded. [1] 9a5d9ba6a363 ("sched/fair: Allow calculate_imbalance() to move idle cpus") Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Morten.Rasmussen@arm.com Cc: dietmar.eggemann@arm.com Cc: efault@gmx.de Cc: kamalesh@linux.vnet.ibm.com Cc: linaro-kernel@lists.linaro.org Cc: nicolas.pitre@linaro.org Cc: preeti@linux.vnet.ibm.com Cc: riel@redhat.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/1425052454-25797-9-git-send-email-vincent.guittot@linaro.org [ Tidied up the changelog. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-02-27 15:54:11 +00:00
static inline bool
group_has_capacity(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
{
if (sgs->sum_nr_running < sgs->group_weight)
sched: Replace capacity_factor by usage The scheduler tries to compute how many tasks a group of CPUs can handle by assuming that a task's load is SCHED_LOAD_SCALE and a CPU's capacity is SCHED_CAPACITY_SCALE. 'struct sg_lb_stats:group_capacity_factor' divides the capacity of the group by SCHED_LOAD_SCALE to estimate how many task can run in the group. Then, it compares this value with the sum of nr_running to decide if the group is overloaded or not. But the 'group_capacity_factor' concept is hardly working for SMT systems, it sometimes works for big cores but fails to do the right thing for little cores. Below are two examples to illustrate the problem that this patch solves: 1- If the original capacity of a CPU is less than SCHED_CAPACITY_SCALE (640 as an example), a group of 3 CPUS will have a max capacity_factor of 2 (div_round_closest(3x640/1024) = 2) which means that it will be seen as overloaded even if we have only one task per CPU. 2 - If the original capacity of a CPU is greater than SCHED_CAPACITY_SCALE (1512 as an example), a group of 4 CPUs will have a capacity_factor of 4 (at max and thanks to the fix [0] for SMT system that prevent the apparition of ghost CPUs) but if one CPU is fully used by rt tasks (and its capacity is reduced to nearly nothing), the capacity factor of the group will still be 4 (div_round_closest(3*1512/1024) = 5 which is cap to 4 with [0]). So, this patch tries to solve this issue by removing capacity_factor and replacing it with the 2 following metrics: - The available CPU's capacity for CFS tasks which is already used by load_balance(). - The usage of the CPU by the CFS tasks. For the latter, utilization_avg_contrib has been re-introduced to compute the usage of a CPU by CFS tasks. 'group_capacity_factor' and 'group_has_free_capacity' has been removed and replaced by 'group_no_capacity'. We compare the number of task with the number of CPUs and we evaluate the level of utilization of the CPUs to define if a group is overloaded or if a group has capacity to handle more tasks. For SD_PREFER_SIBLING, a group is tagged overloaded if it has more than 1 task so it will be selected in priority (among the overloaded groups). Since [1], SD_PREFER_SIBLING is no more concerned by the computation of 'load_above_capacity' because local is not overloaded. [1] 9a5d9ba6a363 ("sched/fair: Allow calculate_imbalance() to move idle cpus") Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Morten.Rasmussen@arm.com Cc: dietmar.eggemann@arm.com Cc: efault@gmx.de Cc: kamalesh@linux.vnet.ibm.com Cc: linaro-kernel@lists.linaro.org Cc: nicolas.pitre@linaro.org Cc: preeti@linux.vnet.ibm.com Cc: riel@redhat.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/1425052454-25797-9-git-send-email-vincent.guittot@linaro.org [ Tidied up the changelog. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-02-27 15:54:11 +00:00
return true;
if ((sgs->group_capacity * imbalance_pct) <
(sgs->group_runnable * 100))
return false;
sched: Replace capacity_factor by usage The scheduler tries to compute how many tasks a group of CPUs can handle by assuming that a task's load is SCHED_LOAD_SCALE and a CPU's capacity is SCHED_CAPACITY_SCALE. 'struct sg_lb_stats:group_capacity_factor' divides the capacity of the group by SCHED_LOAD_SCALE to estimate how many task can run in the group. Then, it compares this value with the sum of nr_running to decide if the group is overloaded or not. But the 'group_capacity_factor' concept is hardly working for SMT systems, it sometimes works for big cores but fails to do the right thing for little cores. Below are two examples to illustrate the problem that this patch solves: 1- If the original capacity of a CPU is less than SCHED_CAPACITY_SCALE (640 as an example), a group of 3 CPUS will have a max capacity_factor of 2 (div_round_closest(3x640/1024) = 2) which means that it will be seen as overloaded even if we have only one task per CPU. 2 - If the original capacity of a CPU is greater than SCHED_CAPACITY_SCALE (1512 as an example), a group of 4 CPUs will have a capacity_factor of 4 (at max and thanks to the fix [0] for SMT system that prevent the apparition of ghost CPUs) but if one CPU is fully used by rt tasks (and its capacity is reduced to nearly nothing), the capacity factor of the group will still be 4 (div_round_closest(3*1512/1024) = 5 which is cap to 4 with [0]). So, this patch tries to solve this issue by removing capacity_factor and replacing it with the 2 following metrics: - The available CPU's capacity for CFS tasks which is already used by load_balance(). - The usage of the CPU by the CFS tasks. For the latter, utilization_avg_contrib has been re-introduced to compute the usage of a CPU by CFS tasks. 'group_capacity_factor' and 'group_has_free_capacity' has been removed and replaced by 'group_no_capacity'. We compare the number of task with the number of CPUs and we evaluate the level of utilization of the CPUs to define if a group is overloaded or if a group has capacity to handle more tasks. For SD_PREFER_SIBLING, a group is tagged overloaded if it has more than 1 task so it will be selected in priority (among the overloaded groups). Since [1], SD_PREFER_SIBLING is no more concerned by the computation of 'load_above_capacity' because local is not overloaded. [1] 9a5d9ba6a363 ("sched/fair: Allow calculate_imbalance() to move idle cpus") Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Morten.Rasmussen@arm.com Cc: dietmar.eggemann@arm.com Cc: efault@gmx.de Cc: kamalesh@linux.vnet.ibm.com Cc: linaro-kernel@lists.linaro.org Cc: nicolas.pitre@linaro.org Cc: preeti@linux.vnet.ibm.com Cc: riel@redhat.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/1425052454-25797-9-git-send-email-vincent.guittot@linaro.org [ Tidied up the changelog. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-02-27 15:54:11 +00:00
if ((sgs->group_capacity * 100) >
(sgs->group_util * imbalance_pct))
sched: Replace capacity_factor by usage The scheduler tries to compute how many tasks a group of CPUs can handle by assuming that a task's load is SCHED_LOAD_SCALE and a CPU's capacity is SCHED_CAPACITY_SCALE. 'struct sg_lb_stats:group_capacity_factor' divides the capacity of the group by SCHED_LOAD_SCALE to estimate how many task can run in the group. Then, it compares this value with the sum of nr_running to decide if the group is overloaded or not. But the 'group_capacity_factor' concept is hardly working for SMT systems, it sometimes works for big cores but fails to do the right thing for little cores. Below are two examples to illustrate the problem that this patch solves: 1- If the original capacity of a CPU is less than SCHED_CAPACITY_SCALE (640 as an example), a group of 3 CPUS will have a max capacity_factor of 2 (div_round_closest(3x640/1024) = 2) which means that it will be seen as overloaded even if we have only one task per CPU. 2 - If the original capacity of a CPU is greater than SCHED_CAPACITY_SCALE (1512 as an example), a group of 4 CPUs will have a capacity_factor of 4 (at max and thanks to the fix [0] for SMT system that prevent the apparition of ghost CPUs) but if one CPU is fully used by rt tasks (and its capacity is reduced to nearly nothing), the capacity factor of the group will still be 4 (div_round_closest(3*1512/1024) = 5 which is cap to 4 with [0]). So, this patch tries to solve this issue by removing capacity_factor and replacing it with the 2 following metrics: - The available CPU's capacity for CFS tasks which is already used by load_balance(). - The usage of the CPU by the CFS tasks. For the latter, utilization_avg_contrib has been re-introduced to compute the usage of a CPU by CFS tasks. 'group_capacity_factor' and 'group_has_free_capacity' has been removed and replaced by 'group_no_capacity'. We compare the number of task with the number of CPUs and we evaluate the level of utilization of the CPUs to define if a group is overloaded or if a group has capacity to handle more tasks. For SD_PREFER_SIBLING, a group is tagged overloaded if it has more than 1 task so it will be selected in priority (among the overloaded groups). Since [1], SD_PREFER_SIBLING is no more concerned by the computation of 'load_above_capacity' because local is not overloaded. [1] 9a5d9ba6a363 ("sched/fair: Allow calculate_imbalance() to move idle cpus") Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Morten.Rasmussen@arm.com Cc: dietmar.eggemann@arm.com Cc: efault@gmx.de Cc: kamalesh@linux.vnet.ibm.com Cc: linaro-kernel@lists.linaro.org Cc: nicolas.pitre@linaro.org Cc: preeti@linux.vnet.ibm.com Cc: riel@redhat.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/1425052454-25797-9-git-send-email-vincent.guittot@linaro.org [ Tidied up the changelog. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-02-27 15:54:11 +00:00
return true;
sched: Replace capacity_factor by usage The scheduler tries to compute how many tasks a group of CPUs can handle by assuming that a task's load is SCHED_LOAD_SCALE and a CPU's capacity is SCHED_CAPACITY_SCALE. 'struct sg_lb_stats:group_capacity_factor' divides the capacity of the group by SCHED_LOAD_SCALE to estimate how many task can run in the group. Then, it compares this value with the sum of nr_running to decide if the group is overloaded or not. But the 'group_capacity_factor' concept is hardly working for SMT systems, it sometimes works for big cores but fails to do the right thing for little cores. Below are two examples to illustrate the problem that this patch solves: 1- If the original capacity of a CPU is less than SCHED_CAPACITY_SCALE (640 as an example), a group of 3 CPUS will have a max capacity_factor of 2 (div_round_closest(3x640/1024) = 2) which means that it will be seen as overloaded even if we have only one task per CPU. 2 - If the original capacity of a CPU is greater than SCHED_CAPACITY_SCALE (1512 as an example), a group of 4 CPUs will have a capacity_factor of 4 (at max and thanks to the fix [0] for SMT system that prevent the apparition of ghost CPUs) but if one CPU is fully used by rt tasks (and its capacity is reduced to nearly nothing), the capacity factor of the group will still be 4 (div_round_closest(3*1512/1024) = 5 which is cap to 4 with [0]). So, this patch tries to solve this issue by removing capacity_factor and replacing it with the 2 following metrics: - The available CPU's capacity for CFS tasks which is already used by load_balance(). - The usage of the CPU by the CFS tasks. For the latter, utilization_avg_contrib has been re-introduced to compute the usage of a CPU by CFS tasks. 'group_capacity_factor' and 'group_has_free_capacity' has been removed and replaced by 'group_no_capacity'. We compare the number of task with the number of CPUs and we evaluate the level of utilization of the CPUs to define if a group is overloaded or if a group has capacity to handle more tasks. For SD_PREFER_SIBLING, a group is tagged overloaded if it has more than 1 task so it will be selected in priority (among the overloaded groups). Since [1], SD_PREFER_SIBLING is no more concerned by the computation of 'load_above_capacity' because local is not overloaded. [1] 9a5d9ba6a363 ("sched/fair: Allow calculate_imbalance() to move idle cpus") Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Morten.Rasmussen@arm.com Cc: dietmar.eggemann@arm.com Cc: efault@gmx.de Cc: kamalesh@linux.vnet.ibm.com Cc: linaro-kernel@lists.linaro.org Cc: nicolas.pitre@linaro.org Cc: preeti@linux.vnet.ibm.com Cc: riel@redhat.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/1425052454-25797-9-git-send-email-vincent.guittot@linaro.org [ Tidied up the changelog. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-02-27 15:54:11 +00:00
return false;
}
/*
* group_is_overloaded returns true if the group has more tasks than it can
* handle.
* group_is_overloaded is not equals to !group_has_capacity because a group
* with the exact right number of tasks, has no more spare capacity but is not
* overloaded so both group_has_capacity and group_is_overloaded return
* false.
*/
static inline bool
group_is_overloaded(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
sched: Replace capacity_factor by usage The scheduler tries to compute how many tasks a group of CPUs can handle by assuming that a task's load is SCHED_LOAD_SCALE and a CPU's capacity is SCHED_CAPACITY_SCALE. 'struct sg_lb_stats:group_capacity_factor' divides the capacity of the group by SCHED_LOAD_SCALE to estimate how many task can run in the group. Then, it compares this value with the sum of nr_running to decide if the group is overloaded or not. But the 'group_capacity_factor' concept is hardly working for SMT systems, it sometimes works for big cores but fails to do the right thing for little cores. Below are two examples to illustrate the problem that this patch solves: 1- If the original capacity of a CPU is less than SCHED_CAPACITY_SCALE (640 as an example), a group of 3 CPUS will have a max capacity_factor of 2 (div_round_closest(3x640/1024) = 2) which means that it will be seen as overloaded even if we have only one task per CPU. 2 - If the original capacity of a CPU is greater than SCHED_CAPACITY_SCALE (1512 as an example), a group of 4 CPUs will have a capacity_factor of 4 (at max and thanks to the fix [0] for SMT system that prevent the apparition of ghost CPUs) but if one CPU is fully used by rt tasks (and its capacity is reduced to nearly nothing), the capacity factor of the group will still be 4 (div_round_closest(3*1512/1024) = 5 which is cap to 4 with [0]). So, this patch tries to solve this issue by removing capacity_factor and replacing it with the 2 following metrics: - The available CPU's capacity for CFS tasks which is already used by load_balance(). - The usage of the CPU by the CFS tasks. For the latter, utilization_avg_contrib has been re-introduced to compute the usage of a CPU by CFS tasks. 'group_capacity_factor' and 'group_has_free_capacity' has been removed and replaced by 'group_no_capacity'. We compare the number of task with the number of CPUs and we evaluate the level of utilization of the CPUs to define if a group is overloaded or if a group has capacity to handle more tasks. For SD_PREFER_SIBLING, a group is tagged overloaded if it has more than 1 task so it will be selected in priority (among the overloaded groups). Since [1], SD_PREFER_SIBLING is no more concerned by the computation of 'load_above_capacity' because local is not overloaded. [1] 9a5d9ba6a363 ("sched/fair: Allow calculate_imbalance() to move idle cpus") Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Morten.Rasmussen@arm.com Cc: dietmar.eggemann@arm.com Cc: efault@gmx.de Cc: kamalesh@linux.vnet.ibm.com Cc: linaro-kernel@lists.linaro.org Cc: nicolas.pitre@linaro.org Cc: preeti@linux.vnet.ibm.com Cc: riel@redhat.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/1425052454-25797-9-git-send-email-vincent.guittot@linaro.org [ Tidied up the changelog. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-02-27 15:54:11 +00:00
{
if (sgs->sum_nr_running <= sgs->group_weight)
sched: Replace capacity_factor by usage The scheduler tries to compute how many tasks a group of CPUs can handle by assuming that a task's load is SCHED_LOAD_SCALE and a CPU's capacity is SCHED_CAPACITY_SCALE. 'struct sg_lb_stats:group_capacity_factor' divides the capacity of the group by SCHED_LOAD_SCALE to estimate how many task can run in the group. Then, it compares this value with the sum of nr_running to decide if the group is overloaded or not. But the 'group_capacity_factor' concept is hardly working for SMT systems, it sometimes works for big cores but fails to do the right thing for little cores. Below are two examples to illustrate the problem that this patch solves: 1- If the original capacity of a CPU is less than SCHED_CAPACITY_SCALE (640 as an example), a group of 3 CPUS will have a max capacity_factor of 2 (div_round_closest(3x640/1024) = 2) which means that it will be seen as overloaded even if we have only one task per CPU. 2 - If the original capacity of a CPU is greater than SCHED_CAPACITY_SCALE (1512 as an example), a group of 4 CPUs will have a capacity_factor of 4 (at max and thanks to the fix [0] for SMT system that prevent the apparition of ghost CPUs) but if one CPU is fully used by rt tasks (and its capacity is reduced to nearly nothing), the capacity factor of the group will still be 4 (div_round_closest(3*1512/1024) = 5 which is cap to 4 with [0]). So, this patch tries to solve this issue by removing capacity_factor and replacing it with the 2 following metrics: - The available CPU's capacity for CFS tasks which is already used by load_balance(). - The usage of the CPU by the CFS tasks. For the latter, utilization_avg_contrib has been re-introduced to compute the usage of a CPU by CFS tasks. 'group_capacity_factor' and 'group_has_free_capacity' has been removed and replaced by 'group_no_capacity'. We compare the number of task with the number of CPUs and we evaluate the level of utilization of the CPUs to define if a group is overloaded or if a group has capacity to handle more tasks. For SD_PREFER_SIBLING, a group is tagged overloaded if it has more than 1 task so it will be selected in priority (among the overloaded groups). Since [1], SD_PREFER_SIBLING is no more concerned by the computation of 'load_above_capacity' because local is not overloaded. [1] 9a5d9ba6a363 ("sched/fair: Allow calculate_imbalance() to move idle cpus") Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Morten.Rasmussen@arm.com Cc: dietmar.eggemann@arm.com Cc: efault@gmx.de Cc: kamalesh@linux.vnet.ibm.com Cc: linaro-kernel@lists.linaro.org Cc: nicolas.pitre@linaro.org Cc: preeti@linux.vnet.ibm.com Cc: riel@redhat.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/1425052454-25797-9-git-send-email-vincent.guittot@linaro.org [ Tidied up the changelog. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-02-27 15:54:11 +00:00
return false;
sched: Replace capacity_factor by usage The scheduler tries to compute how many tasks a group of CPUs can handle by assuming that a task's load is SCHED_LOAD_SCALE and a CPU's capacity is SCHED_CAPACITY_SCALE. 'struct sg_lb_stats:group_capacity_factor' divides the capacity of the group by SCHED_LOAD_SCALE to estimate how many task can run in the group. Then, it compares this value with the sum of nr_running to decide if the group is overloaded or not. But the 'group_capacity_factor' concept is hardly working for SMT systems, it sometimes works for big cores but fails to do the right thing for little cores. Below are two examples to illustrate the problem that this patch solves: 1- If the original capacity of a CPU is less than SCHED_CAPACITY_SCALE (640 as an example), a group of 3 CPUS will have a max capacity_factor of 2 (div_round_closest(3x640/1024) = 2) which means that it will be seen as overloaded even if we have only one task per CPU. 2 - If the original capacity of a CPU is greater than SCHED_CAPACITY_SCALE (1512 as an example), a group of 4 CPUs will have a capacity_factor of 4 (at max and thanks to the fix [0] for SMT system that prevent the apparition of ghost CPUs) but if one CPU is fully used by rt tasks (and its capacity is reduced to nearly nothing), the capacity factor of the group will still be 4 (div_round_closest(3*1512/1024) = 5 which is cap to 4 with [0]). So, this patch tries to solve this issue by removing capacity_factor and replacing it with the 2 following metrics: - The available CPU's capacity for CFS tasks which is already used by load_balance(). - The usage of the CPU by the CFS tasks. For the latter, utilization_avg_contrib has been re-introduced to compute the usage of a CPU by CFS tasks. 'group_capacity_factor' and 'group_has_free_capacity' has been removed and replaced by 'group_no_capacity'. We compare the number of task with the number of CPUs and we evaluate the level of utilization of the CPUs to define if a group is overloaded or if a group has capacity to handle more tasks. For SD_PREFER_SIBLING, a group is tagged overloaded if it has more than 1 task so it will be selected in priority (among the overloaded groups). Since [1], SD_PREFER_SIBLING is no more concerned by the computation of 'load_above_capacity' because local is not overloaded. [1] 9a5d9ba6a363 ("sched/fair: Allow calculate_imbalance() to move idle cpus") Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Morten.Rasmussen@arm.com Cc: dietmar.eggemann@arm.com Cc: efault@gmx.de Cc: kamalesh@linux.vnet.ibm.com Cc: linaro-kernel@lists.linaro.org Cc: nicolas.pitre@linaro.org Cc: preeti@linux.vnet.ibm.com Cc: riel@redhat.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/1425052454-25797-9-git-send-email-vincent.guittot@linaro.org [ Tidied up the changelog. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-02-27 15:54:11 +00:00
if ((sgs->group_capacity * 100) <
(sgs->group_util * imbalance_pct))
sched: Replace capacity_factor by usage The scheduler tries to compute how many tasks a group of CPUs can handle by assuming that a task's load is SCHED_LOAD_SCALE and a CPU's capacity is SCHED_CAPACITY_SCALE. 'struct sg_lb_stats:group_capacity_factor' divides the capacity of the group by SCHED_LOAD_SCALE to estimate how many task can run in the group. Then, it compares this value with the sum of nr_running to decide if the group is overloaded or not. But the 'group_capacity_factor' concept is hardly working for SMT systems, it sometimes works for big cores but fails to do the right thing for little cores. Below are two examples to illustrate the problem that this patch solves: 1- If the original capacity of a CPU is less than SCHED_CAPACITY_SCALE (640 as an example), a group of 3 CPUS will have a max capacity_factor of 2 (div_round_closest(3x640/1024) = 2) which means that it will be seen as overloaded even if we have only one task per CPU. 2 - If the original capacity of a CPU is greater than SCHED_CAPACITY_SCALE (1512 as an example), a group of 4 CPUs will have a capacity_factor of 4 (at max and thanks to the fix [0] for SMT system that prevent the apparition of ghost CPUs) but if one CPU is fully used by rt tasks (and its capacity is reduced to nearly nothing), the capacity factor of the group will still be 4 (div_round_closest(3*1512/1024) = 5 which is cap to 4 with [0]). So, this patch tries to solve this issue by removing capacity_factor and replacing it with the 2 following metrics: - The available CPU's capacity for CFS tasks which is already used by load_balance(). - The usage of the CPU by the CFS tasks. For the latter, utilization_avg_contrib has been re-introduced to compute the usage of a CPU by CFS tasks. 'group_capacity_factor' and 'group_has_free_capacity' has been removed and replaced by 'group_no_capacity'. We compare the number of task with the number of CPUs and we evaluate the level of utilization of the CPUs to define if a group is overloaded or if a group has capacity to handle more tasks. For SD_PREFER_SIBLING, a group is tagged overloaded if it has more than 1 task so it will be selected in priority (among the overloaded groups). Since [1], SD_PREFER_SIBLING is no more concerned by the computation of 'load_above_capacity' because local is not overloaded. [1] 9a5d9ba6a363 ("sched/fair: Allow calculate_imbalance() to move idle cpus") Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Morten.Rasmussen@arm.com Cc: dietmar.eggemann@arm.com Cc: efault@gmx.de Cc: kamalesh@linux.vnet.ibm.com Cc: linaro-kernel@lists.linaro.org Cc: nicolas.pitre@linaro.org Cc: preeti@linux.vnet.ibm.com Cc: riel@redhat.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/1425052454-25797-9-git-send-email-vincent.guittot@linaro.org [ Tidied up the changelog. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-02-27 15:54:11 +00:00
return true;
if ((sgs->group_capacity * imbalance_pct) <
(sgs->group_runnable * 100))
return true;
sched: Replace capacity_factor by usage The scheduler tries to compute how many tasks a group of CPUs can handle by assuming that a task's load is SCHED_LOAD_SCALE and a CPU's capacity is SCHED_CAPACITY_SCALE. 'struct sg_lb_stats:group_capacity_factor' divides the capacity of the group by SCHED_LOAD_SCALE to estimate how many task can run in the group. Then, it compares this value with the sum of nr_running to decide if the group is overloaded or not. But the 'group_capacity_factor' concept is hardly working for SMT systems, it sometimes works for big cores but fails to do the right thing for little cores. Below are two examples to illustrate the problem that this patch solves: 1- If the original capacity of a CPU is less than SCHED_CAPACITY_SCALE (640 as an example), a group of 3 CPUS will have a max capacity_factor of 2 (div_round_closest(3x640/1024) = 2) which means that it will be seen as overloaded even if we have only one task per CPU. 2 - If the original capacity of a CPU is greater than SCHED_CAPACITY_SCALE (1512 as an example), a group of 4 CPUs will have a capacity_factor of 4 (at max and thanks to the fix [0] for SMT system that prevent the apparition of ghost CPUs) but if one CPU is fully used by rt tasks (and its capacity is reduced to nearly nothing), the capacity factor of the group will still be 4 (div_round_closest(3*1512/1024) = 5 which is cap to 4 with [0]). So, this patch tries to solve this issue by removing capacity_factor and replacing it with the 2 following metrics: - The available CPU's capacity for CFS tasks which is already used by load_balance(). - The usage of the CPU by the CFS tasks. For the latter, utilization_avg_contrib has been re-introduced to compute the usage of a CPU by CFS tasks. 'group_capacity_factor' and 'group_has_free_capacity' has been removed and replaced by 'group_no_capacity'. We compare the number of task with the number of CPUs and we evaluate the level of utilization of the CPUs to define if a group is overloaded or if a group has capacity to handle more tasks. For SD_PREFER_SIBLING, a group is tagged overloaded if it has more than 1 task so it will be selected in priority (among the overloaded groups). Since [1], SD_PREFER_SIBLING is no more concerned by the computation of 'load_above_capacity' because local is not overloaded. [1] 9a5d9ba6a363 ("sched/fair: Allow calculate_imbalance() to move idle cpus") Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Morten.Rasmussen@arm.com Cc: dietmar.eggemann@arm.com Cc: efault@gmx.de Cc: kamalesh@linux.vnet.ibm.com Cc: linaro-kernel@lists.linaro.org Cc: nicolas.pitre@linaro.org Cc: preeti@linux.vnet.ibm.com Cc: riel@redhat.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/1425052454-25797-9-git-send-email-vincent.guittot@linaro.org [ Tidied up the changelog. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-02-27 15:54:11 +00:00
return false;
}
/*
* group_smaller_min_cpu_capacity: Returns true if sched_group sg has smaller
* per-CPU capacity than sched_group ref.
*/
static inline bool
group_smaller_min_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
{
return fits_capacity(sg->sgc->min_capacity, ref->sgc->min_capacity);
}
/*
* group_smaller_max_cpu_capacity: Returns true if sched_group sg has smaller
* per-CPU capacity_orig than sched_group ref.
*/
static inline bool
group_smaller_max_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
{
return fits_capacity(sg->sgc->max_capacity, ref->sgc->max_capacity);
}
static inline enum
group_type group_classify(unsigned int imbalance_pct,
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
struct sched_group *group,
struct sg_lb_stats *sgs)
sched/fair: Make update_sd_pick_busiest() return 'true' on a busier sd Currently update_sd_pick_busiest only identifies the busiest sd that is either overloaded, or has a group imbalance. When no sd is imbalanced or overloaded, the load balancer fails to find the busiest domain. This breaks load balancing between domains that are not overloaded, in the !SD_ASYM_PACKING case. This patch makes update_sd_pick_busiest return true when the busiest sd yet is encountered. Groups are ranked in the order overloaded > imbalanced > other, with higher ranked groups getting priority even when their load is lower. This is necessary due to the possibility of unequal capacities and cpumasks between domains within a sched group. Behaviour for SD_ASYM_PACKING does not seem to match the comment, but I have no hardware to test that so I have left the behaviour of that code unchanged. Enum for group classification suggested by Peter Zijlstra. Signed-off-by: Rik van Riel <riel@redhat.com> [peterz: replaced sg_lb_stats::group_imb with the new enum group_type in an attempt to avoid endless recalculation] Signed-off-by: Peter Zijlstra <peterz@infradead.org> Acked-by: Vincent Guittot <vincent.guittot@linaro.org> Acked-by: Michael Neuling <mikey@neuling.org> Cc: ktkhai@parallels.com Cc: tim.c.chen@linux.intel.com Cc: nicolas.pitre@linaro.org Cc: jhladky@redhat.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/20140729152743.GI3935@laptop Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-07-28 18:16:28 +00:00
{
if (group_is_overloaded(imbalance_pct, sgs))
sched/fair: Make update_sd_pick_busiest() return 'true' on a busier sd Currently update_sd_pick_busiest only identifies the busiest sd that is either overloaded, or has a group imbalance. When no sd is imbalanced or overloaded, the load balancer fails to find the busiest domain. This breaks load balancing between domains that are not overloaded, in the !SD_ASYM_PACKING case. This patch makes update_sd_pick_busiest return true when the busiest sd yet is encountered. Groups are ranked in the order overloaded > imbalanced > other, with higher ranked groups getting priority even when their load is lower. This is necessary due to the possibility of unequal capacities and cpumasks between domains within a sched group. Behaviour for SD_ASYM_PACKING does not seem to match the comment, but I have no hardware to test that so I have left the behaviour of that code unchanged. Enum for group classification suggested by Peter Zijlstra. Signed-off-by: Rik van Riel <riel@redhat.com> [peterz: replaced sg_lb_stats::group_imb with the new enum group_type in an attempt to avoid endless recalculation] Signed-off-by: Peter Zijlstra <peterz@infradead.org> Acked-by: Vincent Guittot <vincent.guittot@linaro.org> Acked-by: Michael Neuling <mikey@neuling.org> Cc: ktkhai@parallels.com Cc: tim.c.chen@linux.intel.com Cc: nicolas.pitre@linaro.org Cc: jhladky@redhat.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/20140729152743.GI3935@laptop Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-07-28 18:16:28 +00:00
return group_overloaded;
if (sg_imbalanced(group))
return group_imbalanced;
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
if (sgs->group_asym_packing)
return group_asym_packing;
sched/fair: Add 'group_misfit_task' load-balance type To maximize throughput in systems with asymmetric CPU capacities (e.g. ARM big.LITTLE) load-balancing has to consider task and CPU utilization as well as per-CPU compute capacity when load-balancing in addition to the current average load based load-balancing policy. Tasks with high utilization that are scheduled on a lower capacity CPU need to be identified and migrated to a higher capacity CPU if possible to maximize throughput. To implement this additional policy an additional group_type (load-balance scenario) is added: 'group_misfit_task'. This represents scenarios where a sched_group has one or more tasks that are not suitable for its per-CPU capacity. 'group_misfit_task' is only considered if the system is not overloaded or imbalanced ('group_imbalanced' or 'group_overloaded'). Identifying misfit tasks requires the rq lock to be held. To avoid taking remote rq locks to examine source sched_groups for misfit tasks, each CPU is responsible for tracking misfit tasks themselves and update the rq->misfit_task flag. This means checking task utilization when tasks are scheduled and on sched_tick. Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: gaku.inami.xh@renesas.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1530699470-29808-3-git-send-email-morten.rasmussen@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-07-04 10:17:40 +00:00
if (sgs->group_misfit_task_load)
return group_misfit_task;
if (!group_has_capacity(imbalance_pct, sgs))
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
return group_fully_busy;
return group_has_spare;
sched/fair: Make update_sd_pick_busiest() return 'true' on a busier sd Currently update_sd_pick_busiest only identifies the busiest sd that is either overloaded, or has a group imbalance. When no sd is imbalanced or overloaded, the load balancer fails to find the busiest domain. This breaks load balancing between domains that are not overloaded, in the !SD_ASYM_PACKING case. This patch makes update_sd_pick_busiest return true when the busiest sd yet is encountered. Groups are ranked in the order overloaded > imbalanced > other, with higher ranked groups getting priority even when their load is lower. This is necessary due to the possibility of unequal capacities and cpumasks between domains within a sched group. Behaviour for SD_ASYM_PACKING does not seem to match the comment, but I have no hardware to test that so I have left the behaviour of that code unchanged. Enum for group classification suggested by Peter Zijlstra. Signed-off-by: Rik van Riel <riel@redhat.com> [peterz: replaced sg_lb_stats::group_imb with the new enum group_type in an attempt to avoid endless recalculation] Signed-off-by: Peter Zijlstra <peterz@infradead.org> Acked-by: Vincent Guittot <vincent.guittot@linaro.org> Acked-by: Michael Neuling <mikey@neuling.org> Cc: ktkhai@parallels.com Cc: tim.c.chen@linux.intel.com Cc: nicolas.pitre@linaro.org Cc: jhladky@redhat.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/20140729152743.GI3935@laptop Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-07-28 18:16:28 +00:00
}
static bool update_nohz_stats(struct rq *rq, bool force)
{
#ifdef CONFIG_NO_HZ_COMMON
unsigned int cpu = rq->cpu;
if (!rq->has_blocked_load)
return false;
if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
return false;
if (!force && !time_after(jiffies, rq->last_blocked_load_update_tick))
return true;
update_blocked_averages(cpu);
return rq->has_blocked_load;
#else
return false;
#endif
}
/**
* update_sg_lb_stats - Update sched_group's statistics for load balancing.
* @env: The load balancing environment.
* @group: sched_group whose statistics are to be updated.
* @sgs: variable to hold the statistics for this group.
sched/fair: Clean-up update_sg_lb_stats parameters In preparation for the introduction of a new root domain flag which can be set during load balance (the 'overutilized' flag), clean-up the set of parameters passed to update_sg_lb_stats(). More specifically, the 'local_group' and 'local_idx' parameters can be removed since they can easily be reconstructed from within the function. While at it, transform the 'overload' parameter into a flag stored in the 'sg_status' parameter hence facilitating the definition of new flags when needed. Suggested-by: Peter Zijlstra <peterz@infradead.org> Suggested-by: Valentin Schneider <valentin.schneider@arm.com> Signed-off-by: Quentin Perret <quentin.perret@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: adharmap@codeaurora.org Cc: chris.redpath@arm.com Cc: currojerez@riseup.net Cc: dietmar.eggemann@arm.com Cc: edubezval@gmail.com Cc: gregkh@linuxfoundation.org Cc: javi.merino@kernel.org Cc: joel@joelfernandes.org Cc: juri.lelli@redhat.com Cc: morten.rasmussen@arm.com Cc: patrick.bellasi@arm.com Cc: pkondeti@codeaurora.org Cc: rjw@rjwysocki.net Cc: skannan@codeaurora.org Cc: smuckle@google.com Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Cc: tkjos@google.com Cc: vincent.guittot@linaro.org Cc: viresh.kumar@linaro.org Link: https://lkml.kernel.org/r/20181203095628.11858-12-quentin.perret@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-12-03 09:56:24 +00:00
* @sg_status: Holds flag indicating the status of the sched_group
*/
static inline void update_sg_lb_stats(struct lb_env *env,
sched/fair: Clean-up update_sg_lb_stats parameters In preparation for the introduction of a new root domain flag which can be set during load balance (the 'overutilized' flag), clean-up the set of parameters passed to update_sg_lb_stats(). More specifically, the 'local_group' and 'local_idx' parameters can be removed since they can easily be reconstructed from within the function. While at it, transform the 'overload' parameter into a flag stored in the 'sg_status' parameter hence facilitating the definition of new flags when needed. Suggested-by: Peter Zijlstra <peterz@infradead.org> Suggested-by: Valentin Schneider <valentin.schneider@arm.com> Signed-off-by: Quentin Perret <quentin.perret@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: adharmap@codeaurora.org Cc: chris.redpath@arm.com Cc: currojerez@riseup.net Cc: dietmar.eggemann@arm.com Cc: edubezval@gmail.com Cc: gregkh@linuxfoundation.org Cc: javi.merino@kernel.org Cc: joel@joelfernandes.org Cc: juri.lelli@redhat.com Cc: morten.rasmussen@arm.com Cc: patrick.bellasi@arm.com Cc: pkondeti@codeaurora.org Cc: rjw@rjwysocki.net Cc: skannan@codeaurora.org Cc: smuckle@google.com Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Cc: tkjos@google.com Cc: vincent.guittot@linaro.org Cc: viresh.kumar@linaro.org Link: https://lkml.kernel.org/r/20181203095628.11858-12-quentin.perret@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-12-03 09:56:24 +00:00
struct sched_group *group,
struct sg_lb_stats *sgs,
int *sg_status)
{
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
int i, nr_running, local_group;
memset(sgs, 0, sizeof(*sgs));
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(group));
for_each_cpu_and(i, sched_group_span(group), env->cpus) {
struct rq *rq = cpu_rq(i);
if ((env->flags & LBF_NOHZ_STATS) && update_nohz_stats(rq, false))
env->flags |= LBF_NOHZ_AGAIN;
sgs->group_load += cpu_load(rq);
sgs->group_util += cpu_util(i);
sgs->group_runnable += cpu_runnable(rq);
sgs->sum_h_nr_running += rq->cfs.h_nr_running;
sched/fair: Implement fast idling of CPUs when the system is partially loaded When a system is lightly loaded (i.e. no more than 1 job per cpu), attempt to pull job to a cpu before putting it to idle is unnecessary and can be skipped. This patch adds an indicator so the scheduler can know when there's no more than 1 active job is on any CPU in the system to skip needless job pulls. On a 4 socket machine with a request/response kind of workload from clients, we saw about 0.13 msec delay when we go through a full load balance to try pull job from all the other cpus. While 0.1 msec was spent on processing the request and generating a response, the 0.13 msec load balance overhead was actually more than the actual work being done. This overhead can be skipped much of the time for lightly loaded systems. With this patch, we tested with a netperf request/response workload that has the server busy with half the cpus in a 4 socket system. We found the patch eliminated 75% of the load balance attempts before idling a cpu. The overhead of setting/clearing the indicator is low as we already gather the necessary info while we call add_nr_running() and update_sd_lb_stats.() We switch to full load balance load immediately if any cpu got more than one job on its run queue in add_nr_running. We'll clear the indicator to avoid load balance when we detect no cpu's have more than one job when we scan the work queues in update_sg_lb_stats(). We are aggressive in turning on the load balance and opportunistic in skipping the load balance. Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Acked-by: Rik van Riel <riel@redhat.com> Acked-by: Jason Low <jason.low2@hp.com> Cc: "Paul E.McKenney" <paulmck@linux.vnet.ibm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Davidlohr Bueso <davidlohr@hp.com> Cc: Alex Shi <alex.shi@linaro.org> Cc: Michel Lespinasse <walken@google.com> Cc: Peter Hurley <peter@hurleysoftware.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1403551009.2970.613.camel@schen9-DESK Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-06-23 19:16:49 +00:00
nr_running = rq->nr_running;
sgs->sum_nr_running += nr_running;
if (nr_running > 1)
sched/fair: Clean-up update_sg_lb_stats parameters In preparation for the introduction of a new root domain flag which can be set during load balance (the 'overutilized' flag), clean-up the set of parameters passed to update_sg_lb_stats(). More specifically, the 'local_group' and 'local_idx' parameters can be removed since they can easily be reconstructed from within the function. While at it, transform the 'overload' parameter into a flag stored in the 'sg_status' parameter hence facilitating the definition of new flags when needed. Suggested-by: Peter Zijlstra <peterz@infradead.org> Suggested-by: Valentin Schneider <valentin.schneider@arm.com> Signed-off-by: Quentin Perret <quentin.perret@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: adharmap@codeaurora.org Cc: chris.redpath@arm.com Cc: currojerez@riseup.net Cc: dietmar.eggemann@arm.com Cc: edubezval@gmail.com Cc: gregkh@linuxfoundation.org Cc: javi.merino@kernel.org Cc: joel@joelfernandes.org Cc: juri.lelli@redhat.com Cc: morten.rasmussen@arm.com Cc: patrick.bellasi@arm.com Cc: pkondeti@codeaurora.org Cc: rjw@rjwysocki.net Cc: skannan@codeaurora.org Cc: smuckle@google.com Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Cc: tkjos@google.com Cc: vincent.guittot@linaro.org Cc: viresh.kumar@linaro.org Link: https://lkml.kernel.org/r/20181203095628.11858-12-quentin.perret@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-12-03 09:56:24 +00:00
*sg_status |= SG_OVERLOAD;
sched/fair: Implement fast idling of CPUs when the system is partially loaded When a system is lightly loaded (i.e. no more than 1 job per cpu), attempt to pull job to a cpu before putting it to idle is unnecessary and can be skipped. This patch adds an indicator so the scheduler can know when there's no more than 1 active job is on any CPU in the system to skip needless job pulls. On a 4 socket machine with a request/response kind of workload from clients, we saw about 0.13 msec delay when we go through a full load balance to try pull job from all the other cpus. While 0.1 msec was spent on processing the request and generating a response, the 0.13 msec load balance overhead was actually more than the actual work being done. This overhead can be skipped much of the time for lightly loaded systems. With this patch, we tested with a netperf request/response workload that has the server busy with half the cpus in a 4 socket system. We found the patch eliminated 75% of the load balance attempts before idling a cpu. The overhead of setting/clearing the indicator is low as we already gather the necessary info while we call add_nr_running() and update_sd_lb_stats.() We switch to full load balance load immediately if any cpu got more than one job on its run queue in add_nr_running. We'll clear the indicator to avoid load balance when we detect no cpu's have more than one job when we scan the work queues in update_sg_lb_stats(). We are aggressive in turning on the load balance and opportunistic in skipping the load balance. Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Acked-by: Rik van Riel <riel@redhat.com> Acked-by: Jason Low <jason.low2@hp.com> Cc: "Paul E.McKenney" <paulmck@linux.vnet.ibm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Davidlohr Bueso <davidlohr@hp.com> Cc: Alex Shi <alex.shi@linaro.org> Cc: Michel Lespinasse <walken@google.com> Cc: Peter Hurley <peter@hurleysoftware.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1403551009.2970.613.camel@schen9-DESK Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-06-23 19:16:49 +00:00
sched/fair: Add over-utilization/tipping point indicator Energy-aware scheduling is only meant to be active while the system is _not_ over-utilized. That is, there are spare cycles available to shift tasks around based on their actual utilization to get a more energy-efficient task distribution without depriving any tasks. When above the tipping point task placement is done the traditional way based on load_avg, spreading the tasks across as many cpus as possible based on priority scaled load to preserve smp_nice. Below the tipping point we want to use util_avg instead. We need to define a criteria for when we make the switch. The util_avg for each cpu converges towards 100% regardless of how many additional tasks we may put on it. If we define over-utilized as: sum_{cpus}(rq.cfs.avg.util_avg) + margin > sum_{cpus}(rq.capacity) some individual cpus may be over-utilized running multiple tasks even when the above condition is false. That should be okay as long as we try to spread the tasks out to avoid per-cpu over-utilization as much as possible and if all tasks have the _same_ priority. If the latter isn't true, we have to consider priority to preserve smp_nice. For example, we could have n_cpus nice=-10 util_avg=55% tasks and n_cpus/2 nice=0 util_avg=60% tasks. Balancing based on util_avg we are likely to end up with nice=-10 tasks sharing cpus and nice=0 tasks getting their own as we 1.5*n_cpus tasks in total and 55%+55% is less over-utilized than 55%+60% for those cpus that have to be shared. The system utilization is only 85% of the system capacity, but we are breaking smp_nice. To be sure not to break smp_nice, we have defined over-utilization conservatively as when any cpu in the system is fully utilized at its highest frequency instead: cpu_rq(any).cfs.avg.util_avg + margin > cpu_rq(any).capacity IOW, as soon as one cpu is (nearly) 100% utilized, we switch to load_avg to factor in priority to preserve smp_nice. With this definition, we can skip periodic load-balance as no cpu has an always-running task when the system is not over-utilized. All tasks will be periodic and we can balance them at wake-up. This conservative condition does however mean that some scenarios that could benefit from energy-aware decisions even if one cpu is fully utilized would not get those benefits. For systems where some cpus might have reduced capacity on some cpus (RT-pressure and/or big.LITTLE), we want periodic load-balance checks as soon a just a single cpu is fully utilized as it might one of those with reduced capacity and in that case we want to migrate it. [ peterz: Added a comment explaining why new tasks are not accounted during overutilization detection. ] Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Quentin Perret <quentin.perret@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: adharmap@codeaurora.org Cc: chris.redpath@arm.com Cc: currojerez@riseup.net Cc: dietmar.eggemann@arm.com Cc: edubezval@gmail.com Cc: gregkh@linuxfoundation.org Cc: javi.merino@kernel.org Cc: joel@joelfernandes.org Cc: juri.lelli@redhat.com Cc: patrick.bellasi@arm.com Cc: pkondeti@codeaurora.org Cc: rjw@rjwysocki.net Cc: skannan@codeaurora.org Cc: smuckle@google.com Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Cc: viresh.kumar@linaro.org Link: https://lkml.kernel.org/r/20181203095628.11858-13-quentin.perret@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-12-03 09:56:25 +00:00
if (cpu_overutilized(i))
*sg_status |= SG_OVERUTILIZED;
sched/fair: Implement fast idling of CPUs when the system is partially loaded When a system is lightly loaded (i.e. no more than 1 job per cpu), attempt to pull job to a cpu before putting it to idle is unnecessary and can be skipped. This patch adds an indicator so the scheduler can know when there's no more than 1 active job is on any CPU in the system to skip needless job pulls. On a 4 socket machine with a request/response kind of workload from clients, we saw about 0.13 msec delay when we go through a full load balance to try pull job from all the other cpus. While 0.1 msec was spent on processing the request and generating a response, the 0.13 msec load balance overhead was actually more than the actual work being done. This overhead can be skipped much of the time for lightly loaded systems. With this patch, we tested with a netperf request/response workload that has the server busy with half the cpus in a 4 socket system. We found the patch eliminated 75% of the load balance attempts before idling a cpu. The overhead of setting/clearing the indicator is low as we already gather the necessary info while we call add_nr_running() and update_sd_lb_stats.() We switch to full load balance load immediately if any cpu got more than one job on its run queue in add_nr_running. We'll clear the indicator to avoid load balance when we detect no cpu's have more than one job when we scan the work queues in update_sg_lb_stats(). We are aggressive in turning on the load balance and opportunistic in skipping the load balance. Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Acked-by: Rik van Riel <riel@redhat.com> Acked-by: Jason Low <jason.low2@hp.com> Cc: "Paul E.McKenney" <paulmck@linux.vnet.ibm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Davidlohr Bueso <davidlohr@hp.com> Cc: Alex Shi <alex.shi@linaro.org> Cc: Michel Lespinasse <walken@google.com> Cc: Peter Hurley <peter@hurleysoftware.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1403551009.2970.613.camel@schen9-DESK Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-06-23 19:16:49 +00:00
#ifdef CONFIG_NUMA_BALANCING
sgs->nr_numa_running += rq->nr_numa_running;
sgs->nr_preferred_running += rq->nr_preferred_running;
#endif
/*
* No need to call idle_cpu() if nr_running is not 0
*/
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
if (!nr_running && idle_cpu(i)) {
sched: Use group weight, idle cpu metrics to fix imbalances during idle Currently we consider a sched domain to be well balanced when the imbalance is less than the domain's imablance_pct. As the number of cores and threads are increasing, current values of imbalance_pct (for example 25% for a NUMA domain) are not enough to detect imbalances like: a) On a WSM-EP system (two sockets, each having 6 cores and 12 logical threads), 24 cpu-hogging tasks get scheduled as 13 on one socket and 11 on another socket. Leading to an idle HT cpu. b) On a hypothetial 2 socket NHM-EX system (each socket having 8 cores and 16 logical threads), 16 cpu-hogging tasks can get scheduled as 9 on one socket and 7 on another socket. Leaving one core in a socket idle whereas in another socket we have a core having both its HT siblings busy. While this issue can be fixed by decreasing the domain's imbalance_pct (by making it a function of number of logical cpus in the domain), it can potentially cause more task migrations across sched groups in an overloaded case. Fix this by using imbalance_pct only during newly_idle and busy load balancing. And during idle load balancing, check if there is an imbalance in number of idle cpu's across the busiest and this sched_group or if the busiest group has more tasks than its weight that the idle cpu in this_group can pull. Reported-by: Nikhil Rao <ncrao@google.com> Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> LKML-Reference: <1284760952.2676.11.camel@sbsiddha-MOBL3.sc.intel.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-09-17 22:02:32 +00:00
sgs->idle_cpus++;
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
/* Idle cpu can't have misfit task */
continue;
}
if (local_group)
continue;
sched/fair: Add 'group_misfit_task' load-balance type To maximize throughput in systems with asymmetric CPU capacities (e.g. ARM big.LITTLE) load-balancing has to consider task and CPU utilization as well as per-CPU compute capacity when load-balancing in addition to the current average load based load-balancing policy. Tasks with high utilization that are scheduled on a lower capacity CPU need to be identified and migrated to a higher capacity CPU if possible to maximize throughput. To implement this additional policy an additional group_type (load-balance scenario) is added: 'group_misfit_task'. This represents scenarios where a sched_group has one or more tasks that are not suitable for its per-CPU capacity. 'group_misfit_task' is only considered if the system is not overloaded or imbalanced ('group_imbalanced' or 'group_overloaded'). Identifying misfit tasks requires the rq lock to be held. To avoid taking remote rq locks to examine source sched_groups for misfit tasks, each CPU is responsible for tracking misfit tasks themselves and update the rq->misfit_task flag. This means checking task utilization when tasks are scheduled and on sched_tick. Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: gaku.inami.xh@renesas.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1530699470-29808-3-git-send-email-morten.rasmussen@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-07-04 10:17:40 +00:00
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
/* Check for a misfit task on the cpu */
sched/fair: Add 'group_misfit_task' load-balance type To maximize throughput in systems with asymmetric CPU capacities (e.g. ARM big.LITTLE) load-balancing has to consider task and CPU utilization as well as per-CPU compute capacity when load-balancing in addition to the current average load based load-balancing policy. Tasks with high utilization that are scheduled on a lower capacity CPU need to be identified and migrated to a higher capacity CPU if possible to maximize throughput. To implement this additional policy an additional group_type (load-balance scenario) is added: 'group_misfit_task'. This represents scenarios where a sched_group has one or more tasks that are not suitable for its per-CPU capacity. 'group_misfit_task' is only considered if the system is not overloaded or imbalanced ('group_imbalanced' or 'group_overloaded'). Identifying misfit tasks requires the rq lock to be held. To avoid taking remote rq locks to examine source sched_groups for misfit tasks, each CPU is responsible for tracking misfit tasks themselves and update the rq->misfit_task flag. This means checking task utilization when tasks are scheduled and on sched_tick. Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: gaku.inami.xh@renesas.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1530699470-29808-3-git-send-email-morten.rasmussen@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-07-04 10:17:40 +00:00
if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
sched/fair: Set rq->rd->overload when misfit Idle balance is a great opportunity to pull a misfit task. However, there are scenarios where misfit tasks are present but idle balance is prevented by the overload flag. A good example of this is a workload of n identical tasks. Let's suppose we have a 2+2 Arm big.LITTLE system. We then spawn 4 fairly CPU-intensive tasks - for the sake of simplicity let's say they are just CPU hogs, even when running on big CPUs. They are identical tasks, so on an SMP system they should all end at (roughly) the same time. However, in our case the LITTLE CPUs are less performing than the big CPUs, so tasks running on the LITTLEs will have a longer completion time. This means that the big CPUs will complete their work earlier, at which point they should pull the tasks from the LITTLEs. What we want to happen is summarized as follows: a,b,c,d are our CPU-hogging tasks _ signifies idling LITTLE_0 | a a a a _ _ LITTLE_1 | b b b b _ _ ---------|------------- big_0 | c c c c a a big_1 | d d d d b b ^ ^ Tasks end on the big CPUs, idle balance happens and the misfit tasks are pulled straight away This however won't happen, because currently the overload flag is only set when there is any CPU that has more than one runnable task - which may very well not be the case here if our CPU-hogging workload is all there is to run. As such, this commit sets the overload flag in update_sg_lb_stats when a group is flagged as having a misfit task. Signed-off-by: Valentin Schneider <valentin.schneider@arm.com> Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: gaku.inami.xh@renesas.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1530699470-29808-10-git-send-email-morten.rasmussen@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-07-04 10:17:47 +00:00
sgs->group_misfit_task_load < rq->misfit_task_load) {
sched/fair: Add 'group_misfit_task' load-balance type To maximize throughput in systems with asymmetric CPU capacities (e.g. ARM big.LITTLE) load-balancing has to consider task and CPU utilization as well as per-CPU compute capacity when load-balancing in addition to the current average load based load-balancing policy. Tasks with high utilization that are scheduled on a lower capacity CPU need to be identified and migrated to a higher capacity CPU if possible to maximize throughput. To implement this additional policy an additional group_type (load-balance scenario) is added: 'group_misfit_task'. This represents scenarios where a sched_group has one or more tasks that are not suitable for its per-CPU capacity. 'group_misfit_task' is only considered if the system is not overloaded or imbalanced ('group_imbalanced' or 'group_overloaded'). Identifying misfit tasks requires the rq lock to be held. To avoid taking remote rq locks to examine source sched_groups for misfit tasks, each CPU is responsible for tracking misfit tasks themselves and update the rq->misfit_task flag. This means checking task utilization when tasks are scheduled and on sched_tick. Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: gaku.inami.xh@renesas.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1530699470-29808-3-git-send-email-morten.rasmussen@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-07-04 10:17:40 +00:00
sgs->group_misfit_task_load = rq->misfit_task_load;
sched/fair: Clean-up update_sg_lb_stats parameters In preparation for the introduction of a new root domain flag which can be set during load balance (the 'overutilized' flag), clean-up the set of parameters passed to update_sg_lb_stats(). More specifically, the 'local_group' and 'local_idx' parameters can be removed since they can easily be reconstructed from within the function. While at it, transform the 'overload' parameter into a flag stored in the 'sg_status' parameter hence facilitating the definition of new flags when needed. Suggested-by: Peter Zijlstra <peterz@infradead.org> Suggested-by: Valentin Schneider <valentin.schneider@arm.com> Signed-off-by: Quentin Perret <quentin.perret@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: adharmap@codeaurora.org Cc: chris.redpath@arm.com Cc: currojerez@riseup.net Cc: dietmar.eggemann@arm.com Cc: edubezval@gmail.com Cc: gregkh@linuxfoundation.org Cc: javi.merino@kernel.org Cc: joel@joelfernandes.org Cc: juri.lelli@redhat.com Cc: morten.rasmussen@arm.com Cc: patrick.bellasi@arm.com Cc: pkondeti@codeaurora.org Cc: rjw@rjwysocki.net Cc: skannan@codeaurora.org Cc: smuckle@google.com Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Cc: tkjos@google.com Cc: vincent.guittot@linaro.org Cc: viresh.kumar@linaro.org Link: https://lkml.kernel.org/r/20181203095628.11858-12-quentin.perret@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-12-03 09:56:24 +00:00
*sg_status |= SG_OVERLOAD;
sched/fair: Set rq->rd->overload when misfit Idle balance is a great opportunity to pull a misfit task. However, there are scenarios where misfit tasks are present but idle balance is prevented by the overload flag. A good example of this is a workload of n identical tasks. Let's suppose we have a 2+2 Arm big.LITTLE system. We then spawn 4 fairly CPU-intensive tasks - for the sake of simplicity let's say they are just CPU hogs, even when running on big CPUs. They are identical tasks, so on an SMP system they should all end at (roughly) the same time. However, in our case the LITTLE CPUs are less performing than the big CPUs, so tasks running on the LITTLEs will have a longer completion time. This means that the big CPUs will complete their work earlier, at which point they should pull the tasks from the LITTLEs. What we want to happen is summarized as follows: a,b,c,d are our CPU-hogging tasks _ signifies idling LITTLE_0 | a a a a _ _ LITTLE_1 | b b b b _ _ ---------|------------- big_0 | c c c c a a big_1 | d d d d b b ^ ^ Tasks end on the big CPUs, idle balance happens and the misfit tasks are pulled straight away This however won't happen, because currently the overload flag is only set when there is any CPU that has more than one runnable task - which may very well not be the case here if our CPU-hogging workload is all there is to run. As such, this commit sets the overload flag in update_sg_lb_stats when a group is flagged as having a misfit task. Signed-off-by: Valentin Schneider <valentin.schneider@arm.com> Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: gaku.inami.xh@renesas.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1530699470-29808-10-git-send-email-morten.rasmussen@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-07-04 10:17:47 +00:00
}
}
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
/* Check if dst CPU is idle and preferred to this group */
if (env->sd->flags & SD_ASYM_PACKING &&
env->idle != CPU_NOT_IDLE &&
sgs->sum_h_nr_running &&
sched_asym_prefer(env->dst_cpu, group->asym_prefer_cpu)) {
sgs->group_asym_packing = 1;
}
sgs->group_capacity = group->sgc->capacity;
sched: Use group weight, idle cpu metrics to fix imbalances during idle Currently we consider a sched domain to be well balanced when the imbalance is less than the domain's imablance_pct. As the number of cores and threads are increasing, current values of imbalance_pct (for example 25% for a NUMA domain) are not enough to detect imbalances like: a) On a WSM-EP system (two sockets, each having 6 cores and 12 logical threads), 24 cpu-hogging tasks get scheduled as 13 on one socket and 11 on another socket. Leading to an idle HT cpu. b) On a hypothetial 2 socket NHM-EX system (each socket having 8 cores and 16 logical threads), 16 cpu-hogging tasks can get scheduled as 9 on one socket and 7 on another socket. Leaving one core in a socket idle whereas in another socket we have a core having both its HT siblings busy. While this issue can be fixed by decreasing the domain's imbalance_pct (by making it a function of number of logical cpus in the domain), it can potentially cause more task migrations across sched groups in an overloaded case. Fix this by using imbalance_pct only during newly_idle and busy load balancing. And during idle load balancing, check if there is an imbalance in number of idle cpu's across the busiest and this sched_group or if the busiest group has more tasks than its weight that the idle cpu in this_group can pull. Reported-by: Nikhil Rao <ncrao@google.com> Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> LKML-Reference: <1284760952.2676.11.camel@sbsiddha-MOBL3.sc.intel.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-09-17 22:02:32 +00:00
sgs->group_weight = group->group_weight;
sgs->group_type = group_classify(env->sd->imbalance_pct, group, sgs);
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
/* Computing avg_load makes sense only when group is overloaded */
if (sgs->group_type == group_overloaded)
sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
sgs->group_capacity;
}
sched: Add asymmetric group packing option for sibling domain Check to see if the group is packed in a sched doman. This is primarily intended to used at the sibling level. Some cores like POWER7 prefer to use lower numbered SMT threads. In the case of POWER7, it can move to lower SMT modes only when higher threads are idle. When in lower SMT modes, the threads will perform better since they share less core resources. Hence when we have idle threads, we want them to be the higher ones. This adds a hook into f_b_g() called check_asym_packing() to check the packing. This packing function is run on idle threads. It checks to see if the busiest CPU in this domain (core in the P7 case) has a higher CPU number than what where the packing function is being run on. If it is, calculate the imbalance and return the higher busier thread as the busiest group to f_b_g(). Here we are assuming a lower CPU number will be equivalent to a lower SMT thread number. It also creates a new SD_ASYM_PACKING flag to enable this feature at any scheduler domain level. It also creates an arch hook to enable this feature at the sibling level. The default function doesn't enable this feature. Based heavily on patch from Peter Zijlstra. Fixes from Srivatsa Vaddagiri. Signed-off-by: Michael Neuling <mikey@neuling.org> Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Thomas Gleixner <tglx@linutronix.de> LKML-Reference: <20100608045702.2936CCC897@localhost.localdomain> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-06-08 04:57:02 +00:00
/**
* update_sd_pick_busiest - return 1 on busiest group
* @env: The load balancing environment.
sched: Add asymmetric group packing option for sibling domain Check to see if the group is packed in a sched doman. This is primarily intended to used at the sibling level. Some cores like POWER7 prefer to use lower numbered SMT threads. In the case of POWER7, it can move to lower SMT modes only when higher threads are idle. When in lower SMT modes, the threads will perform better since they share less core resources. Hence when we have idle threads, we want them to be the higher ones. This adds a hook into f_b_g() called check_asym_packing() to check the packing. This packing function is run on idle threads. It checks to see if the busiest CPU in this domain (core in the P7 case) has a higher CPU number than what where the packing function is being run on. If it is, calculate the imbalance and return the higher busier thread as the busiest group to f_b_g(). Here we are assuming a lower CPU number will be equivalent to a lower SMT thread number. It also creates a new SD_ASYM_PACKING flag to enable this feature at any scheduler domain level. It also creates an arch hook to enable this feature at the sibling level. The default function doesn't enable this feature. Based heavily on patch from Peter Zijlstra. Fixes from Srivatsa Vaddagiri. Signed-off-by: Michael Neuling <mikey@neuling.org> Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Thomas Gleixner <tglx@linutronix.de> LKML-Reference: <20100608045702.2936CCC897@localhost.localdomain> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-06-08 04:57:02 +00:00
* @sds: sched_domain statistics
* @sg: sched_group candidate to be checked for being the busiest
* @sgs: sched_group statistics
sched: Add asymmetric group packing option for sibling domain Check to see if the group is packed in a sched doman. This is primarily intended to used at the sibling level. Some cores like POWER7 prefer to use lower numbered SMT threads. In the case of POWER7, it can move to lower SMT modes only when higher threads are idle. When in lower SMT modes, the threads will perform better since they share less core resources. Hence when we have idle threads, we want them to be the higher ones. This adds a hook into f_b_g() called check_asym_packing() to check the packing. This packing function is run on idle threads. It checks to see if the busiest CPU in this domain (core in the P7 case) has a higher CPU number than what where the packing function is being run on. If it is, calculate the imbalance and return the higher busier thread as the busiest group to f_b_g(). Here we are assuming a lower CPU number will be equivalent to a lower SMT thread number. It also creates a new SD_ASYM_PACKING flag to enable this feature at any scheduler domain level. It also creates an arch hook to enable this feature at the sibling level. The default function doesn't enable this feature. Based heavily on patch from Peter Zijlstra. Fixes from Srivatsa Vaddagiri. Signed-off-by: Michael Neuling <mikey@neuling.org> Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Thomas Gleixner <tglx@linutronix.de> LKML-Reference: <20100608045702.2936CCC897@localhost.localdomain> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-06-08 04:57:02 +00:00
*
* Determine if @sg is a busier group than the previously selected
* busiest group.
*
* Return: %true if @sg is a busier group than the previously selected
* busiest group. %false otherwise.
sched: Add asymmetric group packing option for sibling domain Check to see if the group is packed in a sched doman. This is primarily intended to used at the sibling level. Some cores like POWER7 prefer to use lower numbered SMT threads. In the case of POWER7, it can move to lower SMT modes only when higher threads are idle. When in lower SMT modes, the threads will perform better since they share less core resources. Hence when we have idle threads, we want them to be the higher ones. This adds a hook into f_b_g() called check_asym_packing() to check the packing. This packing function is run on idle threads. It checks to see if the busiest CPU in this domain (core in the P7 case) has a higher CPU number than what where the packing function is being run on. If it is, calculate the imbalance and return the higher busier thread as the busiest group to f_b_g(). Here we are assuming a lower CPU number will be equivalent to a lower SMT thread number. It also creates a new SD_ASYM_PACKING flag to enable this feature at any scheduler domain level. It also creates an arch hook to enable this feature at the sibling level. The default function doesn't enable this feature. Based heavily on patch from Peter Zijlstra. Fixes from Srivatsa Vaddagiri. Signed-off-by: Michael Neuling <mikey@neuling.org> Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Thomas Gleixner <tglx@linutronix.de> LKML-Reference: <20100608045702.2936CCC897@localhost.localdomain> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-06-08 04:57:02 +00:00
*/
static bool update_sd_pick_busiest(struct lb_env *env,
sched: Add asymmetric group packing option for sibling domain Check to see if the group is packed in a sched doman. This is primarily intended to used at the sibling level. Some cores like POWER7 prefer to use lower numbered SMT threads. In the case of POWER7, it can move to lower SMT modes only when higher threads are idle. When in lower SMT modes, the threads will perform better since they share less core resources. Hence when we have idle threads, we want them to be the higher ones. This adds a hook into f_b_g() called check_asym_packing() to check the packing. This packing function is run on idle threads. It checks to see if the busiest CPU in this domain (core in the P7 case) has a higher CPU number than what where the packing function is being run on. If it is, calculate the imbalance and return the higher busier thread as the busiest group to f_b_g(). Here we are assuming a lower CPU number will be equivalent to a lower SMT thread number. It also creates a new SD_ASYM_PACKING flag to enable this feature at any scheduler domain level. It also creates an arch hook to enable this feature at the sibling level. The default function doesn't enable this feature. Based heavily on patch from Peter Zijlstra. Fixes from Srivatsa Vaddagiri. Signed-off-by: Michael Neuling <mikey@neuling.org> Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Thomas Gleixner <tglx@linutronix.de> LKML-Reference: <20100608045702.2936CCC897@localhost.localdomain> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-06-08 04:57:02 +00:00
struct sd_lb_stats *sds,
struct sched_group *sg,
struct sg_lb_stats *sgs)
sched: Add asymmetric group packing option for sibling domain Check to see if the group is packed in a sched doman. This is primarily intended to used at the sibling level. Some cores like POWER7 prefer to use lower numbered SMT threads. In the case of POWER7, it can move to lower SMT modes only when higher threads are idle. When in lower SMT modes, the threads will perform better since they share less core resources. Hence when we have idle threads, we want them to be the higher ones. This adds a hook into f_b_g() called check_asym_packing() to check the packing. This packing function is run on idle threads. It checks to see if the busiest CPU in this domain (core in the P7 case) has a higher CPU number than what where the packing function is being run on. If it is, calculate the imbalance and return the higher busier thread as the busiest group to f_b_g(). Here we are assuming a lower CPU number will be equivalent to a lower SMT thread number. It also creates a new SD_ASYM_PACKING flag to enable this feature at any scheduler domain level. It also creates an arch hook to enable this feature at the sibling level. The default function doesn't enable this feature. Based heavily on patch from Peter Zijlstra. Fixes from Srivatsa Vaddagiri. Signed-off-by: Michael Neuling <mikey@neuling.org> Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Thomas Gleixner <tglx@linutronix.de> LKML-Reference: <20100608045702.2936CCC897@localhost.localdomain> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-06-08 04:57:02 +00:00
{
sched/fair: Make update_sd_pick_busiest() return 'true' on a busier sd Currently update_sd_pick_busiest only identifies the busiest sd that is either overloaded, or has a group imbalance. When no sd is imbalanced or overloaded, the load balancer fails to find the busiest domain. This breaks load balancing between domains that are not overloaded, in the !SD_ASYM_PACKING case. This patch makes update_sd_pick_busiest return true when the busiest sd yet is encountered. Groups are ranked in the order overloaded > imbalanced > other, with higher ranked groups getting priority even when their load is lower. This is necessary due to the possibility of unequal capacities and cpumasks between domains within a sched group. Behaviour for SD_ASYM_PACKING does not seem to match the comment, but I have no hardware to test that so I have left the behaviour of that code unchanged. Enum for group classification suggested by Peter Zijlstra. Signed-off-by: Rik van Riel <riel@redhat.com> [peterz: replaced sg_lb_stats::group_imb with the new enum group_type in an attempt to avoid endless recalculation] Signed-off-by: Peter Zijlstra <peterz@infradead.org> Acked-by: Vincent Guittot <vincent.guittot@linaro.org> Acked-by: Michael Neuling <mikey@neuling.org> Cc: ktkhai@parallels.com Cc: tim.c.chen@linux.intel.com Cc: nicolas.pitre@linaro.org Cc: jhladky@redhat.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/20140729152743.GI3935@laptop Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-07-28 18:16:28 +00:00
struct sg_lb_stats *busiest = &sds->busiest_stat;
sched: Add asymmetric group packing option for sibling domain Check to see if the group is packed in a sched doman. This is primarily intended to used at the sibling level. Some cores like POWER7 prefer to use lower numbered SMT threads. In the case of POWER7, it can move to lower SMT modes only when higher threads are idle. When in lower SMT modes, the threads will perform better since they share less core resources. Hence when we have idle threads, we want them to be the higher ones. This adds a hook into f_b_g() called check_asym_packing() to check the packing. This packing function is run on idle threads. It checks to see if the busiest CPU in this domain (core in the P7 case) has a higher CPU number than what where the packing function is being run on. If it is, calculate the imbalance and return the higher busier thread as the busiest group to f_b_g(). Here we are assuming a lower CPU number will be equivalent to a lower SMT thread number. It also creates a new SD_ASYM_PACKING flag to enable this feature at any scheduler domain level. It also creates an arch hook to enable this feature at the sibling level. The default function doesn't enable this feature. Based heavily on patch from Peter Zijlstra. Fixes from Srivatsa Vaddagiri. Signed-off-by: Michael Neuling <mikey@neuling.org> Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Thomas Gleixner <tglx@linutronix.de> LKML-Reference: <20100608045702.2936CCC897@localhost.localdomain> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-06-08 04:57:02 +00:00
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
/* Make sure that there is at least one task to pull */
if (!sgs->sum_h_nr_running)
return false;
sched/fair: Consider misfit tasks when load-balancing On asymmetric CPU capacity systems load intensive tasks can end up on CPUs that don't suit their compute demand. In this scenarios 'misfit' tasks should be migrated to CPUs with higher compute capacity to ensure better throughput. group_misfit_task indicates this scenario, but tweaks to the load-balance code are needed to make the migrations happen. Misfit balancing only makes sense between a source group of lower per-CPU capacity and destination group of higher compute capacity. Otherwise, misfit balancing is ignored. group_misfit_task has lowest priority so any imbalance due to overload is dealt with first. The modifications are: 1. Only pick a group containing misfit tasks as the busiest group if the destination group has higher capacity and has spare capacity. 2. When the busiest group is a 'misfit' group, skip the usual average load and group capacity checks. 3. Set the imbalance for 'misfit' balancing sufficiently high for a task to be pulled ignoring average load. 4. Pick the CPU with the highest misfit load as the source CPU. 5. If the misfit task is alone on the source CPU, go for active balancing. Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: gaku.inami.xh@renesas.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1530699470-29808-5-git-send-email-morten.rasmussen@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-07-04 10:17:42 +00:00
/*
* Don't try to pull misfit tasks we can't help.
* We can use max_capacity here as reduction in capacity on some
* CPUs in the group should either be possible to resolve
* internally or be covered by avg_load imbalance (eventually).
*/
if (sgs->group_type == group_misfit_task &&
(!group_smaller_max_cpu_capacity(sg, sds->local) ||
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
sds->local_stat.group_type != group_has_spare))
sched/fair: Consider misfit tasks when load-balancing On asymmetric CPU capacity systems load intensive tasks can end up on CPUs that don't suit their compute demand. In this scenarios 'misfit' tasks should be migrated to CPUs with higher compute capacity to ensure better throughput. group_misfit_task indicates this scenario, but tweaks to the load-balance code are needed to make the migrations happen. Misfit balancing only makes sense between a source group of lower per-CPU capacity and destination group of higher compute capacity. Otherwise, misfit balancing is ignored. group_misfit_task has lowest priority so any imbalance due to overload is dealt with first. The modifications are: 1. Only pick a group containing misfit tasks as the busiest group if the destination group has higher capacity and has spare capacity. 2. When the busiest group is a 'misfit' group, skip the usual average load and group capacity checks. 3. Set the imbalance for 'misfit' balancing sufficiently high for a task to be pulled ignoring average load. 4. Pick the CPU with the highest misfit load as the source CPU. 5. If the misfit task is alone on the source CPU, go for active balancing. Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: gaku.inami.xh@renesas.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1530699470-29808-5-git-send-email-morten.rasmussen@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-07-04 10:17:42 +00:00
return false;
sched/fair: Make update_sd_pick_busiest() return 'true' on a busier sd Currently update_sd_pick_busiest only identifies the busiest sd that is either overloaded, or has a group imbalance. When no sd is imbalanced or overloaded, the load balancer fails to find the busiest domain. This breaks load balancing between domains that are not overloaded, in the !SD_ASYM_PACKING case. This patch makes update_sd_pick_busiest return true when the busiest sd yet is encountered. Groups are ranked in the order overloaded > imbalanced > other, with higher ranked groups getting priority even when their load is lower. This is necessary due to the possibility of unequal capacities and cpumasks between domains within a sched group. Behaviour for SD_ASYM_PACKING does not seem to match the comment, but I have no hardware to test that so I have left the behaviour of that code unchanged. Enum for group classification suggested by Peter Zijlstra. Signed-off-by: Rik van Riel <riel@redhat.com> [peterz: replaced sg_lb_stats::group_imb with the new enum group_type in an attempt to avoid endless recalculation] Signed-off-by: Peter Zijlstra <peterz@infradead.org> Acked-by: Vincent Guittot <vincent.guittot@linaro.org> Acked-by: Michael Neuling <mikey@neuling.org> Cc: ktkhai@parallels.com Cc: tim.c.chen@linux.intel.com Cc: nicolas.pitre@linaro.org Cc: jhladky@redhat.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/20140729152743.GI3935@laptop Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-07-28 18:16:28 +00:00
if (sgs->group_type > busiest->group_type)
sched: Add asymmetric group packing option for sibling domain Check to see if the group is packed in a sched doman. This is primarily intended to used at the sibling level. Some cores like POWER7 prefer to use lower numbered SMT threads. In the case of POWER7, it can move to lower SMT modes only when higher threads are idle. When in lower SMT modes, the threads will perform better since they share less core resources. Hence when we have idle threads, we want them to be the higher ones. This adds a hook into f_b_g() called check_asym_packing() to check the packing. This packing function is run on idle threads. It checks to see if the busiest CPU in this domain (core in the P7 case) has a higher CPU number than what where the packing function is being run on. If it is, calculate the imbalance and return the higher busier thread as the busiest group to f_b_g(). Here we are assuming a lower CPU number will be equivalent to a lower SMT thread number. It also creates a new SD_ASYM_PACKING flag to enable this feature at any scheduler domain level. It also creates an arch hook to enable this feature at the sibling level. The default function doesn't enable this feature. Based heavily on patch from Peter Zijlstra. Fixes from Srivatsa Vaddagiri. Signed-off-by: Michael Neuling <mikey@neuling.org> Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Thomas Gleixner <tglx@linutronix.de> LKML-Reference: <20100608045702.2936CCC897@localhost.localdomain> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-06-08 04:57:02 +00:00
return true;
sched/fair: Make update_sd_pick_busiest() return 'true' on a busier sd Currently update_sd_pick_busiest only identifies the busiest sd that is either overloaded, or has a group imbalance. When no sd is imbalanced or overloaded, the load balancer fails to find the busiest domain. This breaks load balancing between domains that are not overloaded, in the !SD_ASYM_PACKING case. This patch makes update_sd_pick_busiest return true when the busiest sd yet is encountered. Groups are ranked in the order overloaded > imbalanced > other, with higher ranked groups getting priority even when their load is lower. This is necessary due to the possibility of unequal capacities and cpumasks between domains within a sched group. Behaviour for SD_ASYM_PACKING does not seem to match the comment, but I have no hardware to test that so I have left the behaviour of that code unchanged. Enum for group classification suggested by Peter Zijlstra. Signed-off-by: Rik van Riel <riel@redhat.com> [peterz: replaced sg_lb_stats::group_imb with the new enum group_type in an attempt to avoid endless recalculation] Signed-off-by: Peter Zijlstra <peterz@infradead.org> Acked-by: Vincent Guittot <vincent.guittot@linaro.org> Acked-by: Michael Neuling <mikey@neuling.org> Cc: ktkhai@parallels.com Cc: tim.c.chen@linux.intel.com Cc: nicolas.pitre@linaro.org Cc: jhladky@redhat.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/20140729152743.GI3935@laptop Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-07-28 18:16:28 +00:00
if (sgs->group_type < busiest->group_type)
return false;
/*
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
* The candidate and the current busiest group are the same type of
* group. Let check which one is the busiest according to the type.
*/
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
switch (sgs->group_type) {
case group_overloaded:
/* Select the overloaded group with highest avg_load. */
if (sgs->avg_load <= busiest->avg_load)
return false;
break;
case group_imbalanced:
/*
* Select the 1st imbalanced group as we don't have any way to
* choose one more than another.
*/
return false;
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
case group_asym_packing:
/* Prefer to move from lowest priority CPU's work */
if (sched_asym_prefer(sg->asym_prefer_cpu, sds->busiest->asym_prefer_cpu))
return false;
break;
sched: Add asymmetric group packing option for sibling domain Check to see if the group is packed in a sched doman. This is primarily intended to used at the sibling level. Some cores like POWER7 prefer to use lower numbered SMT threads. In the case of POWER7, it can move to lower SMT modes only when higher threads are idle. When in lower SMT modes, the threads will perform better since they share less core resources. Hence when we have idle threads, we want them to be the higher ones. This adds a hook into f_b_g() called check_asym_packing() to check the packing. This packing function is run on idle threads. It checks to see if the busiest CPU in this domain (core in the P7 case) has a higher CPU number than what where the packing function is being run on. If it is, calculate the imbalance and return the higher busier thread as the busiest group to f_b_g(). Here we are assuming a lower CPU number will be equivalent to a lower SMT thread number. It also creates a new SD_ASYM_PACKING flag to enable this feature at any scheduler domain level. It also creates an arch hook to enable this feature at the sibling level. The default function doesn't enable this feature. Based heavily on patch from Peter Zijlstra. Fixes from Srivatsa Vaddagiri. Signed-off-by: Michael Neuling <mikey@neuling.org> Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Thomas Gleixner <tglx@linutronix.de> LKML-Reference: <20100608045702.2936CCC897@localhost.localdomain> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-06-08 04:57:02 +00:00
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
case group_misfit_task:
/*
* If we have more than one misfit sg go with the biggest
* misfit.
*/
if (sgs->group_misfit_task_load < busiest->group_misfit_task_load)
return false;
break;
sched: Add asymmetric group packing option for sibling domain Check to see if the group is packed in a sched doman. This is primarily intended to used at the sibling level. Some cores like POWER7 prefer to use lower numbered SMT threads. In the case of POWER7, it can move to lower SMT modes only when higher threads are idle. When in lower SMT modes, the threads will perform better since they share less core resources. Hence when we have idle threads, we want them to be the higher ones. This adds a hook into f_b_g() called check_asym_packing() to check the packing. This packing function is run on idle threads. It checks to see if the busiest CPU in this domain (core in the P7 case) has a higher CPU number than what where the packing function is being run on. If it is, calculate the imbalance and return the higher busier thread as the busiest group to f_b_g(). Here we are assuming a lower CPU number will be equivalent to a lower SMT thread number. It also creates a new SD_ASYM_PACKING flag to enable this feature at any scheduler domain level. It also creates an arch hook to enable this feature at the sibling level. The default function doesn't enable this feature. Based heavily on patch from Peter Zijlstra. Fixes from Srivatsa Vaddagiri. Signed-off-by: Michael Neuling <mikey@neuling.org> Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Thomas Gleixner <tglx@linutronix.de> LKML-Reference: <20100608045702.2936CCC897@localhost.localdomain> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-06-08 04:57:02 +00:00
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
case group_fully_busy:
/*
* Select the fully busy group with highest avg_load. In
* theory, there is no need to pull task from such kind of
* group because tasks have all compute capacity that they need
* but we can still improve the overall throughput by reducing
* contention when accessing shared HW resources.
*
* XXX for now avg_load is not computed and always 0 so we
* select the 1st one.
*/
if (sgs->avg_load <= busiest->avg_load)
return false;
break;
case group_has_spare:
/*
* Select not overloaded group with lowest number of idle cpus
* and highest number of running tasks. We could also compare
* the spare capacity which is more stable but it can end up
* that the group has less spare capacity but finally more idle
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
* CPUs which means less opportunity to pull tasks.
*/
if (sgs->idle_cpus > busiest->idle_cpus)
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
return false;
else if ((sgs->idle_cpus == busiest->idle_cpus) &&
(sgs->sum_nr_running <= busiest->sum_nr_running))
return false;
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
break;
sched: Add asymmetric group packing option for sibling domain Check to see if the group is packed in a sched doman. This is primarily intended to used at the sibling level. Some cores like POWER7 prefer to use lower numbered SMT threads. In the case of POWER7, it can move to lower SMT modes only when higher threads are idle. When in lower SMT modes, the threads will perform better since they share less core resources. Hence when we have idle threads, we want them to be the higher ones. This adds a hook into f_b_g() called check_asym_packing() to check the packing. This packing function is run on idle threads. It checks to see if the busiest CPU in this domain (core in the P7 case) has a higher CPU number than what where the packing function is being run on. If it is, calculate the imbalance and return the higher busier thread as the busiest group to f_b_g(). Here we are assuming a lower CPU number will be equivalent to a lower SMT thread number. It also creates a new SD_ASYM_PACKING flag to enable this feature at any scheduler domain level. It also creates an arch hook to enable this feature at the sibling level. The default function doesn't enable this feature. Based heavily on patch from Peter Zijlstra. Fixes from Srivatsa Vaddagiri. Signed-off-by: Michael Neuling <mikey@neuling.org> Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Thomas Gleixner <tglx@linutronix.de> LKML-Reference: <20100608045702.2936CCC897@localhost.localdomain> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-06-08 04:57:02 +00:00
}
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
/*
* Candidate sg has no more than one task per CPU and has higher
* per-CPU capacity. Migrating tasks to less capable CPUs may harm
* throughput. Maximize throughput, power/energy consequences are not
* considered.
*/
if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
(sgs->group_type <= group_fully_busy) &&
(group_smaller_min_cpu_capacity(sds->local, sg)))
return false;
return true;
sched: Add asymmetric group packing option for sibling domain Check to see if the group is packed in a sched doman. This is primarily intended to used at the sibling level. Some cores like POWER7 prefer to use lower numbered SMT threads. In the case of POWER7, it can move to lower SMT modes only when higher threads are idle. When in lower SMT modes, the threads will perform better since they share less core resources. Hence when we have idle threads, we want them to be the higher ones. This adds a hook into f_b_g() called check_asym_packing() to check the packing. This packing function is run on idle threads. It checks to see if the busiest CPU in this domain (core in the P7 case) has a higher CPU number than what where the packing function is being run on. If it is, calculate the imbalance and return the higher busier thread as the busiest group to f_b_g(). Here we are assuming a lower CPU number will be equivalent to a lower SMT thread number. It also creates a new SD_ASYM_PACKING flag to enable this feature at any scheduler domain level. It also creates an arch hook to enable this feature at the sibling level. The default function doesn't enable this feature. Based heavily on patch from Peter Zijlstra. Fixes from Srivatsa Vaddagiri. Signed-off-by: Michael Neuling <mikey@neuling.org> Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Thomas Gleixner <tglx@linutronix.de> LKML-Reference: <20100608045702.2936CCC897@localhost.localdomain> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-06-08 04:57:02 +00:00
}
#ifdef CONFIG_NUMA_BALANCING
static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
{
if (sgs->sum_h_nr_running > sgs->nr_numa_running)
return regular;
if (sgs->sum_h_nr_running > sgs->nr_preferred_running)
return remote;
return all;
}
static inline enum fbq_type fbq_classify_rq(struct rq *rq)
{
if (rq->nr_running > rq->nr_numa_running)
return regular;
if (rq->nr_running > rq->nr_preferred_running)
return remote;
return all;
}
#else
static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
{
return all;
}
static inline enum fbq_type fbq_classify_rq(struct rq *rq)
{
return regular;
}
#endif /* CONFIG_NUMA_BALANCING */
struct sg_lb_stats;
sched/fair: Fix rework of find_idlest_group() The task, for which the scheduler looks for the idlest group of CPUs, must be discounted from all statistics in order to get a fair comparison between groups. This includes utilization, load, nr_running and idle_cpus. Such unfairness can be easily highlighted with the unixbench execl 1 task. This test continuously call execve() and the scheduler looks for the idlest group/CPU on which it should place the task. Because the task runs on the local group/CPU, the latter seems already busy even if there is nothing else running on it. As a result, the scheduler will always select another group/CPU than the local one. This recovers most of the performance regression on my system from the recent load-balancer rewrite. [ mingo: Minor cleanups. ] Reported-by: kernel test robot <rong.a.chen@intel.com> Tested-by: kernel test robot <rong.a.chen@intel.com> Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Fixes: 57abff067a08 ("sched/fair: Rework find_idlest_group()") Link: https://lkml.kernel.org/r/1571762798-25900-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-22 16:46:38 +00:00
/*
* task_running_on_cpu - return 1 if @p is running on @cpu.
*/
static unsigned int task_running_on_cpu(int cpu, struct task_struct *p)
{
/* Task has no contribution or is new */
if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
return 0;
if (task_on_rq_queued(p))
return 1;
return 0;
}
/**
* idle_cpu_without - would a given CPU be idle without p ?
* @cpu: the processor on which idleness is tested.
* @p: task which should be ignored.
*
* Return: 1 if the CPU would be idle. 0 otherwise.
*/
static int idle_cpu_without(int cpu, struct task_struct *p)
{
struct rq *rq = cpu_rq(cpu);
if (rq->curr != rq->idle && rq->curr != p)
return 0;
/*
* rq->nr_running can't be used but an updated version without the
* impact of p on cpu must be used instead. The updated nr_running
* be computed and tested before calling idle_cpu_without().
*/
#ifdef CONFIG_SMP
if (!llist_empty(&rq->wake_list))
return 0;
#endif
return 1;
}
/*
* update_sg_wakeup_stats - Update sched_group's statistics for wakeup.
sched/fair: Fix rework of find_idlest_group() The task, for which the scheduler looks for the idlest group of CPUs, must be discounted from all statistics in order to get a fair comparison between groups. This includes utilization, load, nr_running and idle_cpus. Such unfairness can be easily highlighted with the unixbench execl 1 task. This test continuously call execve() and the scheduler looks for the idlest group/CPU on which it should place the task. Because the task runs on the local group/CPU, the latter seems already busy even if there is nothing else running on it. As a result, the scheduler will always select another group/CPU than the local one. This recovers most of the performance regression on my system from the recent load-balancer rewrite. [ mingo: Minor cleanups. ] Reported-by: kernel test robot <rong.a.chen@intel.com> Tested-by: kernel test robot <rong.a.chen@intel.com> Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Fixes: 57abff067a08 ("sched/fair: Rework find_idlest_group()") Link: https://lkml.kernel.org/r/1571762798-25900-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-22 16:46:38 +00:00
* @sd: The sched_domain level to look for idlest group.
* @group: sched_group whose statistics are to be updated.
* @sgs: variable to hold the statistics for this group.
sched/fair: Fix rework of find_idlest_group() The task, for which the scheduler looks for the idlest group of CPUs, must be discounted from all statistics in order to get a fair comparison between groups. This includes utilization, load, nr_running and idle_cpus. Such unfairness can be easily highlighted with the unixbench execl 1 task. This test continuously call execve() and the scheduler looks for the idlest group/CPU on which it should place the task. Because the task runs on the local group/CPU, the latter seems already busy even if there is nothing else running on it. As a result, the scheduler will always select another group/CPU than the local one. This recovers most of the performance regression on my system from the recent load-balancer rewrite. [ mingo: Minor cleanups. ] Reported-by: kernel test robot <rong.a.chen@intel.com> Tested-by: kernel test robot <rong.a.chen@intel.com> Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Fixes: 57abff067a08 ("sched/fair: Rework find_idlest_group()") Link: https://lkml.kernel.org/r/1571762798-25900-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-22 16:46:38 +00:00
* @p: The task for which we look for the idlest group/CPU.
*/
static inline void update_sg_wakeup_stats(struct sched_domain *sd,
struct sched_group *group,
struct sg_lb_stats *sgs,
struct task_struct *p)
{
int i, nr_running;
memset(sgs, 0, sizeof(*sgs));
for_each_cpu(i, sched_group_span(group)) {
struct rq *rq = cpu_rq(i);
sched/fair: Fix rework of find_idlest_group() The task, for which the scheduler looks for the idlest group of CPUs, must be discounted from all statistics in order to get a fair comparison between groups. This includes utilization, load, nr_running and idle_cpus. Such unfairness can be easily highlighted with the unixbench execl 1 task. This test continuously call execve() and the scheduler looks for the idlest group/CPU on which it should place the task. Because the task runs on the local group/CPU, the latter seems already busy even if there is nothing else running on it. As a result, the scheduler will always select another group/CPU than the local one. This recovers most of the performance regression on my system from the recent load-balancer rewrite. [ mingo: Minor cleanups. ] Reported-by: kernel test robot <rong.a.chen@intel.com> Tested-by: kernel test robot <rong.a.chen@intel.com> Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Fixes: 57abff067a08 ("sched/fair: Rework find_idlest_group()") Link: https://lkml.kernel.org/r/1571762798-25900-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-22 16:46:38 +00:00
unsigned int local;
sched/fair: Fix rework of find_idlest_group() The task, for which the scheduler looks for the idlest group of CPUs, must be discounted from all statistics in order to get a fair comparison between groups. This includes utilization, load, nr_running and idle_cpus. Such unfairness can be easily highlighted with the unixbench execl 1 task. This test continuously call execve() and the scheduler looks for the idlest group/CPU on which it should place the task. Because the task runs on the local group/CPU, the latter seems already busy even if there is nothing else running on it. As a result, the scheduler will always select another group/CPU than the local one. This recovers most of the performance regression on my system from the recent load-balancer rewrite. [ mingo: Minor cleanups. ] Reported-by: kernel test robot <rong.a.chen@intel.com> Tested-by: kernel test robot <rong.a.chen@intel.com> Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Fixes: 57abff067a08 ("sched/fair: Rework find_idlest_group()") Link: https://lkml.kernel.org/r/1571762798-25900-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-22 16:46:38 +00:00
sgs->group_load += cpu_load_without(rq, p);
sgs->group_util += cpu_util_without(i, p);
sgs->group_runnable += cpu_runnable_without(rq, p);
sched/fair: Fix rework of find_idlest_group() The task, for which the scheduler looks for the idlest group of CPUs, must be discounted from all statistics in order to get a fair comparison between groups. This includes utilization, load, nr_running and idle_cpus. Such unfairness can be easily highlighted with the unixbench execl 1 task. This test continuously call execve() and the scheduler looks for the idlest group/CPU on which it should place the task. Because the task runs on the local group/CPU, the latter seems already busy even if there is nothing else running on it. As a result, the scheduler will always select another group/CPU than the local one. This recovers most of the performance regression on my system from the recent load-balancer rewrite. [ mingo: Minor cleanups. ] Reported-by: kernel test robot <rong.a.chen@intel.com> Tested-by: kernel test robot <rong.a.chen@intel.com> Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Fixes: 57abff067a08 ("sched/fair: Rework find_idlest_group()") Link: https://lkml.kernel.org/r/1571762798-25900-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-22 16:46:38 +00:00
local = task_running_on_cpu(i, p);
sgs->sum_h_nr_running += rq->cfs.h_nr_running - local;
sched/fair: Fix rework of find_idlest_group() The task, for which the scheduler looks for the idlest group of CPUs, must be discounted from all statistics in order to get a fair comparison between groups. This includes utilization, load, nr_running and idle_cpus. Such unfairness can be easily highlighted with the unixbench execl 1 task. This test continuously call execve() and the scheduler looks for the idlest group/CPU on which it should place the task. Because the task runs on the local group/CPU, the latter seems already busy even if there is nothing else running on it. As a result, the scheduler will always select another group/CPU than the local one. This recovers most of the performance regression on my system from the recent load-balancer rewrite. [ mingo: Minor cleanups. ] Reported-by: kernel test robot <rong.a.chen@intel.com> Tested-by: kernel test robot <rong.a.chen@intel.com> Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Fixes: 57abff067a08 ("sched/fair: Rework find_idlest_group()") Link: https://lkml.kernel.org/r/1571762798-25900-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-22 16:46:38 +00:00
nr_running = rq->nr_running - local;
sgs->sum_nr_running += nr_running;
/*
sched/fair: Fix rework of find_idlest_group() The task, for which the scheduler looks for the idlest group of CPUs, must be discounted from all statistics in order to get a fair comparison between groups. This includes utilization, load, nr_running and idle_cpus. Such unfairness can be easily highlighted with the unixbench execl 1 task. This test continuously call execve() and the scheduler looks for the idlest group/CPU on which it should place the task. Because the task runs on the local group/CPU, the latter seems already busy even if there is nothing else running on it. As a result, the scheduler will always select another group/CPU than the local one. This recovers most of the performance regression on my system from the recent load-balancer rewrite. [ mingo: Minor cleanups. ] Reported-by: kernel test robot <rong.a.chen@intel.com> Tested-by: kernel test robot <rong.a.chen@intel.com> Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Fixes: 57abff067a08 ("sched/fair: Rework find_idlest_group()") Link: https://lkml.kernel.org/r/1571762798-25900-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-22 16:46:38 +00:00
* No need to call idle_cpu_without() if nr_running is not 0
*/
sched/fair: Fix rework of find_idlest_group() The task, for which the scheduler looks for the idlest group of CPUs, must be discounted from all statistics in order to get a fair comparison between groups. This includes utilization, load, nr_running and idle_cpus. Such unfairness can be easily highlighted with the unixbench execl 1 task. This test continuously call execve() and the scheduler looks for the idlest group/CPU on which it should place the task. Because the task runs on the local group/CPU, the latter seems already busy even if there is nothing else running on it. As a result, the scheduler will always select another group/CPU than the local one. This recovers most of the performance regression on my system from the recent load-balancer rewrite. [ mingo: Minor cleanups. ] Reported-by: kernel test robot <rong.a.chen@intel.com> Tested-by: kernel test robot <rong.a.chen@intel.com> Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Fixes: 57abff067a08 ("sched/fair: Rework find_idlest_group()") Link: https://lkml.kernel.org/r/1571762798-25900-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-22 16:46:38 +00:00
if (!nr_running && idle_cpu_without(i, p))
sgs->idle_cpus++;
}
/* Check if task fits in the group */
if (sd->flags & SD_ASYM_CPUCAPACITY &&
!task_fits_capacity(p, group->sgc->max_capacity)) {
sgs->group_misfit_task_load = 1;
}
sgs->group_capacity = group->sgc->capacity;
sgs->group_weight = group->group_weight;
sgs->group_type = group_classify(sd->imbalance_pct, group, sgs);
/*
* Computing avg_load makes sense only when group is fully busy or
* overloaded
*/
if (sgs->group_type == group_fully_busy ||
sgs->group_type == group_overloaded)
sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
sgs->group_capacity;
}
static bool update_pick_idlest(struct sched_group *idlest,
struct sg_lb_stats *idlest_sgs,
struct sched_group *group,
struct sg_lb_stats *sgs)
{
if (sgs->group_type < idlest_sgs->group_type)
return true;
if (sgs->group_type > idlest_sgs->group_type)
return false;
/*
* The candidate and the current idlest group are the same type of
* group. Let check which one is the idlest according to the type.
*/
switch (sgs->group_type) {
case group_overloaded:
case group_fully_busy:
/* Select the group with lowest avg_load. */
if (idlest_sgs->avg_load <= sgs->avg_load)
return false;
break;
case group_imbalanced:
case group_asym_packing:
/* Those types are not used in the slow wakeup path */
return false;
case group_misfit_task:
/* Select group with the highest max capacity */
if (idlest->sgc->max_capacity >= group->sgc->max_capacity)
return false;
break;
case group_has_spare:
/* Select group with most idle CPUs */
if (idlest_sgs->idle_cpus >= sgs->idle_cpus)
return false;
break;
}
return true;
}
/*
* find_idlest_group() finds and returns the least busy CPU group within the
* domain.
*
* Assumes p is allowed on at least one CPU in sd.
*/
static struct sched_group *
find_idlest_group(struct sched_domain *sd, struct task_struct *p,
int this_cpu, int sd_flag)
{
struct sched_group *idlest = NULL, *local = NULL, *group = sd->groups;
struct sg_lb_stats local_sgs, tmp_sgs;
struct sg_lb_stats *sgs;
unsigned long imbalance;
struct sg_lb_stats idlest_sgs = {
.avg_load = UINT_MAX,
.group_type = group_overloaded,
};
imbalance = scale_load_down(NICE_0_LOAD) *
(sd->imbalance_pct-100) / 100;
do {
int local_group;
/* Skip over this group if it has no CPUs allowed */
if (!cpumask_intersects(sched_group_span(group),
p->cpus_ptr))
continue;
local_group = cpumask_test_cpu(this_cpu,
sched_group_span(group));
if (local_group) {
sgs = &local_sgs;
local = group;
} else {
sgs = &tmp_sgs;
}
update_sg_wakeup_stats(sd, group, sgs, p);
if (!local_group && update_pick_idlest(idlest, &idlest_sgs, group, sgs)) {
idlest = group;
idlest_sgs = *sgs;
}
} while (group = group->next, group != sd->groups);
/* There is no idlest group to push tasks to */
if (!idlest)
return NULL;
/* The local group has been skipped because of CPU affinity */
if (!local)
return idlest;
/*
* If the local group is idler than the selected idlest group
* don't try and push the task.
*/
if (local_sgs.group_type < idlest_sgs.group_type)
return NULL;
/*
* If the local group is busier than the selected idlest group
* try and push the task.
*/
if (local_sgs.group_type > idlest_sgs.group_type)
return idlest;
switch (local_sgs.group_type) {
case group_overloaded:
case group_fully_busy:
/*
* When comparing groups across NUMA domains, it's possible for
* the local domain to be very lightly loaded relative to the
* remote domains but "imbalance" skews the comparison making
* remote CPUs look much more favourable. When considering
* cross-domain, add imbalance to the load on the remote node
* and consider staying local.
*/
if ((sd->flags & SD_NUMA) &&
((idlest_sgs.avg_load + imbalance) >= local_sgs.avg_load))
return NULL;
/*
* If the local group is less loaded than the selected
* idlest group don't try and push any tasks.
*/
if (idlest_sgs.avg_load >= (local_sgs.avg_load + imbalance))
return NULL;
if (100 * local_sgs.avg_load <= sd->imbalance_pct * idlest_sgs.avg_load)
return NULL;
break;
case group_imbalanced:
case group_asym_packing:
/* Those type are not used in the slow wakeup path */
return NULL;
case group_misfit_task:
/* Select group with the highest max capacity */
if (local->sgc->max_capacity >= idlest->sgc->max_capacity)
return NULL;
break;
case group_has_spare:
if (sd->flags & SD_NUMA) {
#ifdef CONFIG_NUMA_BALANCING
int idlest_cpu;
/*
* If there is spare capacity at NUMA, try to select
* the preferred node
*/
if (cpu_to_node(this_cpu) == p->numa_preferred_nid)
return NULL;
idlest_cpu = cpumask_first(sched_group_span(idlest));
if (cpu_to_node(idlest_cpu) == p->numa_preferred_nid)
return idlest;
#endif
/*
* Otherwise, keep the task on this node to stay close
* its wakeup source and improve locality. If there is
* a real need of migration, periodic load balance will
* take care of it.
*/
if (local_sgs.idle_cpus)
return NULL;
}
/*
* Select group with highest number of idle CPUs. We could also
* compare the utilization which is more stable but it can end
* up that the group has less spare capacity but finally more
* idle CPUs which means more opportunity to run task.
*/
if (local_sgs.idle_cpus >= idlest_sgs.idle_cpus)
return NULL;
break;
}
return idlest;
}
/**
* update_sd_lb_stats - Update sched_domain's statistics for load balancing.
* @env: The load balancing environment.
* @sds: variable to hold the statistics for this sched_domain.
*/
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
{
struct sched_domain *child = env->sd->child;
struct sched_group *sg = env->sd->groups;
sched/fair: Prefer sibiling only if local group is under-utilized If the child domain prefers tasks to go siblings, the local group could end up pulling tasks to itself even if the local group is almost equally loaded as the source group. Lets assume a 4 core,smt==2 machine running 5 thread ebizzy workload. Everytime, local group has capacity and source group has atleast 2 threads, local group tries to pull the task. This causes the threads to constantly move between different cores. This is even more profound if the cores have more threads, like in Power 8, smt 8 mode. Fix this by only allowing local group to pull a task, if the source group has more number of tasks than the local group. Here are the relevant perf stat numbers of a 22 core,smt 8 Power 8 machine. Without patch: Performance counter stats for 'ebizzy -t 22 -S 100' (5 runs): 1,440 context-switches # 0.001 K/sec ( +- 1.26% ) 366 cpu-migrations # 0.000 K/sec ( +- 5.58% ) 3,933 page-faults # 0.002 K/sec ( +- 11.08% ) Performance counter stats for 'ebizzy -t 48 -S 100' (5 runs): 6,287 context-switches # 0.001 K/sec ( +- 3.65% ) 3,776 cpu-migrations # 0.001 K/sec ( +- 4.84% ) 5,702 page-faults # 0.001 K/sec ( +- 9.36% ) Performance counter stats for 'ebizzy -t 96 -S 100' (5 runs): 8,776 context-switches # 0.001 K/sec ( +- 0.73% ) 2,790 cpu-migrations # 0.000 K/sec ( +- 0.98% ) 10,540 page-faults # 0.001 K/sec ( +- 3.12% ) With patch: Performance counter stats for 'ebizzy -t 22 -S 100' (5 runs): 1,133 context-switches # 0.001 K/sec ( +- 4.72% ) 123 cpu-migrations # 0.000 K/sec ( +- 3.42% ) 3,858 page-faults # 0.002 K/sec ( +- 8.52% ) Performance counter stats for 'ebizzy -t 48 -S 100' (5 runs): 2,169 context-switches # 0.000 K/sec ( +- 6.19% ) 189 cpu-migrations # 0.000 K/sec ( +- 12.75% ) 5,917 page-faults # 0.001 K/sec ( +- 8.09% ) Performance counter stats for 'ebizzy -t 96 -S 100' (5 runs): 5,333 context-switches # 0.001 K/sec ( +- 5.91% ) 506 cpu-migrations # 0.000 K/sec ( +- 3.35% ) 10,792 page-faults # 0.001 K/sec ( +- 7.75% ) Which show that in these workloads CPU migrations get reduced significantly. Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vincent Guittot <vincent.guittot@linaro.org> Link: http://lkml.kernel.org/r/1490205470-10249-1-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-03-22 17:57:50 +00:00
struct sg_lb_stats *local = &sds->local_stat;
struct sg_lb_stats tmp_sgs;
sched/fair: Clean-up update_sg_lb_stats parameters In preparation for the introduction of a new root domain flag which can be set during load balance (the 'overutilized' flag), clean-up the set of parameters passed to update_sg_lb_stats(). More specifically, the 'local_group' and 'local_idx' parameters can be removed since they can easily be reconstructed from within the function. While at it, transform the 'overload' parameter into a flag stored in the 'sg_status' parameter hence facilitating the definition of new flags when needed. Suggested-by: Peter Zijlstra <peterz@infradead.org> Suggested-by: Valentin Schneider <valentin.schneider@arm.com> Signed-off-by: Quentin Perret <quentin.perret@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: adharmap@codeaurora.org Cc: chris.redpath@arm.com Cc: currojerez@riseup.net Cc: dietmar.eggemann@arm.com Cc: edubezval@gmail.com Cc: gregkh@linuxfoundation.org Cc: javi.merino@kernel.org Cc: joel@joelfernandes.org Cc: juri.lelli@redhat.com Cc: morten.rasmussen@arm.com Cc: patrick.bellasi@arm.com Cc: pkondeti@codeaurora.org Cc: rjw@rjwysocki.net Cc: skannan@codeaurora.org Cc: smuckle@google.com Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Cc: tkjos@google.com Cc: vincent.guittot@linaro.org Cc: viresh.kumar@linaro.org Link: https://lkml.kernel.org/r/20181203095628.11858-12-quentin.perret@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-12-03 09:56:24 +00:00
int sg_status = 0;
#ifdef CONFIG_NO_HZ_COMMON
if (env->idle == CPU_NEWLY_IDLE && READ_ONCE(nohz.has_blocked))
env->flags |= LBF_NOHZ_STATS;
#endif
do {
struct sg_lb_stats *sgs = &tmp_sgs;
int local_group;
local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
if (local_group) {
sds->local = sg;
sched/fair: Prefer sibiling only if local group is under-utilized If the child domain prefers tasks to go siblings, the local group could end up pulling tasks to itself even if the local group is almost equally loaded as the source group. Lets assume a 4 core,smt==2 machine running 5 thread ebizzy workload. Everytime, local group has capacity and source group has atleast 2 threads, local group tries to pull the task. This causes the threads to constantly move between different cores. This is even more profound if the cores have more threads, like in Power 8, smt 8 mode. Fix this by only allowing local group to pull a task, if the source group has more number of tasks than the local group. Here are the relevant perf stat numbers of a 22 core,smt 8 Power 8 machine. Without patch: Performance counter stats for 'ebizzy -t 22 -S 100' (5 runs): 1,440 context-switches # 0.001 K/sec ( +- 1.26% ) 366 cpu-migrations # 0.000 K/sec ( +- 5.58% ) 3,933 page-faults # 0.002 K/sec ( +- 11.08% ) Performance counter stats for 'ebizzy -t 48 -S 100' (5 runs): 6,287 context-switches # 0.001 K/sec ( +- 3.65% ) 3,776 cpu-migrations # 0.001 K/sec ( +- 4.84% ) 5,702 page-faults # 0.001 K/sec ( +- 9.36% ) Performance counter stats for 'ebizzy -t 96 -S 100' (5 runs): 8,776 context-switches # 0.001 K/sec ( +- 0.73% ) 2,790 cpu-migrations # 0.000 K/sec ( +- 0.98% ) 10,540 page-faults # 0.001 K/sec ( +- 3.12% ) With patch: Performance counter stats for 'ebizzy -t 22 -S 100' (5 runs): 1,133 context-switches # 0.001 K/sec ( +- 4.72% ) 123 cpu-migrations # 0.000 K/sec ( +- 3.42% ) 3,858 page-faults # 0.002 K/sec ( +- 8.52% ) Performance counter stats for 'ebizzy -t 48 -S 100' (5 runs): 2,169 context-switches # 0.000 K/sec ( +- 6.19% ) 189 cpu-migrations # 0.000 K/sec ( +- 12.75% ) 5,917 page-faults # 0.001 K/sec ( +- 8.09% ) Performance counter stats for 'ebizzy -t 96 -S 100' (5 runs): 5,333 context-switches # 0.001 K/sec ( +- 5.91% ) 506 cpu-migrations # 0.000 K/sec ( +- 3.35% ) 10,792 page-faults # 0.001 K/sec ( +- 7.75% ) Which show that in these workloads CPU migrations get reduced significantly. Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vincent Guittot <vincent.guittot@linaro.org> Link: http://lkml.kernel.org/r/1490205470-10249-1-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-03-22 17:57:50 +00:00
sgs = local;
if (env->idle != CPU_NEWLY_IDLE ||
time_after_eq(jiffies, sg->sgc->next_update))
update_group_capacity(env->sd, env->dst_cpu);
}
sched/fair: Clean-up update_sg_lb_stats parameters In preparation for the introduction of a new root domain flag which can be set during load balance (the 'overutilized' flag), clean-up the set of parameters passed to update_sg_lb_stats(). More specifically, the 'local_group' and 'local_idx' parameters can be removed since they can easily be reconstructed from within the function. While at it, transform the 'overload' parameter into a flag stored in the 'sg_status' parameter hence facilitating the definition of new flags when needed. Suggested-by: Peter Zijlstra <peterz@infradead.org> Suggested-by: Valentin Schneider <valentin.schneider@arm.com> Signed-off-by: Quentin Perret <quentin.perret@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: adharmap@codeaurora.org Cc: chris.redpath@arm.com Cc: currojerez@riseup.net Cc: dietmar.eggemann@arm.com Cc: edubezval@gmail.com Cc: gregkh@linuxfoundation.org Cc: javi.merino@kernel.org Cc: joel@joelfernandes.org Cc: juri.lelli@redhat.com Cc: morten.rasmussen@arm.com Cc: patrick.bellasi@arm.com Cc: pkondeti@codeaurora.org Cc: rjw@rjwysocki.net Cc: skannan@codeaurora.org Cc: smuckle@google.com Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Cc: tkjos@google.com Cc: vincent.guittot@linaro.org Cc: viresh.kumar@linaro.org Link: https://lkml.kernel.org/r/20181203095628.11858-12-quentin.perret@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-12-03 09:56:24 +00:00
update_sg_lb_stats(env, sg, sgs, &sg_status);
if (local_group)
goto next_group;
if (update_sd_pick_busiest(env, sds, sg, sgs)) {
sched: Add asymmetric group packing option for sibling domain Check to see if the group is packed in a sched doman. This is primarily intended to used at the sibling level. Some cores like POWER7 prefer to use lower numbered SMT threads. In the case of POWER7, it can move to lower SMT modes only when higher threads are idle. When in lower SMT modes, the threads will perform better since they share less core resources. Hence when we have idle threads, we want them to be the higher ones. This adds a hook into f_b_g() called check_asym_packing() to check the packing. This packing function is run on idle threads. It checks to see if the busiest CPU in this domain (core in the P7 case) has a higher CPU number than what where the packing function is being run on. If it is, calculate the imbalance and return the higher busier thread as the busiest group to f_b_g(). Here we are assuming a lower CPU number will be equivalent to a lower SMT thread number. It also creates a new SD_ASYM_PACKING flag to enable this feature at any scheduler domain level. It also creates an arch hook to enable this feature at the sibling level. The default function doesn't enable this feature. Based heavily on patch from Peter Zijlstra. Fixes from Srivatsa Vaddagiri. Signed-off-by: Michael Neuling <mikey@neuling.org> Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Thomas Gleixner <tglx@linutronix.de> LKML-Reference: <20100608045702.2936CCC897@localhost.localdomain> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-06-08 04:57:02 +00:00
sds->busiest = sg;
sds->busiest_stat = *sgs;
}
next_group:
/* Now, start updating sd_lb_stats */
sds->total_load += sgs->group_load;
sds->total_capacity += sgs->group_capacity;
sched: Add asymmetric group packing option for sibling domain Check to see if the group is packed in a sched doman. This is primarily intended to used at the sibling level. Some cores like POWER7 prefer to use lower numbered SMT threads. In the case of POWER7, it can move to lower SMT modes only when higher threads are idle. When in lower SMT modes, the threads will perform better since they share less core resources. Hence when we have idle threads, we want them to be the higher ones. This adds a hook into f_b_g() called check_asym_packing() to check the packing. This packing function is run on idle threads. It checks to see if the busiest CPU in this domain (core in the P7 case) has a higher CPU number than what where the packing function is being run on. If it is, calculate the imbalance and return the higher busier thread as the busiest group to f_b_g(). Here we are assuming a lower CPU number will be equivalent to a lower SMT thread number. It also creates a new SD_ASYM_PACKING flag to enable this feature at any scheduler domain level. It also creates an arch hook to enable this feature at the sibling level. The default function doesn't enable this feature. Based heavily on patch from Peter Zijlstra. Fixes from Srivatsa Vaddagiri. Signed-off-by: Michael Neuling <mikey@neuling.org> Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Thomas Gleixner <tglx@linutronix.de> LKML-Reference: <20100608045702.2936CCC897@localhost.localdomain> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-06-08 04:57:02 +00:00
sg = sg->next;
} while (sg != env->sd->groups);
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
/* Tag domain that child domain prefers tasks go to siblings first */
sds->prefer_sibling = child && child->flags & SD_PREFER_SIBLING;
#ifdef CONFIG_NO_HZ_COMMON
if ((env->flags & LBF_NOHZ_AGAIN) &&
cpumask_subset(nohz.idle_cpus_mask, sched_domain_span(env->sd))) {
WRITE_ONCE(nohz.next_blocked,
jiffies + msecs_to_jiffies(LOAD_AVG_PERIOD));
}
#endif
if (env->sd->flags & SD_NUMA)
env->fbq_type = fbq_classify_group(&sds->busiest_stat);
sched/fair: Implement fast idling of CPUs when the system is partially loaded When a system is lightly loaded (i.e. no more than 1 job per cpu), attempt to pull job to a cpu before putting it to idle is unnecessary and can be skipped. This patch adds an indicator so the scheduler can know when there's no more than 1 active job is on any CPU in the system to skip needless job pulls. On a 4 socket machine with a request/response kind of workload from clients, we saw about 0.13 msec delay when we go through a full load balance to try pull job from all the other cpus. While 0.1 msec was spent on processing the request and generating a response, the 0.13 msec load balance overhead was actually more than the actual work being done. This overhead can be skipped much of the time for lightly loaded systems. With this patch, we tested with a netperf request/response workload that has the server busy with half the cpus in a 4 socket system. We found the patch eliminated 75% of the load balance attempts before idling a cpu. The overhead of setting/clearing the indicator is low as we already gather the necessary info while we call add_nr_running() and update_sd_lb_stats.() We switch to full load balance load immediately if any cpu got more than one job on its run queue in add_nr_running. We'll clear the indicator to avoid load balance when we detect no cpu's have more than one job when we scan the work queues in update_sg_lb_stats(). We are aggressive in turning on the load balance and opportunistic in skipping the load balance. Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Acked-by: Rik van Riel <riel@redhat.com> Acked-by: Jason Low <jason.low2@hp.com> Cc: "Paul E.McKenney" <paulmck@linux.vnet.ibm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Davidlohr Bueso <davidlohr@hp.com> Cc: Alex Shi <alex.shi@linaro.org> Cc: Michel Lespinasse <walken@google.com> Cc: Peter Hurley <peter@hurleysoftware.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1403551009.2970.613.camel@schen9-DESK Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-06-23 19:16:49 +00:00
if (!env->sd->parent) {
sched/fair: Add over-utilization/tipping point indicator Energy-aware scheduling is only meant to be active while the system is _not_ over-utilized. That is, there are spare cycles available to shift tasks around based on their actual utilization to get a more energy-efficient task distribution without depriving any tasks. When above the tipping point task placement is done the traditional way based on load_avg, spreading the tasks across as many cpus as possible based on priority scaled load to preserve smp_nice. Below the tipping point we want to use util_avg instead. We need to define a criteria for when we make the switch. The util_avg for each cpu converges towards 100% regardless of how many additional tasks we may put on it. If we define over-utilized as: sum_{cpus}(rq.cfs.avg.util_avg) + margin > sum_{cpus}(rq.capacity) some individual cpus may be over-utilized running multiple tasks even when the above condition is false. That should be okay as long as we try to spread the tasks out to avoid per-cpu over-utilization as much as possible and if all tasks have the _same_ priority. If the latter isn't true, we have to consider priority to preserve smp_nice. For example, we could have n_cpus nice=-10 util_avg=55% tasks and n_cpus/2 nice=0 util_avg=60% tasks. Balancing based on util_avg we are likely to end up with nice=-10 tasks sharing cpus and nice=0 tasks getting their own as we 1.5*n_cpus tasks in total and 55%+55% is less over-utilized than 55%+60% for those cpus that have to be shared. The system utilization is only 85% of the system capacity, but we are breaking smp_nice. To be sure not to break smp_nice, we have defined over-utilization conservatively as when any cpu in the system is fully utilized at its highest frequency instead: cpu_rq(any).cfs.avg.util_avg + margin > cpu_rq(any).capacity IOW, as soon as one cpu is (nearly) 100% utilized, we switch to load_avg to factor in priority to preserve smp_nice. With this definition, we can skip periodic load-balance as no cpu has an always-running task when the system is not over-utilized. All tasks will be periodic and we can balance them at wake-up. This conservative condition does however mean that some scenarios that could benefit from energy-aware decisions even if one cpu is fully utilized would not get those benefits. For systems where some cpus might have reduced capacity on some cpus (RT-pressure and/or big.LITTLE), we want periodic load-balance checks as soon a just a single cpu is fully utilized as it might one of those with reduced capacity and in that case we want to migrate it. [ peterz: Added a comment explaining why new tasks are not accounted during overutilization detection. ] Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Quentin Perret <quentin.perret@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: adharmap@codeaurora.org Cc: chris.redpath@arm.com Cc: currojerez@riseup.net Cc: dietmar.eggemann@arm.com Cc: edubezval@gmail.com Cc: gregkh@linuxfoundation.org Cc: javi.merino@kernel.org Cc: joel@joelfernandes.org Cc: juri.lelli@redhat.com Cc: patrick.bellasi@arm.com Cc: pkondeti@codeaurora.org Cc: rjw@rjwysocki.net Cc: skannan@codeaurora.org Cc: smuckle@google.com Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Cc: viresh.kumar@linaro.org Link: https://lkml.kernel.org/r/20181203095628.11858-13-quentin.perret@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-12-03 09:56:25 +00:00
struct root_domain *rd = env->dst_rq->rd;
sched/fair: Implement fast idling of CPUs when the system is partially loaded When a system is lightly loaded (i.e. no more than 1 job per cpu), attempt to pull job to a cpu before putting it to idle is unnecessary and can be skipped. This patch adds an indicator so the scheduler can know when there's no more than 1 active job is on any CPU in the system to skip needless job pulls. On a 4 socket machine with a request/response kind of workload from clients, we saw about 0.13 msec delay when we go through a full load balance to try pull job from all the other cpus. While 0.1 msec was spent on processing the request and generating a response, the 0.13 msec load balance overhead was actually more than the actual work being done. This overhead can be skipped much of the time for lightly loaded systems. With this patch, we tested with a netperf request/response workload that has the server busy with half the cpus in a 4 socket system. We found the patch eliminated 75% of the load balance attempts before idling a cpu. The overhead of setting/clearing the indicator is low as we already gather the necessary info while we call add_nr_running() and update_sd_lb_stats.() We switch to full load balance load immediately if any cpu got more than one job on its run queue in add_nr_running. We'll clear the indicator to avoid load balance when we detect no cpu's have more than one job when we scan the work queues in update_sg_lb_stats(). We are aggressive in turning on the load balance and opportunistic in skipping the load balance. Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Acked-by: Rik van Riel <riel@redhat.com> Acked-by: Jason Low <jason.low2@hp.com> Cc: "Paul E.McKenney" <paulmck@linux.vnet.ibm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Davidlohr Bueso <davidlohr@hp.com> Cc: Alex Shi <alex.shi@linaro.org> Cc: Michel Lespinasse <walken@google.com> Cc: Peter Hurley <peter@hurleysoftware.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1403551009.2970.613.camel@schen9-DESK Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-06-23 19:16:49 +00:00
/* update overload indicator if we are at root domain */
sched/fair: Add over-utilization/tipping point indicator Energy-aware scheduling is only meant to be active while the system is _not_ over-utilized. That is, there are spare cycles available to shift tasks around based on their actual utilization to get a more energy-efficient task distribution without depriving any tasks. When above the tipping point task placement is done the traditional way based on load_avg, spreading the tasks across as many cpus as possible based on priority scaled load to preserve smp_nice. Below the tipping point we want to use util_avg instead. We need to define a criteria for when we make the switch. The util_avg for each cpu converges towards 100% regardless of how many additional tasks we may put on it. If we define over-utilized as: sum_{cpus}(rq.cfs.avg.util_avg) + margin > sum_{cpus}(rq.capacity) some individual cpus may be over-utilized running multiple tasks even when the above condition is false. That should be okay as long as we try to spread the tasks out to avoid per-cpu over-utilization as much as possible and if all tasks have the _same_ priority. If the latter isn't true, we have to consider priority to preserve smp_nice. For example, we could have n_cpus nice=-10 util_avg=55% tasks and n_cpus/2 nice=0 util_avg=60% tasks. Balancing based on util_avg we are likely to end up with nice=-10 tasks sharing cpus and nice=0 tasks getting their own as we 1.5*n_cpus tasks in total and 55%+55% is less over-utilized than 55%+60% for those cpus that have to be shared. The system utilization is only 85% of the system capacity, but we are breaking smp_nice. To be sure not to break smp_nice, we have defined over-utilization conservatively as when any cpu in the system is fully utilized at its highest frequency instead: cpu_rq(any).cfs.avg.util_avg + margin > cpu_rq(any).capacity IOW, as soon as one cpu is (nearly) 100% utilized, we switch to load_avg to factor in priority to preserve smp_nice. With this definition, we can skip periodic load-balance as no cpu has an always-running task when the system is not over-utilized. All tasks will be periodic and we can balance them at wake-up. This conservative condition does however mean that some scenarios that could benefit from energy-aware decisions even if one cpu is fully utilized would not get those benefits. For systems where some cpus might have reduced capacity on some cpus (RT-pressure and/or big.LITTLE), we want periodic load-balance checks as soon a just a single cpu is fully utilized as it might one of those with reduced capacity and in that case we want to migrate it. [ peterz: Added a comment explaining why new tasks are not accounted during overutilization detection. ] Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Quentin Perret <quentin.perret@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: adharmap@codeaurora.org Cc: chris.redpath@arm.com Cc: currojerez@riseup.net Cc: dietmar.eggemann@arm.com Cc: edubezval@gmail.com Cc: gregkh@linuxfoundation.org Cc: javi.merino@kernel.org Cc: joel@joelfernandes.org Cc: juri.lelli@redhat.com Cc: patrick.bellasi@arm.com Cc: pkondeti@codeaurora.org Cc: rjw@rjwysocki.net Cc: skannan@codeaurora.org Cc: smuckle@google.com Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Cc: viresh.kumar@linaro.org Link: https://lkml.kernel.org/r/20181203095628.11858-13-quentin.perret@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-12-03 09:56:25 +00:00
WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD);
/* Update over-utilization (tipping point, U >= 0) indicator */
WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED);
trace_sched_overutilized_tp(rd, sg_status & SG_OVERUTILIZED);
sched/fair: Add over-utilization/tipping point indicator Energy-aware scheduling is only meant to be active while the system is _not_ over-utilized. That is, there are spare cycles available to shift tasks around based on their actual utilization to get a more energy-efficient task distribution without depriving any tasks. When above the tipping point task placement is done the traditional way based on load_avg, spreading the tasks across as many cpus as possible based on priority scaled load to preserve smp_nice. Below the tipping point we want to use util_avg instead. We need to define a criteria for when we make the switch. The util_avg for each cpu converges towards 100% regardless of how many additional tasks we may put on it. If we define over-utilized as: sum_{cpus}(rq.cfs.avg.util_avg) + margin > sum_{cpus}(rq.capacity) some individual cpus may be over-utilized running multiple tasks even when the above condition is false. That should be okay as long as we try to spread the tasks out to avoid per-cpu over-utilization as much as possible and if all tasks have the _same_ priority. If the latter isn't true, we have to consider priority to preserve smp_nice. For example, we could have n_cpus nice=-10 util_avg=55% tasks and n_cpus/2 nice=0 util_avg=60% tasks. Balancing based on util_avg we are likely to end up with nice=-10 tasks sharing cpus and nice=0 tasks getting their own as we 1.5*n_cpus tasks in total and 55%+55% is less over-utilized than 55%+60% for those cpus that have to be shared. The system utilization is only 85% of the system capacity, but we are breaking smp_nice. To be sure not to break smp_nice, we have defined over-utilization conservatively as when any cpu in the system is fully utilized at its highest frequency instead: cpu_rq(any).cfs.avg.util_avg + margin > cpu_rq(any).capacity IOW, as soon as one cpu is (nearly) 100% utilized, we switch to load_avg to factor in priority to preserve smp_nice. With this definition, we can skip periodic load-balance as no cpu has an always-running task when the system is not over-utilized. All tasks will be periodic and we can balance them at wake-up. This conservative condition does however mean that some scenarios that could benefit from energy-aware decisions even if one cpu is fully utilized would not get those benefits. For systems where some cpus might have reduced capacity on some cpus (RT-pressure and/or big.LITTLE), we want periodic load-balance checks as soon a just a single cpu is fully utilized as it might one of those with reduced capacity and in that case we want to migrate it. [ peterz: Added a comment explaining why new tasks are not accounted during overutilization detection. ] Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Quentin Perret <quentin.perret@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: adharmap@codeaurora.org Cc: chris.redpath@arm.com Cc: currojerez@riseup.net Cc: dietmar.eggemann@arm.com Cc: edubezval@gmail.com Cc: gregkh@linuxfoundation.org Cc: javi.merino@kernel.org Cc: joel@joelfernandes.org Cc: juri.lelli@redhat.com Cc: patrick.bellasi@arm.com Cc: pkondeti@codeaurora.org Cc: rjw@rjwysocki.net Cc: skannan@codeaurora.org Cc: smuckle@google.com Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Cc: viresh.kumar@linaro.org Link: https://lkml.kernel.org/r/20181203095628.11858-13-quentin.perret@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-12-03 09:56:25 +00:00
} else if (sg_status & SG_OVERUTILIZED) {
struct root_domain *rd = env->dst_rq->rd;
WRITE_ONCE(rd->overutilized, SG_OVERUTILIZED);
trace_sched_overutilized_tp(rd, SG_OVERUTILIZED);
sched/fair: Implement fast idling of CPUs when the system is partially loaded When a system is lightly loaded (i.e. no more than 1 job per cpu), attempt to pull job to a cpu before putting it to idle is unnecessary and can be skipped. This patch adds an indicator so the scheduler can know when there's no more than 1 active job is on any CPU in the system to skip needless job pulls. On a 4 socket machine with a request/response kind of workload from clients, we saw about 0.13 msec delay when we go through a full load balance to try pull job from all the other cpus. While 0.1 msec was spent on processing the request and generating a response, the 0.13 msec load balance overhead was actually more than the actual work being done. This overhead can be skipped much of the time for lightly loaded systems. With this patch, we tested with a netperf request/response workload that has the server busy with half the cpus in a 4 socket system. We found the patch eliminated 75% of the load balance attempts before idling a cpu. The overhead of setting/clearing the indicator is low as we already gather the necessary info while we call add_nr_running() and update_sd_lb_stats.() We switch to full load balance load immediately if any cpu got more than one job on its run queue in add_nr_running. We'll clear the indicator to avoid load balance when we detect no cpu's have more than one job when we scan the work queues in update_sg_lb_stats(). We are aggressive in turning on the load balance and opportunistic in skipping the load balance. Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Acked-by: Rik van Riel <riel@redhat.com> Acked-by: Jason Low <jason.low2@hp.com> Cc: "Paul E.McKenney" <paulmck@linux.vnet.ibm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Davidlohr Bueso <davidlohr@hp.com> Cc: Alex Shi <alex.shi@linaro.org> Cc: Michel Lespinasse <walken@google.com> Cc: Peter Hurley <peter@hurleysoftware.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1403551009.2970.613.camel@schen9-DESK Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-06-23 19:16:49 +00:00
}
sched: Add asymmetric group packing option for sibling domain Check to see if the group is packed in a sched doman. This is primarily intended to used at the sibling level. Some cores like POWER7 prefer to use lower numbered SMT threads. In the case of POWER7, it can move to lower SMT modes only when higher threads are idle. When in lower SMT modes, the threads will perform better since they share less core resources. Hence when we have idle threads, we want them to be the higher ones. This adds a hook into f_b_g() called check_asym_packing() to check the packing. This packing function is run on idle threads. It checks to see if the busiest CPU in this domain (core in the P7 case) has a higher CPU number than what where the packing function is being run on. If it is, calculate the imbalance and return the higher busier thread as the busiest group to f_b_g(). Here we are assuming a lower CPU number will be equivalent to a lower SMT thread number. It also creates a new SD_ASYM_PACKING flag to enable this feature at any scheduler domain level. It also creates an arch hook to enable this feature at the sibling level. The default function doesn't enable this feature. Based heavily on patch from Peter Zijlstra. Fixes from Srivatsa Vaddagiri. Signed-off-by: Michael Neuling <mikey@neuling.org> Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Thomas Gleixner <tglx@linutronix.de> LKML-Reference: <20100608045702.2936CCC897@localhost.localdomain> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-06-08 04:57:02 +00:00
}
static inline long adjust_numa_imbalance(int imbalance, int src_nr_running)
{
unsigned int imbalance_min;
/*
* Allow a small imbalance based on a simple pair of communicating
* tasks that remain local when the source domain is almost idle.
*/
imbalance_min = 2;
if (src_nr_running <= imbalance_min)
return 0;
return imbalance;
}
/**
* calculate_imbalance - Calculate the amount of imbalance present within the
* groups of a given sched_domain during load balance.
* @env: load balance environment
* @sds: statistics of the sched_domain whose imbalance is to be calculated.
*/
static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
{
struct sg_lb_stats *local, *busiest;
local = &sds->local_stat;
busiest = &sds->busiest_stat;
sched: Fix SCHED_MC regression caused by change in sched cpu_power On platforms like dual socket quad-core platform, the scheduler load balancer is not detecting the load imbalances in certain scenarios. This is leading to scenarios like where one socket is completely busy (with all the 4 cores running with 4 tasks) and leaving another socket completely idle. This causes performance issues as those 4 tasks share the memory controller, last-level cache bandwidth etc. Also we won't be taking advantage of turbo-mode as much as we would like, etc. Some of the comparisons in the scheduler load balancing code are comparing the "weighted cpu load that is scaled wrt sched_group's cpu_power" with the "weighted average load per task that is not scaled wrt sched_group's cpu_power". While this has probably been broken for a longer time (for multi socket numa nodes etc), the problem got aggrevated via this recent change: | | commit f93e65c186ab3c05ce2068733ca10e34fd00125e | Author: Peter Zijlstra <a.p.zijlstra@chello.nl> | Date: Tue Sep 1 10:34:32 2009 +0200 | | sched: Restore __cpu_power to a straight sum of power | Also with this change, the sched group cpu power alone no longer reflects the group capacity that is needed to implement MC, MT performance (default) and power-savings (user-selectable) policies. We need to use the computed group capacity (sgs.group_capacity, that is computed using the SD_PREFER_SIBLING logic in update_sd_lb_stats()) to find out if the group with the max load is above its capacity and how much load to move etc. Reported-by: Ma Ling <ling.ma@intel.com> Initial-Analysis-by: Zhang, Yanmin <yanmin_zhang@linux.intel.com> Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> [ -v2: build fix ] Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: <stable@kernel.org> # [2.6.32.x, 2.6.33.x] LKML-Reference: <1266970432.11588.22.camel@sbs-t61.sc.intel.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-02-24 00:13:52 +00:00
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
if (busiest->group_type == group_misfit_task) {
/* Set imbalance to allow misfit tasks to be balanced. */
env->migration_type = migrate_misfit;
env->imbalance = 1;
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
return;
}
if (busiest->group_type == group_asym_packing) {
/*
* In case of asym capacity, we will try to migrate all load to
* the preferred CPU.
*/
env->migration_type = migrate_task;
env->imbalance = busiest->sum_h_nr_running;
return;
}
if (busiest->group_type == group_imbalanced) {
/*
* In the group_imb case we cannot rely on group-wide averages
* to ensure CPU-load equilibrium, try to move any task to fix
* the imbalance. The next load balance will take care of
* balancing back the system.
*/
env->migration_type = migrate_task;
env->imbalance = 1;
return;
}
/*
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
* Try to use spare capacity of local group without overloading it or
* emptying busiest.
*/
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
if (local->group_type == group_has_spare) {
if (busiest->group_type > group_fully_busy) {
/*
* If busiest is overloaded, try to fill spare
* capacity. This might end up creating spare capacity
* in busiest or busiest still being overloaded but
* there is no simple way to directly compute the
* amount of load to migrate in order to balance the
* system.
*/
env->migration_type = migrate_util;
env->imbalance = max(local->group_capacity, local->group_util) -
local->group_util;
/*
* In some cases, the group's utilization is max or even
* higher than capacity because of migrations but the
* local CPU is (newly) idle. There is at least one
* waiting task in this overloaded busiest group. Let's
* try to pull it.
*/
if (env->idle != CPU_NOT_IDLE && env->imbalance == 0) {
env->migration_type = migrate_task;
env->imbalance = 1;
}
return;
}
if (busiest->group_weight == 1 || sds->prefer_sibling) {
unsigned int nr_diff = busiest->sum_nr_running;
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
/*
* When prefer sibling, evenly spread running tasks on
* groups.
*/
env->migration_type = migrate_task;
lsub_positive(&nr_diff, local->sum_nr_running);
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
env->imbalance = nr_diff >> 1;
sched/fair: Allow a small load imbalance between low utilisation SD_NUMA domains The CPU load balancer balances between different domains to spread load and strives to have equal balance everywhere. Communicating tasks can migrate so they are topologically close to each other but these decisions are independent. On a lightly loaded NUMA machine, two communicating tasks pulled together at wakeup time can be pushed apart by the load balancer. In isolation, the load balancer decision is fine but it ignores the tasks data locality and the wakeup/LB paths continually conflict. NUMA balancing is also a factor but it also simply conflicts with the load balancer. This patch allows a fixed degree of imbalance of two tasks to exist between NUMA domains regardless of utilisation levels. In many cases, this prevents communicating tasks being pulled apart. It was evaluated whether the imbalance should be scaled to the domain size. However, no additional benefit was measured across a range of workloads and machines and scaling adds the risk that lower domains have to be rebalanced. While this could change again in the future, such a change should specify the use case and benefit. The most obvious impact is on netperf TCP_STREAM -- two simple communicating tasks with some softirq offload depending on the transmission rate. 2-socket Haswell machine 48 core, HT enabled netperf-tcp -- mmtests config config-network-netperf-unbound baseline lbnuma-v3 Hmean 64 568.73 ( 0.00%) 577.56 * 1.55%* Hmean 128 1089.98 ( 0.00%) 1128.06 * 3.49%* Hmean 256 2061.72 ( 0.00%) 2104.39 * 2.07%* Hmean 1024 7254.27 ( 0.00%) 7557.52 * 4.18%* Hmean 2048 11729.20 ( 0.00%) 13350.67 * 13.82%* Hmean 3312 15309.08 ( 0.00%) 18058.95 * 17.96%* Hmean 4096 17338.75 ( 0.00%) 20483.66 * 18.14%* Hmean 8192 25047.12 ( 0.00%) 27806.84 * 11.02%* Hmean 16384 27359.55 ( 0.00%) 33071.88 * 20.88%* Stddev 64 2.16 ( 0.00%) 2.02 ( 6.53%) Stddev 128 2.31 ( 0.00%) 2.19 ( 5.05%) Stddev 256 11.88 ( 0.00%) 3.22 ( 72.88%) Stddev 1024 23.68 ( 0.00%) 7.24 ( 69.43%) Stddev 2048 79.46 ( 0.00%) 71.49 ( 10.03%) Stddev 3312 26.71 ( 0.00%) 57.80 (-116.41%) Stddev 4096 185.57 ( 0.00%) 96.15 ( 48.19%) Stddev 8192 245.80 ( 0.00%) 100.73 ( 59.02%) Stddev 16384 207.31 ( 0.00%) 141.65 ( 31.67%) In this case, there was a sizable improvement to performance and a general reduction in variance. However, this is not univeral. For most machines, the impact was roughly a 3% performance gain. Ops NUMA base-page range updates 19796.00 292.00 Ops NUMA PTE updates 19796.00 292.00 Ops NUMA PMD updates 0.00 0.00 Ops NUMA hint faults 16113.00 143.00 Ops NUMA hint local faults % 8407.00 142.00 Ops NUMA hint local percent 52.18 99.30 Ops NUMA pages migrated 4244.00 1.00 Without the patch, only 52.18% of sampled accesses are local. In an earlier changelog, 100% of sampled accesses are local and indeed on most machines, this was still the case. In this specific case, the local sampled rates was 99.3% but note the "base-page range updates" and "PTE updates". The activity with the patch is negligible as were the number of faults. The small number of pages migrated were related to shared libraries. A 2-socket Broadwell showed better results on average but are not presented for brevity as the performance was similar except it showed 100% of the sampled NUMA hints were local. The patch holds up for a 4-socket Haswell, an AMD EPYC and AMD Epyc 2 machine. For dbench, the impact depends on the filesystem used and the number of clients. On XFS, there is little difference as the clients typically communicate with workqueues which have a separate class of scheduler problem at the moment. For ext4, performance is generally better, particularly for small numbers of clients as NUMA balancing activity is negligible with the patch applied. A more interesting example is the Facebook schbench which uses a number of messaging threads to communicate with worker threads. In this configuration, one messaging thread is used per NUMA node and the number of worker threads is varied. The 50, 75, 90, 95, 99, 99.5 and 99.9 percentiles for response latency is then reported. Lat 50.00th-qrtle-1 44.00 ( 0.00%) 37.00 ( 15.91%) Lat 75.00th-qrtle-1 53.00 ( 0.00%) 41.00 ( 22.64%) Lat 90.00th-qrtle-1 57.00 ( 0.00%) 42.00 ( 26.32%) Lat 95.00th-qrtle-1 63.00 ( 0.00%) 43.00 ( 31.75%) Lat 99.00th-qrtle-1 76.00 ( 0.00%) 51.00 ( 32.89%) Lat 99.50th-qrtle-1 89.00 ( 0.00%) 52.00 ( 41.57%) Lat 99.90th-qrtle-1 98.00 ( 0.00%) 55.00 ( 43.88%) Lat 50.00th-qrtle-2 42.00 ( 0.00%) 42.00 ( 0.00%) Lat 75.00th-qrtle-2 48.00 ( 0.00%) 47.00 ( 2.08%) Lat 90.00th-qrtle-2 53.00 ( 0.00%) 52.00 ( 1.89%) Lat 95.00th-qrtle-2 55.00 ( 0.00%) 53.00 ( 3.64%) Lat 99.00th-qrtle-2 62.00 ( 0.00%) 60.00 ( 3.23%) Lat 99.50th-qrtle-2 63.00 ( 0.00%) 63.00 ( 0.00%) Lat 99.90th-qrtle-2 68.00 ( 0.00%) 66.00 ( 2.94% For higher worker threads, the differences become negligible but it's interesting to note the difference in wakeup latency at low utilisation and mpstat confirms that activity was almost all on one node until the number of worker threads increase. Hackbench generally showed neutral results across a range of machines. This is different to earlier versions of the patch which allowed imbalances for higher degrees of utilisation. perf bench pipe showed negligible differences in overall performance as the differences are very close to the noise. An earlier prototype of the patch showed major regressions for NAS C-class when running with only half of the available CPUs -- 20-30% performance hits were measured at the time. With this version of the patch, the impact is negligible with small gains/losses within the noise measured. This is because the number of threads far exceeds the small imbalance the aptch cares about. Similarly, there were report of regressions for the autonuma benchmark against earlier versions but again, normal load balancing now applies for that workload. In general, the patch simply seeks to avoid unnecessary cross-node migrations in the basic case where imbalances are very small. For low utilisation communicating workloads, this patch generally behaves better with less NUMA balancing activity. For high utilisation, there is no change in behaviour. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: Valentin Schneider <valentin.schneider@arm.com> Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org> Reviewed-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Acked-by: Phil Auld <pauld@redhat.com> Tested-by: Phil Auld <pauld@redhat.com> Link: https://lkml.kernel.org/r/20200114101319.GO3466@techsingularity.net
2020-01-14 10:13:20 +00:00
} else {
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
sched/fair: Allow a small load imbalance between low utilisation SD_NUMA domains The CPU load balancer balances between different domains to spread load and strives to have equal balance everywhere. Communicating tasks can migrate so they are topologically close to each other but these decisions are independent. On a lightly loaded NUMA machine, two communicating tasks pulled together at wakeup time can be pushed apart by the load balancer. In isolation, the load balancer decision is fine but it ignores the tasks data locality and the wakeup/LB paths continually conflict. NUMA balancing is also a factor but it also simply conflicts with the load balancer. This patch allows a fixed degree of imbalance of two tasks to exist between NUMA domains regardless of utilisation levels. In many cases, this prevents communicating tasks being pulled apart. It was evaluated whether the imbalance should be scaled to the domain size. However, no additional benefit was measured across a range of workloads and machines and scaling adds the risk that lower domains have to be rebalanced. While this could change again in the future, such a change should specify the use case and benefit. The most obvious impact is on netperf TCP_STREAM -- two simple communicating tasks with some softirq offload depending on the transmission rate. 2-socket Haswell machine 48 core, HT enabled netperf-tcp -- mmtests config config-network-netperf-unbound baseline lbnuma-v3 Hmean 64 568.73 ( 0.00%) 577.56 * 1.55%* Hmean 128 1089.98 ( 0.00%) 1128.06 * 3.49%* Hmean 256 2061.72 ( 0.00%) 2104.39 * 2.07%* Hmean 1024 7254.27 ( 0.00%) 7557.52 * 4.18%* Hmean 2048 11729.20 ( 0.00%) 13350.67 * 13.82%* Hmean 3312 15309.08 ( 0.00%) 18058.95 * 17.96%* Hmean 4096 17338.75 ( 0.00%) 20483.66 * 18.14%* Hmean 8192 25047.12 ( 0.00%) 27806.84 * 11.02%* Hmean 16384 27359.55 ( 0.00%) 33071.88 * 20.88%* Stddev 64 2.16 ( 0.00%) 2.02 ( 6.53%) Stddev 128 2.31 ( 0.00%) 2.19 ( 5.05%) Stddev 256 11.88 ( 0.00%) 3.22 ( 72.88%) Stddev 1024 23.68 ( 0.00%) 7.24 ( 69.43%) Stddev 2048 79.46 ( 0.00%) 71.49 ( 10.03%) Stddev 3312 26.71 ( 0.00%) 57.80 (-116.41%) Stddev 4096 185.57 ( 0.00%) 96.15 ( 48.19%) Stddev 8192 245.80 ( 0.00%) 100.73 ( 59.02%) Stddev 16384 207.31 ( 0.00%) 141.65 ( 31.67%) In this case, there was a sizable improvement to performance and a general reduction in variance. However, this is not univeral. For most machines, the impact was roughly a 3% performance gain. Ops NUMA base-page range updates 19796.00 292.00 Ops NUMA PTE updates 19796.00 292.00 Ops NUMA PMD updates 0.00 0.00 Ops NUMA hint faults 16113.00 143.00 Ops NUMA hint local faults % 8407.00 142.00 Ops NUMA hint local percent 52.18 99.30 Ops NUMA pages migrated 4244.00 1.00 Without the patch, only 52.18% of sampled accesses are local. In an earlier changelog, 100% of sampled accesses are local and indeed on most machines, this was still the case. In this specific case, the local sampled rates was 99.3% but note the "base-page range updates" and "PTE updates". The activity with the patch is negligible as were the number of faults. The small number of pages migrated were related to shared libraries. A 2-socket Broadwell showed better results on average but are not presented for brevity as the performance was similar except it showed 100% of the sampled NUMA hints were local. The patch holds up for a 4-socket Haswell, an AMD EPYC and AMD Epyc 2 machine. For dbench, the impact depends on the filesystem used and the number of clients. On XFS, there is little difference as the clients typically communicate with workqueues which have a separate class of scheduler problem at the moment. For ext4, performance is generally better, particularly for small numbers of clients as NUMA balancing activity is negligible with the patch applied. A more interesting example is the Facebook schbench which uses a number of messaging threads to communicate with worker threads. In this configuration, one messaging thread is used per NUMA node and the number of worker threads is varied. The 50, 75, 90, 95, 99, 99.5 and 99.9 percentiles for response latency is then reported. Lat 50.00th-qrtle-1 44.00 ( 0.00%) 37.00 ( 15.91%) Lat 75.00th-qrtle-1 53.00 ( 0.00%) 41.00 ( 22.64%) Lat 90.00th-qrtle-1 57.00 ( 0.00%) 42.00 ( 26.32%) Lat 95.00th-qrtle-1 63.00 ( 0.00%) 43.00 ( 31.75%) Lat 99.00th-qrtle-1 76.00 ( 0.00%) 51.00 ( 32.89%) Lat 99.50th-qrtle-1 89.00 ( 0.00%) 52.00 ( 41.57%) Lat 99.90th-qrtle-1 98.00 ( 0.00%) 55.00 ( 43.88%) Lat 50.00th-qrtle-2 42.00 ( 0.00%) 42.00 ( 0.00%) Lat 75.00th-qrtle-2 48.00 ( 0.00%) 47.00 ( 2.08%) Lat 90.00th-qrtle-2 53.00 ( 0.00%) 52.00 ( 1.89%) Lat 95.00th-qrtle-2 55.00 ( 0.00%) 53.00 ( 3.64%) Lat 99.00th-qrtle-2 62.00 ( 0.00%) 60.00 ( 3.23%) Lat 99.50th-qrtle-2 63.00 ( 0.00%) 63.00 ( 0.00%) Lat 99.90th-qrtle-2 68.00 ( 0.00%) 66.00 ( 2.94% For higher worker threads, the differences become negligible but it's interesting to note the difference in wakeup latency at low utilisation and mpstat confirms that activity was almost all on one node until the number of worker threads increase. Hackbench generally showed neutral results across a range of machines. This is different to earlier versions of the patch which allowed imbalances for higher degrees of utilisation. perf bench pipe showed negligible differences in overall performance as the differences are very close to the noise. An earlier prototype of the patch showed major regressions for NAS C-class when running with only half of the available CPUs -- 20-30% performance hits were measured at the time. With this version of the patch, the impact is negligible with small gains/losses within the noise measured. This is because the number of threads far exceeds the small imbalance the aptch cares about. Similarly, there were report of regressions for the autonuma benchmark against earlier versions but again, normal load balancing now applies for that workload. In general, the patch simply seeks to avoid unnecessary cross-node migrations in the basic case where imbalances are very small. For low utilisation communicating workloads, this patch generally behaves better with less NUMA balancing activity. For high utilisation, there is no change in behaviour. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: Valentin Schneider <valentin.schneider@arm.com> Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org> Reviewed-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Acked-by: Phil Auld <pauld@redhat.com> Tested-by: Phil Auld <pauld@redhat.com> Link: https://lkml.kernel.org/r/20200114101319.GO3466@techsingularity.net
2020-01-14 10:13:20 +00:00
/*
* If there is no overload, we just want to even the number of
* idle cpus.
*/
env->migration_type = migrate_task;
env->imbalance = max_t(long, 0, (local->idle_cpus -
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
busiest->idle_cpus) >> 1);
sched/fair: Allow a small load imbalance between low utilisation SD_NUMA domains The CPU load balancer balances between different domains to spread load and strives to have equal balance everywhere. Communicating tasks can migrate so they are topologically close to each other but these decisions are independent. On a lightly loaded NUMA machine, two communicating tasks pulled together at wakeup time can be pushed apart by the load balancer. In isolation, the load balancer decision is fine but it ignores the tasks data locality and the wakeup/LB paths continually conflict. NUMA balancing is also a factor but it also simply conflicts with the load balancer. This patch allows a fixed degree of imbalance of two tasks to exist between NUMA domains regardless of utilisation levels. In many cases, this prevents communicating tasks being pulled apart. It was evaluated whether the imbalance should be scaled to the domain size. However, no additional benefit was measured across a range of workloads and machines and scaling adds the risk that lower domains have to be rebalanced. While this could change again in the future, such a change should specify the use case and benefit. The most obvious impact is on netperf TCP_STREAM -- two simple communicating tasks with some softirq offload depending on the transmission rate. 2-socket Haswell machine 48 core, HT enabled netperf-tcp -- mmtests config config-network-netperf-unbound baseline lbnuma-v3 Hmean 64 568.73 ( 0.00%) 577.56 * 1.55%* Hmean 128 1089.98 ( 0.00%) 1128.06 * 3.49%* Hmean 256 2061.72 ( 0.00%) 2104.39 * 2.07%* Hmean 1024 7254.27 ( 0.00%) 7557.52 * 4.18%* Hmean 2048 11729.20 ( 0.00%) 13350.67 * 13.82%* Hmean 3312 15309.08 ( 0.00%) 18058.95 * 17.96%* Hmean 4096 17338.75 ( 0.00%) 20483.66 * 18.14%* Hmean 8192 25047.12 ( 0.00%) 27806.84 * 11.02%* Hmean 16384 27359.55 ( 0.00%) 33071.88 * 20.88%* Stddev 64 2.16 ( 0.00%) 2.02 ( 6.53%) Stddev 128 2.31 ( 0.00%) 2.19 ( 5.05%) Stddev 256 11.88 ( 0.00%) 3.22 ( 72.88%) Stddev 1024 23.68 ( 0.00%) 7.24 ( 69.43%) Stddev 2048 79.46 ( 0.00%) 71.49 ( 10.03%) Stddev 3312 26.71 ( 0.00%) 57.80 (-116.41%) Stddev 4096 185.57 ( 0.00%) 96.15 ( 48.19%) Stddev 8192 245.80 ( 0.00%) 100.73 ( 59.02%) Stddev 16384 207.31 ( 0.00%) 141.65 ( 31.67%) In this case, there was a sizable improvement to performance and a general reduction in variance. However, this is not univeral. For most machines, the impact was roughly a 3% performance gain. Ops NUMA base-page range updates 19796.00 292.00 Ops NUMA PTE updates 19796.00 292.00 Ops NUMA PMD updates 0.00 0.00 Ops NUMA hint faults 16113.00 143.00 Ops NUMA hint local faults % 8407.00 142.00 Ops NUMA hint local percent 52.18 99.30 Ops NUMA pages migrated 4244.00 1.00 Without the patch, only 52.18% of sampled accesses are local. In an earlier changelog, 100% of sampled accesses are local and indeed on most machines, this was still the case. In this specific case, the local sampled rates was 99.3% but note the "base-page range updates" and "PTE updates". The activity with the patch is negligible as were the number of faults. The small number of pages migrated were related to shared libraries. A 2-socket Broadwell showed better results on average but are not presented for brevity as the performance was similar except it showed 100% of the sampled NUMA hints were local. The patch holds up for a 4-socket Haswell, an AMD EPYC and AMD Epyc 2 machine. For dbench, the impact depends on the filesystem used and the number of clients. On XFS, there is little difference as the clients typically communicate with workqueues which have a separate class of scheduler problem at the moment. For ext4, performance is generally better, particularly for small numbers of clients as NUMA balancing activity is negligible with the patch applied. A more interesting example is the Facebook schbench which uses a number of messaging threads to communicate with worker threads. In this configuration, one messaging thread is used per NUMA node and the number of worker threads is varied. The 50, 75, 90, 95, 99, 99.5 and 99.9 percentiles for response latency is then reported. Lat 50.00th-qrtle-1 44.00 ( 0.00%) 37.00 ( 15.91%) Lat 75.00th-qrtle-1 53.00 ( 0.00%) 41.00 ( 22.64%) Lat 90.00th-qrtle-1 57.00 ( 0.00%) 42.00 ( 26.32%) Lat 95.00th-qrtle-1 63.00 ( 0.00%) 43.00 ( 31.75%) Lat 99.00th-qrtle-1 76.00 ( 0.00%) 51.00 ( 32.89%) Lat 99.50th-qrtle-1 89.00 ( 0.00%) 52.00 ( 41.57%) Lat 99.90th-qrtle-1 98.00 ( 0.00%) 55.00 ( 43.88%) Lat 50.00th-qrtle-2 42.00 ( 0.00%) 42.00 ( 0.00%) Lat 75.00th-qrtle-2 48.00 ( 0.00%) 47.00 ( 2.08%) Lat 90.00th-qrtle-2 53.00 ( 0.00%) 52.00 ( 1.89%) Lat 95.00th-qrtle-2 55.00 ( 0.00%) 53.00 ( 3.64%) Lat 99.00th-qrtle-2 62.00 ( 0.00%) 60.00 ( 3.23%) Lat 99.50th-qrtle-2 63.00 ( 0.00%) 63.00 ( 0.00%) Lat 99.90th-qrtle-2 68.00 ( 0.00%) 66.00 ( 2.94% For higher worker threads, the differences become negligible but it's interesting to note the difference in wakeup latency at low utilisation and mpstat confirms that activity was almost all on one node until the number of worker threads increase. Hackbench generally showed neutral results across a range of machines. This is different to earlier versions of the patch which allowed imbalances for higher degrees of utilisation. perf bench pipe showed negligible differences in overall performance as the differences are very close to the noise. An earlier prototype of the patch showed major regressions for NAS C-class when running with only half of the available CPUs -- 20-30% performance hits were measured at the time. With this version of the patch, the impact is negligible with small gains/losses within the noise measured. This is because the number of threads far exceeds the small imbalance the aptch cares about. Similarly, there were report of regressions for the autonuma benchmark against earlier versions but again, normal load balancing now applies for that workload. In general, the patch simply seeks to avoid unnecessary cross-node migrations in the basic case where imbalances are very small. For low utilisation communicating workloads, this patch generally behaves better with less NUMA balancing activity. For high utilisation, there is no change in behaviour. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: Valentin Schneider <valentin.schneider@arm.com> Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org> Reviewed-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Acked-by: Phil Auld <pauld@redhat.com> Tested-by: Phil Auld <pauld@redhat.com> Link: https://lkml.kernel.org/r/20200114101319.GO3466@techsingularity.net
2020-01-14 10:13:20 +00:00
}
/* Consider allowing a small imbalance between NUMA groups */
if (env->sd->flags & SD_NUMA)
env->imbalance = adjust_numa_imbalance(env->imbalance,
busiest->sum_nr_running);
sched/fair: Allow a small load imbalance between low utilisation SD_NUMA domains The CPU load balancer balances between different domains to spread load and strives to have equal balance everywhere. Communicating tasks can migrate so they are topologically close to each other but these decisions are independent. On a lightly loaded NUMA machine, two communicating tasks pulled together at wakeup time can be pushed apart by the load balancer. In isolation, the load balancer decision is fine but it ignores the tasks data locality and the wakeup/LB paths continually conflict. NUMA balancing is also a factor but it also simply conflicts with the load balancer. This patch allows a fixed degree of imbalance of two tasks to exist between NUMA domains regardless of utilisation levels. In many cases, this prevents communicating tasks being pulled apart. It was evaluated whether the imbalance should be scaled to the domain size. However, no additional benefit was measured across a range of workloads and machines and scaling adds the risk that lower domains have to be rebalanced. While this could change again in the future, such a change should specify the use case and benefit. The most obvious impact is on netperf TCP_STREAM -- two simple communicating tasks with some softirq offload depending on the transmission rate. 2-socket Haswell machine 48 core, HT enabled netperf-tcp -- mmtests config config-network-netperf-unbound baseline lbnuma-v3 Hmean 64 568.73 ( 0.00%) 577.56 * 1.55%* Hmean 128 1089.98 ( 0.00%) 1128.06 * 3.49%* Hmean 256 2061.72 ( 0.00%) 2104.39 * 2.07%* Hmean 1024 7254.27 ( 0.00%) 7557.52 * 4.18%* Hmean 2048 11729.20 ( 0.00%) 13350.67 * 13.82%* Hmean 3312 15309.08 ( 0.00%) 18058.95 * 17.96%* Hmean 4096 17338.75 ( 0.00%) 20483.66 * 18.14%* Hmean 8192 25047.12 ( 0.00%) 27806.84 * 11.02%* Hmean 16384 27359.55 ( 0.00%) 33071.88 * 20.88%* Stddev 64 2.16 ( 0.00%) 2.02 ( 6.53%) Stddev 128 2.31 ( 0.00%) 2.19 ( 5.05%) Stddev 256 11.88 ( 0.00%) 3.22 ( 72.88%) Stddev 1024 23.68 ( 0.00%) 7.24 ( 69.43%) Stddev 2048 79.46 ( 0.00%) 71.49 ( 10.03%) Stddev 3312 26.71 ( 0.00%) 57.80 (-116.41%) Stddev 4096 185.57 ( 0.00%) 96.15 ( 48.19%) Stddev 8192 245.80 ( 0.00%) 100.73 ( 59.02%) Stddev 16384 207.31 ( 0.00%) 141.65 ( 31.67%) In this case, there was a sizable improvement to performance and a general reduction in variance. However, this is not univeral. For most machines, the impact was roughly a 3% performance gain. Ops NUMA base-page range updates 19796.00 292.00 Ops NUMA PTE updates 19796.00 292.00 Ops NUMA PMD updates 0.00 0.00 Ops NUMA hint faults 16113.00 143.00 Ops NUMA hint local faults % 8407.00 142.00 Ops NUMA hint local percent 52.18 99.30 Ops NUMA pages migrated 4244.00 1.00 Without the patch, only 52.18% of sampled accesses are local. In an earlier changelog, 100% of sampled accesses are local and indeed on most machines, this was still the case. In this specific case, the local sampled rates was 99.3% but note the "base-page range updates" and "PTE updates". The activity with the patch is negligible as were the number of faults. The small number of pages migrated were related to shared libraries. A 2-socket Broadwell showed better results on average but are not presented for brevity as the performance was similar except it showed 100% of the sampled NUMA hints were local. The patch holds up for a 4-socket Haswell, an AMD EPYC and AMD Epyc 2 machine. For dbench, the impact depends on the filesystem used and the number of clients. On XFS, there is little difference as the clients typically communicate with workqueues which have a separate class of scheduler problem at the moment. For ext4, performance is generally better, particularly for small numbers of clients as NUMA balancing activity is negligible with the patch applied. A more interesting example is the Facebook schbench which uses a number of messaging threads to communicate with worker threads. In this configuration, one messaging thread is used per NUMA node and the number of worker threads is varied. The 50, 75, 90, 95, 99, 99.5 and 99.9 percentiles for response latency is then reported. Lat 50.00th-qrtle-1 44.00 ( 0.00%) 37.00 ( 15.91%) Lat 75.00th-qrtle-1 53.00 ( 0.00%) 41.00 ( 22.64%) Lat 90.00th-qrtle-1 57.00 ( 0.00%) 42.00 ( 26.32%) Lat 95.00th-qrtle-1 63.00 ( 0.00%) 43.00 ( 31.75%) Lat 99.00th-qrtle-1 76.00 ( 0.00%) 51.00 ( 32.89%) Lat 99.50th-qrtle-1 89.00 ( 0.00%) 52.00 ( 41.57%) Lat 99.90th-qrtle-1 98.00 ( 0.00%) 55.00 ( 43.88%) Lat 50.00th-qrtle-2 42.00 ( 0.00%) 42.00 ( 0.00%) Lat 75.00th-qrtle-2 48.00 ( 0.00%) 47.00 ( 2.08%) Lat 90.00th-qrtle-2 53.00 ( 0.00%) 52.00 ( 1.89%) Lat 95.00th-qrtle-2 55.00 ( 0.00%) 53.00 ( 3.64%) Lat 99.00th-qrtle-2 62.00 ( 0.00%) 60.00 ( 3.23%) Lat 99.50th-qrtle-2 63.00 ( 0.00%) 63.00 ( 0.00%) Lat 99.90th-qrtle-2 68.00 ( 0.00%) 66.00 ( 2.94% For higher worker threads, the differences become negligible but it's interesting to note the difference in wakeup latency at low utilisation and mpstat confirms that activity was almost all on one node until the number of worker threads increase. Hackbench generally showed neutral results across a range of machines. This is different to earlier versions of the patch which allowed imbalances for higher degrees of utilisation. perf bench pipe showed negligible differences in overall performance as the differences are very close to the noise. An earlier prototype of the patch showed major regressions for NAS C-class when running with only half of the available CPUs -- 20-30% performance hits were measured at the time. With this version of the patch, the impact is negligible with small gains/losses within the noise measured. This is because the number of threads far exceeds the small imbalance the aptch cares about. Similarly, there were report of regressions for the autonuma benchmark against earlier versions but again, normal load balancing now applies for that workload. In general, the patch simply seeks to avoid unnecessary cross-node migrations in the basic case where imbalances are very small. For low utilisation communicating workloads, this patch generally behaves better with less NUMA balancing activity. For high utilisation, there is no change in behaviour. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: Valentin Schneider <valentin.schneider@arm.com> Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org> Reviewed-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Acked-by: Phil Auld <pauld@redhat.com> Tested-by: Phil Auld <pauld@redhat.com> Link: https://lkml.kernel.org/r/20200114101319.GO3466@techsingularity.net
2020-01-14 10:13:20 +00:00
return;
}
/*
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
* Local is fully busy but has to take more load to relieve the
* busiest group
*/
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
if (local->group_type < group_overloaded) {
/*
* Local will become overloaded so the avg_load metrics are
* finally needed.
*/
local->avg_load = (local->group_load * SCHED_CAPACITY_SCALE) /
local->group_capacity;
sds->avg_load = (sds->total_load * SCHED_CAPACITY_SCALE) /
sds->total_capacity;
/*
* If the local group is more loaded than the selected
* busiest group don't try to pull any tasks.
*/
if (local->avg_load >= busiest->avg_load) {
env->imbalance = 0;
return;
}
sched: Fix SCHED_MC regression caused by change in sched cpu_power On platforms like dual socket quad-core platform, the scheduler load balancer is not detecting the load imbalances in certain scenarios. This is leading to scenarios like where one socket is completely busy (with all the 4 cores running with 4 tasks) and leaving another socket completely idle. This causes performance issues as those 4 tasks share the memory controller, last-level cache bandwidth etc. Also we won't be taking advantage of turbo-mode as much as we would like, etc. Some of the comparisons in the scheduler load balancing code are comparing the "weighted cpu load that is scaled wrt sched_group's cpu_power" with the "weighted average load per task that is not scaled wrt sched_group's cpu_power". While this has probably been broken for a longer time (for multi socket numa nodes etc), the problem got aggrevated via this recent change: | | commit f93e65c186ab3c05ce2068733ca10e34fd00125e | Author: Peter Zijlstra <a.p.zijlstra@chello.nl> | Date: Tue Sep 1 10:34:32 2009 +0200 | | sched: Restore __cpu_power to a straight sum of power | Also with this change, the sched group cpu power alone no longer reflects the group capacity that is needed to implement MC, MT performance (default) and power-savings (user-selectable) policies. We need to use the computed group capacity (sgs.group_capacity, that is computed using the SD_PREFER_SIBLING logic in update_sd_lb_stats()) to find out if the group with the max load is above its capacity and how much load to move etc. Reported-by: Ma Ling <ling.ma@intel.com> Initial-Analysis-by: Zhang, Yanmin <yanmin_zhang@linux.intel.com> Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> [ -v2: build fix ] Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: <stable@kernel.org> # [2.6.32.x, 2.6.33.x] LKML-Reference: <1266970432.11588.22.camel@sbs-t61.sc.intel.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-02-24 00:13:52 +00:00
}
/*
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
* Both group are or will become overloaded and we're trying to get all
* the CPUs to the average_load, so we don't want to push ourselves
* above the average load, nor do we wish to reduce the max loaded CPU
* below the average load. At the same time, we also don't want to
* reduce the group load below the group capacity. Thus we look for
* the minimum possible imbalance.
sched: Fix SCHED_MC regression caused by change in sched cpu_power On platforms like dual socket quad-core platform, the scheduler load balancer is not detecting the load imbalances in certain scenarios. This is leading to scenarios like where one socket is completely busy (with all the 4 cores running with 4 tasks) and leaving another socket completely idle. This causes performance issues as those 4 tasks share the memory controller, last-level cache bandwidth etc. Also we won't be taking advantage of turbo-mode as much as we would like, etc. Some of the comparisons in the scheduler load balancing code are comparing the "weighted cpu load that is scaled wrt sched_group's cpu_power" with the "weighted average load per task that is not scaled wrt sched_group's cpu_power". While this has probably been broken for a longer time (for multi socket numa nodes etc), the problem got aggrevated via this recent change: | | commit f93e65c186ab3c05ce2068733ca10e34fd00125e | Author: Peter Zijlstra <a.p.zijlstra@chello.nl> | Date: Tue Sep 1 10:34:32 2009 +0200 | | sched: Restore __cpu_power to a straight sum of power | Also with this change, the sched group cpu power alone no longer reflects the group capacity that is needed to implement MC, MT performance (default) and power-savings (user-selectable) policies. We need to use the computed group capacity (sgs.group_capacity, that is computed using the SD_PREFER_SIBLING logic in update_sd_lb_stats()) to find out if the group with the max load is above its capacity and how much load to move etc. Reported-by: Ma Ling <ling.ma@intel.com> Initial-Analysis-by: Zhang, Yanmin <yanmin_zhang@linux.intel.com> Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> [ -v2: build fix ] Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: <stable@kernel.org> # [2.6.32.x, 2.6.33.x] LKML-Reference: <1266970432.11588.22.camel@sbs-t61.sc.intel.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-02-24 00:13:52 +00:00
*/
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
env->migration_type = migrate_load;
env->imbalance = min(
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
(busiest->avg_load - sds->avg_load) * busiest->group_capacity,
(sds->avg_load - local->avg_load) * local->group_capacity
sched: Final power vs. capacity cleanups It is better not to think about compute capacity as being equivalent to "CPU power". The upcoming "power aware" scheduler work may create confusion with the notion of energy consumption if "power" is used too liberally. This contains the architecture visible changes. Incidentally, only ARM takes advantage of the available pow^H^H^Hcapacity scaling hooks and therefore those changes outside kernel/sched/ are confined to one ARM specific file. The default arch_scale_smt_power() hook is not overridden by anyone. Replacements are as follows: arch_scale_freq_power --> arch_scale_freq_capacity arch_scale_smt_power --> arch_scale_smt_capacity SCHED_POWER_SCALE --> SCHED_CAPACITY_SCALE SCHED_POWER_SHIFT --> SCHED_CAPACITY_SHIFT The local usage of "power" in arch/arm/kernel/topology.c is also changed to "capacity" as appropriate. Signed-off-by: Nicolas Pitre <nico@linaro.org> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Daniel Lezcano <daniel.lezcano@linaro.org> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: "Rafael J. Wysocki" <rjw@rjwysocki.net> Cc: linaro-kernel@lists.linaro.org Cc: Arnd Bergmann <arnd@arndb.de> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Grant Likely <grant.likely@linaro.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mark Brown <broonie@linaro.org> Cc: Rob Herring <robh+dt@kernel.org> Cc: Russell King <linux@arm.linux.org.uk> Cc: Sudeep KarkadaNagesha <sudeep.karkadanagesha@arm.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: devicetree@vger.kernel.org Cc: linux-arm-kernel@lists.infradead.org Cc: linux-kernel@vger.kernel.org Link: http://lkml.kernel.org/n/tip-48zba9qbznvglwelgq2cfygh@git.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-05-26 22:19:39 +00:00
) / SCHED_CAPACITY_SCALE;
}
/******* find_busiest_group() helpers end here *********************/
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
/*
* Decision matrix according to the local and busiest group type:
*
* busiest \ local has_spare fully_busy misfit asym imbalanced overloaded
* has_spare nr_idle balanced N/A N/A balanced balanced
* fully_busy nr_idle nr_idle N/A N/A balanced balanced
* misfit_task force N/A N/A N/A force force
* asym_packing force force N/A N/A force force
* imbalanced force force N/A N/A force force
* overloaded force force N/A N/A force avg_load
*
* N/A : Not Applicable because already filtered while updating
* statistics.
* balanced : The system is balanced for these 2 groups.
* force : Calculate the imbalance as load migration is probably needed.
* avg_load : Only if imbalance is significant enough.
* nr_idle : dst_cpu is not busy and the number of idle CPUs is quite
* different in groups.
*/
/**
* find_busiest_group - Returns the busiest group within the sched_domain
* if there is an imbalance.
*
* Also calculates the amount of runnable load which should be moved
* to restore balance.
*
* @env: The load balancing environment.
*
* Return: - The busiest group if imbalance exists.
*/
static struct sched_group *find_busiest_group(struct lb_env *env)
{
struct sg_lb_stats *local, *busiest;
struct sd_lb_stats sds;
init_sd_lb_stats(&sds);
/*
* Compute the various statistics relevant for load balancing at
* this level.
*/
update_sd_lb_stats(env, &sds);
sched/fair: Add over-utilization/tipping point indicator Energy-aware scheduling is only meant to be active while the system is _not_ over-utilized. That is, there are spare cycles available to shift tasks around based on their actual utilization to get a more energy-efficient task distribution without depriving any tasks. When above the tipping point task placement is done the traditional way based on load_avg, spreading the tasks across as many cpus as possible based on priority scaled load to preserve smp_nice. Below the tipping point we want to use util_avg instead. We need to define a criteria for when we make the switch. The util_avg for each cpu converges towards 100% regardless of how many additional tasks we may put on it. If we define over-utilized as: sum_{cpus}(rq.cfs.avg.util_avg) + margin > sum_{cpus}(rq.capacity) some individual cpus may be over-utilized running multiple tasks even when the above condition is false. That should be okay as long as we try to spread the tasks out to avoid per-cpu over-utilization as much as possible and if all tasks have the _same_ priority. If the latter isn't true, we have to consider priority to preserve smp_nice. For example, we could have n_cpus nice=-10 util_avg=55% tasks and n_cpus/2 nice=0 util_avg=60% tasks. Balancing based on util_avg we are likely to end up with nice=-10 tasks sharing cpus and nice=0 tasks getting their own as we 1.5*n_cpus tasks in total and 55%+55% is less over-utilized than 55%+60% for those cpus that have to be shared. The system utilization is only 85% of the system capacity, but we are breaking smp_nice. To be sure not to break smp_nice, we have defined over-utilization conservatively as when any cpu in the system is fully utilized at its highest frequency instead: cpu_rq(any).cfs.avg.util_avg + margin > cpu_rq(any).capacity IOW, as soon as one cpu is (nearly) 100% utilized, we switch to load_avg to factor in priority to preserve smp_nice. With this definition, we can skip periodic load-balance as no cpu has an always-running task when the system is not over-utilized. All tasks will be periodic and we can balance them at wake-up. This conservative condition does however mean that some scenarios that could benefit from energy-aware decisions even if one cpu is fully utilized would not get those benefits. For systems where some cpus might have reduced capacity on some cpus (RT-pressure and/or big.LITTLE), we want periodic load-balance checks as soon a just a single cpu is fully utilized as it might one of those with reduced capacity and in that case we want to migrate it. [ peterz: Added a comment explaining why new tasks are not accounted during overutilization detection. ] Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Quentin Perret <quentin.perret@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: adharmap@codeaurora.org Cc: chris.redpath@arm.com Cc: currojerez@riseup.net Cc: dietmar.eggemann@arm.com Cc: edubezval@gmail.com Cc: gregkh@linuxfoundation.org Cc: javi.merino@kernel.org Cc: joel@joelfernandes.org Cc: juri.lelli@redhat.com Cc: patrick.bellasi@arm.com Cc: pkondeti@codeaurora.org Cc: rjw@rjwysocki.net Cc: skannan@codeaurora.org Cc: smuckle@google.com Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Cc: viresh.kumar@linaro.org Link: https://lkml.kernel.org/r/20181203095628.11858-13-quentin.perret@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-12-03 09:56:25 +00:00
if (sched_energy_enabled()) {
sched/fair: Add over-utilization/tipping point indicator Energy-aware scheduling is only meant to be active while the system is _not_ over-utilized. That is, there are spare cycles available to shift tasks around based on their actual utilization to get a more energy-efficient task distribution without depriving any tasks. When above the tipping point task placement is done the traditional way based on load_avg, spreading the tasks across as many cpus as possible based on priority scaled load to preserve smp_nice. Below the tipping point we want to use util_avg instead. We need to define a criteria for when we make the switch. The util_avg for each cpu converges towards 100% regardless of how many additional tasks we may put on it. If we define over-utilized as: sum_{cpus}(rq.cfs.avg.util_avg) + margin > sum_{cpus}(rq.capacity) some individual cpus may be over-utilized running multiple tasks even when the above condition is false. That should be okay as long as we try to spread the tasks out to avoid per-cpu over-utilization as much as possible and if all tasks have the _same_ priority. If the latter isn't true, we have to consider priority to preserve smp_nice. For example, we could have n_cpus nice=-10 util_avg=55% tasks and n_cpus/2 nice=0 util_avg=60% tasks. Balancing based on util_avg we are likely to end up with nice=-10 tasks sharing cpus and nice=0 tasks getting their own as we 1.5*n_cpus tasks in total and 55%+55% is less over-utilized than 55%+60% for those cpus that have to be shared. The system utilization is only 85% of the system capacity, but we are breaking smp_nice. To be sure not to break smp_nice, we have defined over-utilization conservatively as when any cpu in the system is fully utilized at its highest frequency instead: cpu_rq(any).cfs.avg.util_avg + margin > cpu_rq(any).capacity IOW, as soon as one cpu is (nearly) 100% utilized, we switch to load_avg to factor in priority to preserve smp_nice. With this definition, we can skip periodic load-balance as no cpu has an always-running task when the system is not over-utilized. All tasks will be periodic and we can balance them at wake-up. This conservative condition does however mean that some scenarios that could benefit from energy-aware decisions even if one cpu is fully utilized would not get those benefits. For systems where some cpus might have reduced capacity on some cpus (RT-pressure and/or big.LITTLE), we want periodic load-balance checks as soon a just a single cpu is fully utilized as it might one of those with reduced capacity and in that case we want to migrate it. [ peterz: Added a comment explaining why new tasks are not accounted during overutilization detection. ] Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Quentin Perret <quentin.perret@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: adharmap@codeaurora.org Cc: chris.redpath@arm.com Cc: currojerez@riseup.net Cc: dietmar.eggemann@arm.com Cc: edubezval@gmail.com Cc: gregkh@linuxfoundation.org Cc: javi.merino@kernel.org Cc: joel@joelfernandes.org Cc: juri.lelli@redhat.com Cc: patrick.bellasi@arm.com Cc: pkondeti@codeaurora.org Cc: rjw@rjwysocki.net Cc: skannan@codeaurora.org Cc: smuckle@google.com Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Cc: viresh.kumar@linaro.org Link: https://lkml.kernel.org/r/20181203095628.11858-13-quentin.perret@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-12-03 09:56:25 +00:00
struct root_domain *rd = env->dst_rq->rd;
if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized))
goto out_balanced;
}
local = &sds.local_stat;
busiest = &sds.busiest_stat;
/* There is no busy sibling group to pull tasks from */
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
if (!sds.busiest)
goto out_balanced;
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
/* Misfit tasks should be dealt with regardless of the avg load */
if (busiest->group_type == group_misfit_task)
goto force_balance;
/* ASYM feature bypasses nice load balance check */
if (busiest->group_type == group_asym_packing)
goto force_balance;
/*
* If the busiest group is imbalanced the below checks don't
* work because they assume all things are equal, which typically
* isn't true due to cpus_ptr constraints and the like.
*/
sched/fair: Make update_sd_pick_busiest() return 'true' on a busier sd Currently update_sd_pick_busiest only identifies the busiest sd that is either overloaded, or has a group imbalance. When no sd is imbalanced or overloaded, the load balancer fails to find the busiest domain. This breaks load balancing between domains that are not overloaded, in the !SD_ASYM_PACKING case. This patch makes update_sd_pick_busiest return true when the busiest sd yet is encountered. Groups are ranked in the order overloaded > imbalanced > other, with higher ranked groups getting priority even when their load is lower. This is necessary due to the possibility of unequal capacities and cpumasks between domains within a sched group. Behaviour for SD_ASYM_PACKING does not seem to match the comment, but I have no hardware to test that so I have left the behaviour of that code unchanged. Enum for group classification suggested by Peter Zijlstra. Signed-off-by: Rik van Riel <riel@redhat.com> [peterz: replaced sg_lb_stats::group_imb with the new enum group_type in an attempt to avoid endless recalculation] Signed-off-by: Peter Zijlstra <peterz@infradead.org> Acked-by: Vincent Guittot <vincent.guittot@linaro.org> Acked-by: Michael Neuling <mikey@neuling.org> Cc: ktkhai@parallels.com Cc: tim.c.chen@linux.intel.com Cc: nicolas.pitre@linaro.org Cc: jhladky@redhat.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/20140729152743.GI3935@laptop Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-07-28 18:16:28 +00:00
if (busiest->group_type == group_imbalanced)
goto force_balance;
/*
* If the local group is busier than the selected busiest group
* don't try and pull any tasks.
*/
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
if (local->group_type > busiest->group_type)
goto out_balanced;
/*
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
* When groups are overloaded, use the avg_load to ensure fairness
* between tasks.
*/
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
if (local->group_type == group_overloaded) {
/*
* If the local group is more loaded than the selected
* busiest group don't try to pull any tasks.
*/
if (local->avg_load >= busiest->avg_load)
goto out_balanced;
/* XXX broken for overlapping NUMA groups */
sds.avg_load = (sds.total_load * SCHED_CAPACITY_SCALE) /
sds.total_capacity;
sched: Use group weight, idle cpu metrics to fix imbalances during idle Currently we consider a sched domain to be well balanced when the imbalance is less than the domain's imablance_pct. As the number of cores and threads are increasing, current values of imbalance_pct (for example 25% for a NUMA domain) are not enough to detect imbalances like: a) On a WSM-EP system (two sockets, each having 6 cores and 12 logical threads), 24 cpu-hogging tasks get scheduled as 13 on one socket and 11 on another socket. Leading to an idle HT cpu. b) On a hypothetial 2 socket NHM-EX system (each socket having 8 cores and 16 logical threads), 16 cpu-hogging tasks can get scheduled as 9 on one socket and 7 on another socket. Leaving one core in a socket idle whereas in another socket we have a core having both its HT siblings busy. While this issue can be fixed by decreasing the domain's imbalance_pct (by making it a function of number of logical cpus in the domain), it can potentially cause more task migrations across sched groups in an overloaded case. Fix this by using imbalance_pct only during newly_idle and busy load balancing. And during idle load balancing, check if there is an imbalance in number of idle cpu's across the busiest and this sched_group or if the busiest group has more tasks than its weight that the idle cpu in this_group can pull. Reported-by: Nikhil Rao <ncrao@google.com> Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> LKML-Reference: <1284760952.2676.11.camel@sbsiddha-MOBL3.sc.intel.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-09-17 22:02:32 +00:00
/*
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
* Don't pull any tasks if this group is already above the
* domain average load.
sched: Use group weight, idle cpu metrics to fix imbalances during idle Currently we consider a sched domain to be well balanced when the imbalance is less than the domain's imablance_pct. As the number of cores and threads are increasing, current values of imbalance_pct (for example 25% for a NUMA domain) are not enough to detect imbalances like: a) On a WSM-EP system (two sockets, each having 6 cores and 12 logical threads), 24 cpu-hogging tasks get scheduled as 13 on one socket and 11 on another socket. Leading to an idle HT cpu. b) On a hypothetial 2 socket NHM-EX system (each socket having 8 cores and 16 logical threads), 16 cpu-hogging tasks can get scheduled as 9 on one socket and 7 on another socket. Leaving one core in a socket idle whereas in another socket we have a core having both its HT siblings busy. While this issue can be fixed by decreasing the domain's imbalance_pct (by making it a function of number of logical cpus in the domain), it can potentially cause more task migrations across sched groups in an overloaded case. Fix this by using imbalance_pct only during newly_idle and busy load balancing. And during idle load balancing, check if there is an imbalance in number of idle cpu's across the busiest and this sched_group or if the busiest group has more tasks than its weight that the idle cpu in this_group can pull. Reported-by: Nikhil Rao <ncrao@google.com> Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> LKML-Reference: <1284760952.2676.11.camel@sbsiddha-MOBL3.sc.intel.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-09-17 22:02:32 +00:00
*/
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
if (local->avg_load >= sds.avg_load)
sched: Use group weight, idle cpu metrics to fix imbalances during idle Currently we consider a sched domain to be well balanced when the imbalance is less than the domain's imablance_pct. As the number of cores and threads are increasing, current values of imbalance_pct (for example 25% for a NUMA domain) are not enough to detect imbalances like: a) On a WSM-EP system (two sockets, each having 6 cores and 12 logical threads), 24 cpu-hogging tasks get scheduled as 13 on one socket and 11 on another socket. Leading to an idle HT cpu. b) On a hypothetial 2 socket NHM-EX system (each socket having 8 cores and 16 logical threads), 16 cpu-hogging tasks can get scheduled as 9 on one socket and 7 on another socket. Leaving one core in a socket idle whereas in another socket we have a core having both its HT siblings busy. While this issue can be fixed by decreasing the domain's imbalance_pct (by making it a function of number of logical cpus in the domain), it can potentially cause more task migrations across sched groups in an overloaded case. Fix this by using imbalance_pct only during newly_idle and busy load balancing. And during idle load balancing, check if there is an imbalance in number of idle cpu's across the busiest and this sched_group or if the busiest group has more tasks than its weight that the idle cpu in this_group can pull. Reported-by: Nikhil Rao <ncrao@google.com> Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> LKML-Reference: <1284760952.2676.11.camel@sbsiddha-MOBL3.sc.intel.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-09-17 22:02:32 +00:00
goto out_balanced;
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
/*
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
* If the busiest group is more loaded, use imbalance_pct to be
* conservative.
*/
if (100 * busiest->avg_load <=
env->sd->imbalance_pct * local->avg_load)
goto out_balanced;
sched: Use group weight, idle cpu metrics to fix imbalances during idle Currently we consider a sched domain to be well balanced when the imbalance is less than the domain's imablance_pct. As the number of cores and threads are increasing, current values of imbalance_pct (for example 25% for a NUMA domain) are not enough to detect imbalances like: a) On a WSM-EP system (two sockets, each having 6 cores and 12 logical threads), 24 cpu-hogging tasks get scheduled as 13 on one socket and 11 on another socket. Leading to an idle HT cpu. b) On a hypothetial 2 socket NHM-EX system (each socket having 8 cores and 16 logical threads), 16 cpu-hogging tasks can get scheduled as 9 on one socket and 7 on another socket. Leaving one core in a socket idle whereas in another socket we have a core having both its HT siblings busy. While this issue can be fixed by decreasing the domain's imbalance_pct (by making it a function of number of logical cpus in the domain), it can potentially cause more task migrations across sched groups in an overloaded case. Fix this by using imbalance_pct only during newly_idle and busy load balancing. And during idle load balancing, check if there is an imbalance in number of idle cpu's across the busiest and this sched_group or if the busiest group has more tasks than its weight that the idle cpu in this_group can pull. Reported-by: Nikhil Rao <ncrao@google.com> Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> LKML-Reference: <1284760952.2676.11.camel@sbsiddha-MOBL3.sc.intel.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-09-17 22:02:32 +00:00
}
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
/* Try to move all excess tasks to child's sibling domain */
if (sds.prefer_sibling && local->group_type == group_has_spare &&
busiest->sum_nr_running > local->sum_nr_running + 1)
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
goto force_balance;
if (busiest->group_type != group_overloaded) {
if (env->idle == CPU_NOT_IDLE)
/*
* If the busiest group is not overloaded (and as a
* result the local one too) but this CPU is already
* busy, let another idle CPU try to pull task.
*/
goto out_balanced;
if (busiest->group_weight > 1 &&
local->idle_cpus <= (busiest->idle_cpus + 1))
/*
* If the busiest group is not overloaded
* and there is no imbalance between this and busiest
* group wrt idle CPUs, it is balanced. The imbalance
* becomes significant if the diff is greater than 1
* otherwise we might end up to just move the imbalance
* on another group. Of course this applies only if
* there is more than 1 CPU per group.
*/
goto out_balanced;
if (busiest->sum_h_nr_running == 1)
/*
* busiest doesn't have any tasks waiting to run
*/
goto out_balanced;
}
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
force_balance:
/* Looks like there is an imbalance. Compute it */
calculate_imbalance(env, &sds);
return env->imbalance ? sds.busiest : NULL;
out_balanced:
env->imbalance = 0;
return NULL;
}
/*
* find_busiest_queue - find the busiest runqueue among the CPUs in the group.
*/
static struct rq *find_busiest_queue(struct lb_env *env,
struct sched_group *group)
{
struct rq *busiest = NULL, *rq;
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
unsigned long busiest_util = 0, busiest_load = 0, busiest_capacity = 1;
unsigned int busiest_nr = 0;
int i;
for_each_cpu_and(i, sched_group_span(group), env->cpus) {
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
unsigned long capacity, load, util;
unsigned int nr_running;
enum fbq_type rt;
rq = cpu_rq(i);
rt = fbq_classify_rq(rq);
/*
* We classify groups/runqueues into three groups:
* - regular: there are !numa tasks
* - remote: there are numa tasks that run on the 'wrong' node
* - all: there is no distinction
*
* In order to avoid migrating ideally placed numa tasks,
* ignore those when there's better options.
*
* If we ignore the actual busiest queue to migrate another
* task, the next balance pass can still reduce the busiest
* queue by moving tasks around inside the node.
*
* If we cannot move enough load due to this classification
* the next pass will adjust the group classification and
* allow migration of more tasks.
*
* Both cases only affect the total convergence complexity.
*/
if (rt > env->fbq_type)
continue;
capacity = capacity_of(i);
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
nr_running = rq->cfs.h_nr_running;
/*
* For ASYM_CPUCAPACITY domains, don't pick a CPU that could
* eventually lead to active_balancing high->low capacity.
* Higher per-CPU capacity is considered better than balancing
* average load.
*/
if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
capacity_of(env->dst_cpu) < capacity &&
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
nr_running == 1)
continue;
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
switch (env->migration_type) {
case migrate_load:
/*
* When comparing with load imbalance, use cpu_load()
* which is not scaled with the CPU capacity.
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
*/
load = cpu_load(rq);
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
if (nr_running == 1 && load > env->imbalance &&
!check_cpu_capacity(rq, env->sd))
break;
sched: Replace capacity_factor by usage The scheduler tries to compute how many tasks a group of CPUs can handle by assuming that a task's load is SCHED_LOAD_SCALE and a CPU's capacity is SCHED_CAPACITY_SCALE. 'struct sg_lb_stats:group_capacity_factor' divides the capacity of the group by SCHED_LOAD_SCALE to estimate how many task can run in the group. Then, it compares this value with the sum of nr_running to decide if the group is overloaded or not. But the 'group_capacity_factor' concept is hardly working for SMT systems, it sometimes works for big cores but fails to do the right thing for little cores. Below are two examples to illustrate the problem that this patch solves: 1- If the original capacity of a CPU is less than SCHED_CAPACITY_SCALE (640 as an example), a group of 3 CPUS will have a max capacity_factor of 2 (div_round_closest(3x640/1024) = 2) which means that it will be seen as overloaded even if we have only one task per CPU. 2 - If the original capacity of a CPU is greater than SCHED_CAPACITY_SCALE (1512 as an example), a group of 4 CPUs will have a capacity_factor of 4 (at max and thanks to the fix [0] for SMT system that prevent the apparition of ghost CPUs) but if one CPU is fully used by rt tasks (and its capacity is reduced to nearly nothing), the capacity factor of the group will still be 4 (div_round_closest(3*1512/1024) = 5 which is cap to 4 with [0]). So, this patch tries to solve this issue by removing capacity_factor and replacing it with the 2 following metrics: - The available CPU's capacity for CFS tasks which is already used by load_balance(). - The usage of the CPU by the CFS tasks. For the latter, utilization_avg_contrib has been re-introduced to compute the usage of a CPU by CFS tasks. 'group_capacity_factor' and 'group_has_free_capacity' has been removed and replaced by 'group_no_capacity'. We compare the number of task with the number of CPUs and we evaluate the level of utilization of the CPUs to define if a group is overloaded or if a group has capacity to handle more tasks. For SD_PREFER_SIBLING, a group is tagged overloaded if it has more than 1 task so it will be selected in priority (among the overloaded groups). Since [1], SD_PREFER_SIBLING is no more concerned by the computation of 'load_above_capacity' because local is not overloaded. [1] 9a5d9ba6a363 ("sched/fair: Allow calculate_imbalance() to move idle cpus") Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Morten.Rasmussen@arm.com Cc: dietmar.eggemann@arm.com Cc: efault@gmx.de Cc: kamalesh@linux.vnet.ibm.com Cc: linaro-kernel@lists.linaro.org Cc: nicolas.pitre@linaro.org Cc: preeti@linux.vnet.ibm.com Cc: riel@redhat.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/1425052454-25797-9-git-send-email-vincent.guittot@linaro.org [ Tidied up the changelog. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-02-27 15:54:11 +00:00
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
/*
* For the load comparisons with the other CPUs,
* consider the cpu_load() scaled with the CPU
* capacity, so that the load can be moved away
* from the CPU that is potentially running at a
* lower capacity.
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
*
* Thus we're looking for max(load_i / capacity_i),
* crosswise multiplication to rid ourselves of the
* division works out to:
* load_i * capacity_j > load_j * capacity_i;
* where j is our previous maximum.
*/
if (load * busiest_capacity > busiest_load * capacity) {
busiest_load = load;
busiest_capacity = capacity;
busiest = rq;
}
break;
case migrate_util:
util = cpu_util(cpu_of(rq));
/*
* Don't try to pull utilization from a CPU with one
* running task. Whatever its utilization, we will fail
* detach the task.
*/
if (nr_running <= 1)
continue;
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
if (busiest_util < util) {
busiest_util = util;
busiest = rq;
}
break;
case migrate_task:
if (busiest_nr < nr_running) {
busiest_nr = nr_running;
busiest = rq;
}
break;
case migrate_misfit:
/*
* For ASYM_CPUCAPACITY domains with misfit tasks we
* simply seek the "biggest" misfit task.
*/
if (rq->misfit_task_load > busiest_load) {
busiest_load = rq->misfit_task_load;
busiest = rq;
}
break;
}
}
return busiest;
}
/*
* Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
* so long as it is large enough.
*/
#define MAX_PINNED_INTERVAL 512
static inline bool
asym_active_balance(struct lb_env *env)
{
/*
* ASYM_PACKING needs to force migrate tasks from busy but
* lower priority CPUs in order to pack all tasks in the
* highest priority CPUs.
*/
return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) &&
sched_asym_prefer(env->dst_cpu, env->src_cpu);
}
static inline bool
voluntary_active_balance(struct lb_env *env)
{
struct sched_domain *sd = env->sd;
sched: Add asymmetric group packing option for sibling domain Check to see if the group is packed in a sched doman. This is primarily intended to used at the sibling level. Some cores like POWER7 prefer to use lower numbered SMT threads. In the case of POWER7, it can move to lower SMT modes only when higher threads are idle. When in lower SMT modes, the threads will perform better since they share less core resources. Hence when we have idle threads, we want them to be the higher ones. This adds a hook into f_b_g() called check_asym_packing() to check the packing. This packing function is run on idle threads. It checks to see if the busiest CPU in this domain (core in the P7 case) has a higher CPU number than what where the packing function is being run on. If it is, calculate the imbalance and return the higher busier thread as the busiest group to f_b_g(). Here we are assuming a lower CPU number will be equivalent to a lower SMT thread number. It also creates a new SD_ASYM_PACKING flag to enable this feature at any scheduler domain level. It also creates an arch hook to enable this feature at the sibling level. The default function doesn't enable this feature. Based heavily on patch from Peter Zijlstra. Fixes from Srivatsa Vaddagiri. Signed-off-by: Michael Neuling <mikey@neuling.org> Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Thomas Gleixner <tglx@linutronix.de> LKML-Reference: <20100608045702.2936CCC897@localhost.localdomain> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-06-08 04:57:02 +00:00
if (asym_active_balance(env))
return 1;
sched: Move CFS tasks to CPUs with higher capacity When a CPU is used to handle a lot of IRQs or some RT tasks, the remaining capacity for CFS tasks can be significantly reduced. Once we detect such situation by comparing cpu_capacity_orig and cpu_capacity, we trig an idle load balance to check if it's worth moving its tasks on an idle CPU. It's worth trying to move the task before the CPU is fully utilized to minimize the preemption by irq or RT tasks. Once the idle load_balance has selected the busiest CPU, it will look for an active load balance for only two cases: - There is only 1 task on the busiest CPU. - We haven't been able to move a task of the busiest rq. A CPU with a reduced capacity is included in the 1st case, and it's worth to actively migrate its task if the idle CPU has got more available capacity for CFS tasks. This test has been added in need_active_balance. As a sidenote, this will not generate more spurious ilb because we already trig an ilb if there is more than 1 busy cpu. If this cpu is the only one that has a task, we will trig the ilb once for migrating the task. The nohz_kick_needed function has been cleaned up a bit while adding the new test env.src_cpu and env.src_rq must be set unconditionnally because they are used in need_active_balance which is called even if busiest->nr_running equals 1 Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Morten.Rasmussen@arm.com Cc: dietmar.eggemann@arm.com Cc: efault@gmx.de Cc: kamalesh@linux.vnet.ibm.com Cc: linaro-kernel@lists.linaro.org Cc: nicolas.pitre@linaro.org Cc: preeti@linux.vnet.ibm.com Cc: riel@redhat.com Link: http://lkml.kernel.org/r/1425052454-25797-12-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-02-27 15:54:14 +00:00
/*
* The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
* It's worth migrating the task if the src_cpu's capacity is reduced
* because of other sched_class or IRQs if more capacity stays
* available on dst_cpu.
*/
if ((env->idle != CPU_NOT_IDLE) &&
(env->src_rq->cfs.h_nr_running == 1)) {
if ((check_cpu_capacity(env->src_rq, sd)) &&
(capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
return 1;
}
sched/fair: Rework load_balance() The load_balance() algorithm contains some heuristics which have become meaningless since the rework of the scheduler's metrics like the introduction of PELT. Furthermore, load is an ill-suited metric for solving certain task placement imbalance scenarios. For instance, in the presence of idle CPUs, we should simply try to get at least one task per CPU, whereas the current load-based algorithm can actually leave idle CPUs alone simply because the load is somewhat balanced. The current algorithm ends up creating virtual and meaningless values like the avg_load_per_task or tweaks the state of a group to make it overloaded whereas it's not, in order to try to migrate tasks. load_balance() should better qualify the imbalance of the group and clearly define what has to be moved to fix this imbalance. The type of sched_group has been extended to better reflect the type of imbalance. We now have: group_has_spare group_fully_busy group_misfit_task group_asym_packing group_imbalanced group_overloaded Based on the type of sched_group, load_balance now sets what it wants to move in order to fix the imbalance. It can be some load as before but also some utilization, a number of task or a type of task: migrate_task migrate_util migrate_load migrate_misfit This new load_balance() algorithm fixes several pending wrong tasks placement: - the 1 task per CPU case with asymmetric system - the case of cfs task preempted by other class - the case of tasks not evenly spread on groups with spare capacity Also the load balance decisions have been consolidated in the 3 functions below after removing the few bypasses and hacks of the current code: - update_sd_pick_busiest() select the busiest sched_group. - find_busiest_group() checks if there is an imbalance between local and busiest group. - calculate_imbalance() decides what have to be moved. Finally, the now unused field total_running of struct sd_lb_stats has been removed. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Ben Segall <bsegall@google.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: hdanton@sina.com Cc: parth@linux.ibm.com Cc: pauld@redhat.com Cc: quentin.perret@arm.com Cc: riel@surriel.com Cc: srikar@linux.vnet.ibm.com Cc: valentin.schneider@arm.com Link: https://lkml.kernel.org/r/1571405198-27570-5-git-send-email-vincent.guittot@linaro.org [ Small readability and spelling updates. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-10-18 13:26:31 +00:00
if (env->migration_type == migrate_misfit)
sched/fair: Consider misfit tasks when load-balancing On asymmetric CPU capacity systems load intensive tasks can end up on CPUs that don't suit their compute demand. In this scenarios 'misfit' tasks should be migrated to CPUs with higher compute capacity to ensure better throughput. group_misfit_task indicates this scenario, but tweaks to the load-balance code are needed to make the migrations happen. Misfit balancing only makes sense between a source group of lower per-CPU capacity and destination group of higher compute capacity. Otherwise, misfit balancing is ignored. group_misfit_task has lowest priority so any imbalance due to overload is dealt with first. The modifications are: 1. Only pick a group containing misfit tasks as the busiest group if the destination group has higher capacity and has spare capacity. 2. When the busiest group is a 'misfit' group, skip the usual average load and group capacity checks. 3. Set the imbalance for 'misfit' balancing sufficiently high for a task to be pulled ignoring average load. 4. Pick the CPU with the highest misfit load as the source CPU. 5. If the misfit task is alone on the source CPU, go for active balancing. Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: gaku.inami.xh@renesas.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1530699470-29808-5-git-send-email-morten.rasmussen@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-07-04 10:17:42 +00:00
return 1;
return 0;
}
static int need_active_balance(struct lb_env *env)
{
struct sched_domain *sd = env->sd;
if (voluntary_active_balance(env))
return 1;
return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
}
sched: replace migration_thread with cpu_stop Currently migration_thread is serving three purposes - migration pusher, context to execute active_load_balance() and forced context switcher for expedited RCU synchronize_sched. All three roles are hardcoded into migration_thread() and determining which job is scheduled is slightly messy. This patch kills migration_thread and replaces all three uses with cpu_stop. The three different roles of migration_thread() are splitted into three separate cpu_stop callbacks - migration_cpu_stop(), active_load_balance_cpu_stop() and synchronize_sched_expedited_cpu_stop() - and each use case now simply asks cpu_stop to execute the callback as necessary. synchronize_sched_expedited() was implemented with private preallocated resources and custom multi-cpu queueing and waiting logic, both of which are provided by cpu_stop. synchronize_sched_expedited_count is made atomic and all other shared resources along with the mutex are dropped. synchronize_sched_expedited() also implemented a check to detect cases where not all the callback got executed on their assigned cpus and fall back to synchronize_sched(). If called with cpu hotplug blocked, cpu_stop already guarantees that and the condition cannot happen; otherwise, stop_machine() would break. However, this patch preserves the paranoid check using a cpumask to record on which cpus the stopper ran so that it can serve as a bisection point if something actually goes wrong theree. Because the internal execution state is no longer visible, rcu_expedited_torture_stats() is removed. This patch also renames cpu_stop threads to from "stopper/%d" to "migration/%d". The names of these threads ultimately don't matter and there's no reason to make unnecessary userland visible changes. With this patch applied, stop_machine() and sched now share the same resources. stop_machine() is faster without wasting any resources and sched migration users are much cleaner. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Peter Zijlstra <peterz@infradead.org> Cc: Ingo Molnar <mingo@elte.hu> Cc: Dipankar Sarma <dipankar@in.ibm.com> Cc: Josh Triplett <josh@freedesktop.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Dimitri Sivanich <sivanich@sgi.com>
2010-05-06 16:49:21 +00:00
static int active_load_balance_cpu_stop(void *data);
static int should_we_balance(struct lb_env *env)
{
struct sched_group *sg = env->sd->groups;
int cpu, balance_cpu = -1;
/*
* Ensure the balancing environment is consistent; can happen
* when the softirq triggers 'during' hotplug.
*/
if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
return 0;
/*
* In the newly idle case, we will allow all the CPUs
* to do the newly idle load balance.
*/
if (env->idle == CPU_NEWLY_IDLE)
return 1;
/* Try to find first idle CPU */
for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
if (!idle_cpu(cpu))
continue;
balance_cpu = cpu;
break;
}
if (balance_cpu == -1)
balance_cpu = group_balance_cpu(sg);
/*
* First idle CPU or the first CPU(busiest) in this sched group
* is eligible for doing load balancing at this and above domains.
*/
return balance_cpu == env->dst_cpu;
}
/*
* Check this_cpu to ensure it is balanced within domain. Attempt to move
* tasks if there is an imbalance.
*/
static int load_balance(int this_cpu, struct rq *this_rq,
struct sched_domain *sd, enum cpu_idle_type idle,
int *continue_balancing)
{
sched: Improve balance_cpu() to consider other cpus in its group as target of (pinned) task Current load balance scheme requires only one cpu in a sched_group (balance_cpu) to look at other peer sched_groups for imbalance and pull tasks towards itself from a busy cpu. Tasks thus pulled by balance_cpu could later get picked up by cpus that are in the same sched_group as that of balance_cpu. This scheme however fails to pull tasks that are not allowed to run on balance_cpu (but are allowed to run on other cpus in its sched_group). That can affect fairness and in some worst case scenarios cause starvation. Consider a two core (2 threads/core) system running tasks as below: Core0 Core1 / \ / \ C0 C1 C2 C3 | | | | v v v v F0 T1 F1 [idle] T2 F0 = SCHED_FIFO task (pinned to C0) F1 = SCHED_FIFO task (pinned to C2) T1 = SCHED_OTHER task (pinned to C1) T2 = SCHED_OTHER task (pinned to C1 and C2) F1 could become a cpu hog, which will starve T2 unless C1 pulls it. Between C0 and C1 however, C0 is required to look for imbalance between cores, which will fail to pull T2 towards Core0. T2 will starve eternally in this case. The same scenario can arise in presence of non-rt tasks as well (say we replace F1 with high irq load). We tackle this problem by having balance_cpu move pinned tasks to one of its sibling cpus (where they can run). We first check if load balance goal can be met by ignoring pinned tasks, failing which we retry move_tasks() with a new env->dst_cpu. This patch modifies load balance semantics on who can move load towards a given cpu in a given sched_domain. Before this patch, a given_cpu or a ilb_cpu acting on behalf of an idle given_cpu is responsible for moving load to given_cpu. With this patch applied, balance_cpu can in addition decide on moving some load to a given_cpu. There is a remote possibility that excess load could get moved as a result of this (balance_cpu and given_cpu/ilb_cpu deciding *independently* and at *same* time to move some load to a given_cpu). However we should see less of such conflicting decisions in practice and moreover subsequent load balance cycles should correct the excess load moved to given_cpu. Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Signed-off-by: Prashanth Nageshappa <prashanth@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/4FE06CDB.2060605@linux.vnet.ibm.com [ minor edits ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2012-06-19 12:13:15 +00:00
int ld_moved, cur_ld_moved, active_balance = 0;
struct sched_domain *sd_parent = sd->parent;
struct sched_group *group;
struct rq *busiest;
struct rq_flags rf;
struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
struct lb_env env = {
.sd = sd,
.dst_cpu = this_cpu,
.dst_rq = this_rq,
.dst_grpmask = sched_group_span(sd->groups),
.idle = idle,
.loop_break = sched_nr_migrate_break,
.cpus = cpus,
.fbq_type = all,
.tasks = LIST_HEAD_INIT(env.tasks),
};
sched/fair: Fix load_balance() affinity redo path If load_balance() fails to migrate any tasks because all tasks were affined, load_balance() removes the source CPU from consideration and attempts to redo and balance among the new subset of CPUs. There is a bug in this code path where the algorithm considers all active CPUs in the system (minus the source that was just masked out). This is not valid for two reasons: some active CPUs may not be in the current scheduling domain and one of the active CPUs is dst_cpu. These CPUs should not be considered, as we cannot pull load from them. Instead of failing out of load_balance(), we may end up redoing the search with no valid CPUs and incorrectly concluding the domain is balanced. Additionally, if the group_imbalance flag was just set, it may also be incorrectly unset, thus the flag will not be seen by other CPUs in future load_balance() runs as that algorithm intends. Fix the check by removing CPUs not in the current domain and the dst_cpu from considertation, thus limiting the evaluation to valid remaining CPUs from which load might be migrated. Co-authored-by: Austin Christ <austinwc@codeaurora.org> Co-authored-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Tested-by: Tyler Baicar <tbaicar@codeaurora.org> Signed-off-by: Jeffrey Hugo <jhugo@codeaurora.org> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Austin Christ <austinwc@codeaurora.org> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Timur Tabi <timur@codeaurora.org> Link: http://lkml.kernel.org/r/1496863138-11322-2-git-send-email-jhugo@codeaurora.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-06-07 19:18:57 +00:00
cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
schedstat_inc(sd->lb_count[idle]);
redo:
if (!should_we_balance(&env)) {
*continue_balancing = 0;
goto out_balanced;
}
group = find_busiest_group(&env);
if (!group) {
schedstat_inc(sd->lb_nobusyg[idle]);
goto out_balanced;
}
busiest = find_busiest_queue(&env, group);
if (!busiest) {
schedstat_inc(sd->lb_nobusyq[idle]);
goto out_balanced;
}
BUG_ON(busiest == env.dst_rq);
schedstat_add(sd->lb_imbalance[idle], env.imbalance);
sched: Move CFS tasks to CPUs with higher capacity When a CPU is used to handle a lot of IRQs or some RT tasks, the remaining capacity for CFS tasks can be significantly reduced. Once we detect such situation by comparing cpu_capacity_orig and cpu_capacity, we trig an idle load balance to check if it's worth moving its tasks on an idle CPU. It's worth trying to move the task before the CPU is fully utilized to minimize the preemption by irq or RT tasks. Once the idle load_balance has selected the busiest CPU, it will look for an active load balance for only two cases: - There is only 1 task on the busiest CPU. - We haven't been able to move a task of the busiest rq. A CPU with a reduced capacity is included in the 1st case, and it's worth to actively migrate its task if the idle CPU has got more available capacity for CFS tasks. This test has been added in need_active_balance. As a sidenote, this will not generate more spurious ilb because we already trig an ilb if there is more than 1 busy cpu. If this cpu is the only one that has a task, we will trig the ilb once for migrating the task. The nohz_kick_needed function has been cleaned up a bit while adding the new test env.src_cpu and env.src_rq must be set unconditionnally because they are used in need_active_balance which is called even if busiest->nr_running equals 1 Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Morten.Rasmussen@arm.com Cc: dietmar.eggemann@arm.com Cc: efault@gmx.de Cc: kamalesh@linux.vnet.ibm.com Cc: linaro-kernel@lists.linaro.org Cc: nicolas.pitre@linaro.org Cc: preeti@linux.vnet.ibm.com Cc: riel@redhat.com Link: http://lkml.kernel.org/r/1425052454-25797-12-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-02-27 15:54:14 +00:00
env.src_cpu = busiest->cpu;
env.src_rq = busiest;
ld_moved = 0;
if (busiest->nr_running > 1) {
/*
* Attempt to move tasks. If find_busiest_group has found
* an imbalance but busiest->nr_running <= 1, the group is
* still unbalanced. ld_moved simply stays zero, so it is
* correctly treated as an imbalance.
*/
env.flags |= LBF_ALL_PINNED;
env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
more_balance:
rq_lock_irqsave(busiest, &rf);
update_rq_clock(busiest);
sched: Improve balance_cpu() to consider other cpus in its group as target of (pinned) task Current load balance scheme requires only one cpu in a sched_group (balance_cpu) to look at other peer sched_groups for imbalance and pull tasks towards itself from a busy cpu. Tasks thus pulled by balance_cpu could later get picked up by cpus that are in the same sched_group as that of balance_cpu. This scheme however fails to pull tasks that are not allowed to run on balance_cpu (but are allowed to run on other cpus in its sched_group). That can affect fairness and in some worst case scenarios cause starvation. Consider a two core (2 threads/core) system running tasks as below: Core0 Core1 / \ / \ C0 C1 C2 C3 | | | | v v v v F0 T1 F1 [idle] T2 F0 = SCHED_FIFO task (pinned to C0) F1 = SCHED_FIFO task (pinned to C2) T1 = SCHED_OTHER task (pinned to C1) T2 = SCHED_OTHER task (pinned to C1 and C2) F1 could become a cpu hog, which will starve T2 unless C1 pulls it. Between C0 and C1 however, C0 is required to look for imbalance between cores, which will fail to pull T2 towards Core0. T2 will starve eternally in this case. The same scenario can arise in presence of non-rt tasks as well (say we replace F1 with high irq load). We tackle this problem by having balance_cpu move pinned tasks to one of its sibling cpus (where they can run). We first check if load balance goal can be met by ignoring pinned tasks, failing which we retry move_tasks() with a new env->dst_cpu. This patch modifies load balance semantics on who can move load towards a given cpu in a given sched_domain. Before this patch, a given_cpu or a ilb_cpu acting on behalf of an idle given_cpu is responsible for moving load to given_cpu. With this patch applied, balance_cpu can in addition decide on moving some load to a given_cpu. There is a remote possibility that excess load could get moved as a result of this (balance_cpu and given_cpu/ilb_cpu deciding *independently* and at *same* time to move some load to a given_cpu). However we should see less of such conflicting decisions in practice and moreover subsequent load balance cycles should correct the excess load moved to given_cpu. Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Signed-off-by: Prashanth Nageshappa <prashanth@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/4FE06CDB.2060605@linux.vnet.ibm.com [ minor edits ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2012-06-19 12:13:15 +00:00
/*
* cur_ld_moved - load moved in current iteration
* ld_moved - cumulative load moved across iterations
*/
cur_ld_moved = detach_tasks(&env);
/*
* We've detached some tasks from busiest_rq. Every
* task is masked "TASK_ON_RQ_MIGRATING", so we can safely
* unlock busiest->lock, and we are able to be sure
* that nobody can manipulate the tasks in parallel.
* See task_rq_lock() family for the details.
*/
rq_unlock(busiest, &rf);
if (cur_ld_moved) {
attach_tasks(&env);
ld_moved += cur_ld_moved;
}
local_irq_restore(rf.flags);
sched: Improve balance_cpu() to consider other cpus in its group as target of (pinned) task Current load balance scheme requires only one cpu in a sched_group (balance_cpu) to look at other peer sched_groups for imbalance and pull tasks towards itself from a busy cpu. Tasks thus pulled by balance_cpu could later get picked up by cpus that are in the same sched_group as that of balance_cpu. This scheme however fails to pull tasks that are not allowed to run on balance_cpu (but are allowed to run on other cpus in its sched_group). That can affect fairness and in some worst case scenarios cause starvation. Consider a two core (2 threads/core) system running tasks as below: Core0 Core1 / \ / \ C0 C1 C2 C3 | | | | v v v v F0 T1 F1 [idle] T2 F0 = SCHED_FIFO task (pinned to C0) F1 = SCHED_FIFO task (pinned to C2) T1 = SCHED_OTHER task (pinned to C1) T2 = SCHED_OTHER task (pinned to C1 and C2) F1 could become a cpu hog, which will starve T2 unless C1 pulls it. Between C0 and C1 however, C0 is required to look for imbalance between cores, which will fail to pull T2 towards Core0. T2 will starve eternally in this case. The same scenario can arise in presence of non-rt tasks as well (say we replace F1 with high irq load). We tackle this problem by having balance_cpu move pinned tasks to one of its sibling cpus (where they can run). We first check if load balance goal can be met by ignoring pinned tasks, failing which we retry move_tasks() with a new env->dst_cpu. This patch modifies load balance semantics on who can move load towards a given cpu in a given sched_domain. Before this patch, a given_cpu or a ilb_cpu acting on behalf of an idle given_cpu is responsible for moving load to given_cpu. With this patch applied, balance_cpu can in addition decide on moving some load to a given_cpu. There is a remote possibility that excess load could get moved as a result of this (balance_cpu and given_cpu/ilb_cpu deciding *independently* and at *same* time to move some load to a given_cpu). However we should see less of such conflicting decisions in practice and moreover subsequent load balance cycles should correct the excess load moved to given_cpu. Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Signed-off-by: Prashanth Nageshappa <prashanth@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/4FE06CDB.2060605@linux.vnet.ibm.com [ minor edits ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2012-06-19 12:13:15 +00:00
if (env.flags & LBF_NEED_BREAK) {
env.flags &= ~LBF_NEED_BREAK;
goto more_balance;
}
sched: Improve balance_cpu() to consider other cpus in its group as target of (pinned) task Current load balance scheme requires only one cpu in a sched_group (balance_cpu) to look at other peer sched_groups for imbalance and pull tasks towards itself from a busy cpu. Tasks thus pulled by balance_cpu could later get picked up by cpus that are in the same sched_group as that of balance_cpu. This scheme however fails to pull tasks that are not allowed to run on balance_cpu (but are allowed to run on other cpus in its sched_group). That can affect fairness and in some worst case scenarios cause starvation. Consider a two core (2 threads/core) system running tasks as below: Core0 Core1 / \ / \ C0 C1 C2 C3 | | | | v v v v F0 T1 F1 [idle] T2 F0 = SCHED_FIFO task (pinned to C0) F1 = SCHED_FIFO task (pinned to C2) T1 = SCHED_OTHER task (pinned to C1) T2 = SCHED_OTHER task (pinned to C1 and C2) F1 could become a cpu hog, which will starve T2 unless C1 pulls it. Between C0 and C1 however, C0 is required to look for imbalance between cores, which will fail to pull T2 towards Core0. T2 will starve eternally in this case. The same scenario can arise in presence of non-rt tasks as well (say we replace F1 with high irq load). We tackle this problem by having balance_cpu move pinned tasks to one of its sibling cpus (where they can run). We first check if load balance goal can be met by ignoring pinned tasks, failing which we retry move_tasks() with a new env->dst_cpu. This patch modifies load balance semantics on who can move load towards a given cpu in a given sched_domain. Before this patch, a given_cpu or a ilb_cpu acting on behalf of an idle given_cpu is responsible for moving load to given_cpu. With this patch applied, balance_cpu can in addition decide on moving some load to a given_cpu. There is a remote possibility that excess load could get moved as a result of this (balance_cpu and given_cpu/ilb_cpu deciding *independently* and at *same* time to move some load to a given_cpu). However we should see less of such conflicting decisions in practice and moreover subsequent load balance cycles should correct the excess load moved to given_cpu. Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Signed-off-by: Prashanth Nageshappa <prashanth@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/4FE06CDB.2060605@linux.vnet.ibm.com [ minor edits ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2012-06-19 12:13:15 +00:00
/*
* Revisit (affine) tasks on src_cpu that couldn't be moved to
* us and move them to an alternate dst_cpu in our sched_group
* where they can run. The upper limit on how many times we
* iterate on same src_cpu is dependent on number of CPUs in our
sched: Improve balance_cpu() to consider other cpus in its group as target of (pinned) task Current load balance scheme requires only one cpu in a sched_group (balance_cpu) to look at other peer sched_groups for imbalance and pull tasks towards itself from a busy cpu. Tasks thus pulled by balance_cpu could later get picked up by cpus that are in the same sched_group as that of balance_cpu. This scheme however fails to pull tasks that are not allowed to run on balance_cpu (but are allowed to run on other cpus in its sched_group). That can affect fairness and in some worst case scenarios cause starvation. Consider a two core (2 threads/core) system running tasks as below: Core0 Core1 / \ / \ C0 C1 C2 C3 | | | | v v v v F0 T1 F1 [idle] T2 F0 = SCHED_FIFO task (pinned to C0) F1 = SCHED_FIFO task (pinned to C2) T1 = SCHED_OTHER task (pinned to C1) T2 = SCHED_OTHER task (pinned to C1 and C2) F1 could become a cpu hog, which will starve T2 unless C1 pulls it. Between C0 and C1 however, C0 is required to look for imbalance between cores, which will fail to pull T2 towards Core0. T2 will starve eternally in this case. The same scenario can arise in presence of non-rt tasks as well (say we replace F1 with high irq load). We tackle this problem by having balance_cpu move pinned tasks to one of its sibling cpus (where they can run). We first check if load balance goal can be met by ignoring pinned tasks, failing which we retry move_tasks() with a new env->dst_cpu. This patch modifies load balance semantics on who can move load towards a given cpu in a given sched_domain. Before this patch, a given_cpu or a ilb_cpu acting on behalf of an idle given_cpu is responsible for moving load to given_cpu. With this patch applied, balance_cpu can in addition decide on moving some load to a given_cpu. There is a remote possibility that excess load could get moved as a result of this (balance_cpu and given_cpu/ilb_cpu deciding *independently* and at *same* time to move some load to a given_cpu). However we should see less of such conflicting decisions in practice and moreover subsequent load balance cycles should correct the excess load moved to given_cpu. Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Signed-off-by: Prashanth Nageshappa <prashanth@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/4FE06CDB.2060605@linux.vnet.ibm.com [ minor edits ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2012-06-19 12:13:15 +00:00
* sched_group.
*
* This changes load balance semantics a bit on who can move
* load to a given_cpu. In addition to the given_cpu itself
* (or a ilb_cpu acting on its behalf where given_cpu is
* nohz-idle), we now have balance_cpu in a position to move
* load to given_cpu. In rare situations, this may cause
* conflicts (balance_cpu and given_cpu/ilb_cpu deciding
* _independently_ and at _same_ time to move some load to
* given_cpu) causing exceess load to be moved to given_cpu.
* This however should not happen so much in practice and
* moreover subsequent load balance cycles should correct the
* excess load moved.
*/
if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
sched: Improve balance_cpu() to consider other cpus in its group as target of (pinned) task Current load balance scheme requires only one cpu in a sched_group (balance_cpu) to look at other peer sched_groups for imbalance and pull tasks towards itself from a busy cpu. Tasks thus pulled by balance_cpu could later get picked up by cpus that are in the same sched_group as that of balance_cpu. This scheme however fails to pull tasks that are not allowed to run on balance_cpu (but are allowed to run on other cpus in its sched_group). That can affect fairness and in some worst case scenarios cause starvation. Consider a two core (2 threads/core) system running tasks as below: Core0 Core1 / \ / \ C0 C1 C2 C3 | | | | v v v v F0 T1 F1 [idle] T2 F0 = SCHED_FIFO task (pinned to C0) F1 = SCHED_FIFO task (pinned to C2) T1 = SCHED_OTHER task (pinned to C1) T2 = SCHED_OTHER task (pinned to C1 and C2) F1 could become a cpu hog, which will starve T2 unless C1 pulls it. Between C0 and C1 however, C0 is required to look for imbalance between cores, which will fail to pull T2 towards Core0. T2 will starve eternally in this case. The same scenario can arise in presence of non-rt tasks as well (say we replace F1 with high irq load). We tackle this problem by having balance_cpu move pinned tasks to one of its sibling cpus (where they can run). We first check if load balance goal can be met by ignoring pinned tasks, failing which we retry move_tasks() with a new env->dst_cpu. This patch modifies load balance semantics on who can move load towards a given cpu in a given sched_domain. Before this patch, a given_cpu or a ilb_cpu acting on behalf of an idle given_cpu is responsible for moving load to given_cpu. With this patch applied, balance_cpu can in addition decide on moving some load to a given_cpu. There is a remote possibility that excess load could get moved as a result of this (balance_cpu and given_cpu/ilb_cpu deciding *independently* and at *same* time to move some load to a given_cpu). However we should see less of such conflicting decisions in practice and moreover subsequent load balance cycles should correct the excess load moved to given_cpu. Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Signed-off-by: Prashanth Nageshappa <prashanth@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/4FE06CDB.2060605@linux.vnet.ibm.com [ minor edits ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2012-06-19 12:13:15 +00:00
/* Prevent to re-select dst_cpu via env's CPUs */
__cpumask_clear_cpu(env.dst_cpu, env.cpus);
env.dst_rq = cpu_rq(env.new_dst_cpu);
sched: Improve balance_cpu() to consider other cpus in its group as target of (pinned) task Current load balance scheme requires only one cpu in a sched_group (balance_cpu) to look at other peer sched_groups for imbalance and pull tasks towards itself from a busy cpu. Tasks thus pulled by balance_cpu could later get picked up by cpus that are in the same sched_group as that of balance_cpu. This scheme however fails to pull tasks that are not allowed to run on balance_cpu (but are allowed to run on other cpus in its sched_group). That can affect fairness and in some worst case scenarios cause starvation. Consider a two core (2 threads/core) system running tasks as below: Core0 Core1 / \ / \ C0 C1 C2 C3 | | | | v v v v F0 T1 F1 [idle] T2 F0 = SCHED_FIFO task (pinned to C0) F1 = SCHED_FIFO task (pinned to C2) T1 = SCHED_OTHER task (pinned to C1) T2 = SCHED_OTHER task (pinned to C1 and C2) F1 could become a cpu hog, which will starve T2 unless C1 pulls it. Between C0 and C1 however, C0 is required to look for imbalance between cores, which will fail to pull T2 towards Core0. T2 will starve eternally in this case. The same scenario can arise in presence of non-rt tasks as well (say we replace F1 with high irq load). We tackle this problem by having balance_cpu move pinned tasks to one of its sibling cpus (where they can run). We first check if load balance goal can be met by ignoring pinned tasks, failing which we retry move_tasks() with a new env->dst_cpu. This patch modifies load balance semantics on who can move load towards a given cpu in a given sched_domain. Before this patch, a given_cpu or a ilb_cpu acting on behalf of an idle given_cpu is responsible for moving load to given_cpu. With this patch applied, balance_cpu can in addition decide on moving some load to a given_cpu. There is a remote possibility that excess load could get moved as a result of this (balance_cpu and given_cpu/ilb_cpu deciding *independently* and at *same* time to move some load to a given_cpu). However we should see less of such conflicting decisions in practice and moreover subsequent load balance cycles should correct the excess load moved to given_cpu. Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Signed-off-by: Prashanth Nageshappa <prashanth@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/4FE06CDB.2060605@linux.vnet.ibm.com [ minor edits ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2012-06-19 12:13:15 +00:00
env.dst_cpu = env.new_dst_cpu;
env.flags &= ~LBF_DST_PINNED;
sched: Improve balance_cpu() to consider other cpus in its group as target of (pinned) task Current load balance scheme requires only one cpu in a sched_group (balance_cpu) to look at other peer sched_groups for imbalance and pull tasks towards itself from a busy cpu. Tasks thus pulled by balance_cpu could later get picked up by cpus that are in the same sched_group as that of balance_cpu. This scheme however fails to pull tasks that are not allowed to run on balance_cpu (but are allowed to run on other cpus in its sched_group). That can affect fairness and in some worst case scenarios cause starvation. Consider a two core (2 threads/core) system running tasks as below: Core0 Core1 / \ / \ C0 C1 C2 C3 | | | | v v v v F0 T1 F1 [idle] T2 F0 = SCHED_FIFO task (pinned to C0) F1 = SCHED_FIFO task (pinned to C2) T1 = SCHED_OTHER task (pinned to C1) T2 = SCHED_OTHER task (pinned to C1 and C2) F1 could become a cpu hog, which will starve T2 unless C1 pulls it. Between C0 and C1 however, C0 is required to look for imbalance between cores, which will fail to pull T2 towards Core0. T2 will starve eternally in this case. The same scenario can arise in presence of non-rt tasks as well (say we replace F1 with high irq load). We tackle this problem by having balance_cpu move pinned tasks to one of its sibling cpus (where they can run). We first check if load balance goal can be met by ignoring pinned tasks, failing which we retry move_tasks() with a new env->dst_cpu. This patch modifies load balance semantics on who can move load towards a given cpu in a given sched_domain. Before this patch, a given_cpu or a ilb_cpu acting on behalf of an idle given_cpu is responsible for moving load to given_cpu. With this patch applied, balance_cpu can in addition decide on moving some load to a given_cpu. There is a remote possibility that excess load could get moved as a result of this (balance_cpu and given_cpu/ilb_cpu deciding *independently* and at *same* time to move some load to a given_cpu). However we should see less of such conflicting decisions in practice and moreover subsequent load balance cycles should correct the excess load moved to given_cpu. Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Signed-off-by: Prashanth Nageshappa <prashanth@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/4FE06CDB.2060605@linux.vnet.ibm.com [ minor edits ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2012-06-19 12:13:15 +00:00
env.loop = 0;
env.loop_break = sched_nr_migrate_break;
sched: Improve balance_cpu() to consider other cpus in its group as target of (pinned) task Current load balance scheme requires only one cpu in a sched_group (balance_cpu) to look at other peer sched_groups for imbalance and pull tasks towards itself from a busy cpu. Tasks thus pulled by balance_cpu could later get picked up by cpus that are in the same sched_group as that of balance_cpu. This scheme however fails to pull tasks that are not allowed to run on balance_cpu (but are allowed to run on other cpus in its sched_group). That can affect fairness and in some worst case scenarios cause starvation. Consider a two core (2 threads/core) system running tasks as below: Core0 Core1 / \ / \ C0 C1 C2 C3 | | | | v v v v F0 T1 F1 [idle] T2 F0 = SCHED_FIFO task (pinned to C0) F1 = SCHED_FIFO task (pinned to C2) T1 = SCHED_OTHER task (pinned to C1) T2 = SCHED_OTHER task (pinned to C1 and C2) F1 could become a cpu hog, which will starve T2 unless C1 pulls it. Between C0 and C1 however, C0 is required to look for imbalance between cores, which will fail to pull T2 towards Core0. T2 will starve eternally in this case. The same scenario can arise in presence of non-rt tasks as well (say we replace F1 with high irq load). We tackle this problem by having balance_cpu move pinned tasks to one of its sibling cpus (where they can run). We first check if load balance goal can be met by ignoring pinned tasks, failing which we retry move_tasks() with a new env->dst_cpu. This patch modifies load balance semantics on who can move load towards a given cpu in a given sched_domain. Before this patch, a given_cpu or a ilb_cpu acting on behalf of an idle given_cpu is responsible for moving load to given_cpu. With this patch applied, balance_cpu can in addition decide on moving some load to a given_cpu. There is a remote possibility that excess load could get moved as a result of this (balance_cpu and given_cpu/ilb_cpu deciding *independently* and at *same* time to move some load to a given_cpu). However we should see less of such conflicting decisions in practice and moreover subsequent load balance cycles should correct the excess load moved to given_cpu. Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Signed-off-by: Prashanth Nageshappa <prashanth@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/4FE06CDB.2060605@linux.vnet.ibm.com [ minor edits ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2012-06-19 12:13:15 +00:00
/*
* Go back to "more_balance" rather than "redo" since we
* need to continue with same src_cpu.
*/
goto more_balance;
}
/*
* We failed to reach balance because of affinity.
*/
if (sd_parent) {
int *group_imbalance = &sd_parent->groups->sgc->imbalance;
sched: Fix imbalance flag reset The imbalance flag can stay set whereas there is no imbalance. Let assume that we have 3 tasks that run on a dual cores /dual cluster system. We will have some idle load balance which are triggered during tick. Unfortunately, the tick is also used to queue background work so we can reach the situation where short work has been queued on a CPU which already runs a task. The load balance will detect this imbalance (2 tasks on 1 CPU and an idle CPU) and will try to pull the waiting task on the idle CPU. The waiting task is a worker thread that is pinned on a CPU so an imbalance due to pinned task is detected and the imbalance flag is set. Then, we will not be able to clear the flag because we have at most 1 task on each CPU but the imbalance flag will trig to useless active load balance between the idle CPU and the busy CPU. We need to reset of the imbalance flag as soon as we have reached a balanced state. If all tasks are pinned, we don't consider that as a balanced state and let the imbalance flag set. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Reviewed-by: Preeti U Murthy <preeti@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: riel@redhat.com Cc: Morten.Rasmussen@arm.com Cc: efault@gmx.de Cc: nicolas.pitre@linaro.org Cc: daniel.lezcano@linaro.org Cc: dietmar.eggemann@arm.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/1409051215-16788-2-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-08-26 11:06:44 +00:00
if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
*group_imbalance = 1;
}
/* All tasks on this runqueue were pinned by CPU affinity */
if (unlikely(env.flags & LBF_ALL_PINNED)) {
__cpumask_clear_cpu(cpu_of(busiest), cpus);
sched/fair: Fix load_balance() affinity redo path If load_balance() fails to migrate any tasks because all tasks were affined, load_balance() removes the source CPU from consideration and attempts to redo and balance among the new subset of CPUs. There is a bug in this code path where the algorithm considers all active CPUs in the system (minus the source that was just masked out). This is not valid for two reasons: some active CPUs may not be in the current scheduling domain and one of the active CPUs is dst_cpu. These CPUs should not be considered, as we cannot pull load from them. Instead of failing out of load_balance(), we may end up redoing the search with no valid CPUs and incorrectly concluding the domain is balanced. Additionally, if the group_imbalance flag was just set, it may also be incorrectly unset, thus the flag will not be seen by other CPUs in future load_balance() runs as that algorithm intends. Fix the check by removing CPUs not in the current domain and the dst_cpu from considertation, thus limiting the evaluation to valid remaining CPUs from which load might be migrated. Co-authored-by: Austin Christ <austinwc@codeaurora.org> Co-authored-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Tested-by: Tyler Baicar <tbaicar@codeaurora.org> Signed-off-by: Jeffrey Hugo <jhugo@codeaurora.org> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Austin Christ <austinwc@codeaurora.org> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Timur Tabi <timur@codeaurora.org> Link: http://lkml.kernel.org/r/1496863138-11322-2-git-send-email-jhugo@codeaurora.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-06-07 19:18:57 +00:00
/*
* Attempting to continue load balancing at the current
* sched_domain level only makes sense if there are
* active CPUs remaining as possible busiest CPUs to
* pull load from which are not contained within the
* destination group that is receiving any migrated
* load.
*/
if (!cpumask_subset(cpus, env.dst_grpmask)) {
env.loop = 0;
env.loop_break = sched_nr_migrate_break;
goto redo;
}
sched: Fix imbalance flag reset The imbalance flag can stay set whereas there is no imbalance. Let assume that we have 3 tasks that run on a dual cores /dual cluster system. We will have some idle load balance which are triggered during tick. Unfortunately, the tick is also used to queue background work so we can reach the situation where short work has been queued on a CPU which already runs a task. The load balance will detect this imbalance (2 tasks on 1 CPU and an idle CPU) and will try to pull the waiting task on the idle CPU. The waiting task is a worker thread that is pinned on a CPU so an imbalance due to pinned task is detected and the imbalance flag is set. Then, we will not be able to clear the flag because we have at most 1 task on each CPU but the imbalance flag will trig to useless active load balance between the idle CPU and the busy CPU. We need to reset of the imbalance flag as soon as we have reached a balanced state. If all tasks are pinned, we don't consider that as a balanced state and let the imbalance flag set. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Reviewed-by: Preeti U Murthy <preeti@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: riel@redhat.com Cc: Morten.Rasmussen@arm.com Cc: efault@gmx.de Cc: nicolas.pitre@linaro.org Cc: daniel.lezcano@linaro.org Cc: dietmar.eggemann@arm.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/1409051215-16788-2-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-08-26 11:06:44 +00:00
goto out_all_pinned;
}
}
if (!ld_moved) {
schedstat_inc(sd->lb_failed[idle]);
sched: Increment cache_nice_tries only on periodic lb scheduler uses cache_nice_tries as an indicator to do cache_hot and active load balance, when normal load balance fails. Currently, this value is changed on any failed load balance attempt. That ends up being not so nice to workloads that enter/exit idle often, as they do more frequent new_idle balance and that pretty soon results in cache hot tasks being pulled in. Making the cache_nice_tries ignore failed new_idle balance seems to make better sense. With that only the failed load balance in periodic load balance gets accounted and the rate of accumulation of cache_nice_tries will not depend on idle entry/exit (short running sleep-wakeup kind of tasks). This reduces movement of cache_hot tasks. schedstat diff (after-before) excerpt from a workload that has frequent and short wakeup-idle pattern (:2 in cpu col below refers to NEWIDLE idx) This snapshot was across ~400 seconds. Without this change: domainstats: domain0 cpu cnt bln fld imb gain hgain nobusyq nobusyg 0:2 306487 219575 73167 110069413 44583 19070 1172 218403 1:2 292139 194853 81421 120893383 50745 21902 1259 193594 2:2 283166 174607 91359 129699642 54931 23688 1287 173320 3:2 273998 161788 93991 132757146 57122 24351 1366 160422 4:2 289851 215692 62190 83398383 36377 13680 851 214841 5:2 316312 222146 77605 117582154 49948 20281 988 221158 6:2 297172 195596 83623 122133390 52801 21301 929 194667 7:2 283391 178078 86378 126622761 55122 22239 928 177150 8:2 297655 210359 72995 110246694 45798 19777 1125 209234 9:2 297357 202011 79363 119753474 50953 22088 1089 200922 10:2 278797 178703 83180 122514385 52969 22726 1128 177575 11:2 272661 167669 86978 127342327 55857 24342 1195 166474 12:2 293039 204031 73211 110282059 47285 19651 948 203083 13:2 289502 196762 76803 114712942 49339 20547 1016 195746 14:2 264446 169609 78292 115715605 50459 21017 982 168627 15:2 260968 163660 80142 116811793 51483 21281 1064 162596 With this change: domainstats: domain0 cpu cnt bln fld imb gain hgain nobusyq nobusyg 0:2 272347 187380 77455 105420270 24975 1 953 186427 1:2 267276 172360 86234 116242264 28087 6 1028 171332 2:2 259769 156777 93281 123243134 30555 1 1043 155734 3:2 250870 143129 97627 127370868 32026 6 1188 141941 4:2 248422 177116 64096 78261112 22202 2 757 176359 5:2 275595 180683 84950 116075022 29400 6 778 179905 6:2 262418 162609 88944 119256898 31056 4 817 161792 7:2 252204 147946 92646 122388300 32879 4 824 147122 8:2 262335 172239 81631 110477214 26599 4 864 171375 9:2 261563 164775 88016 117203621 28331 3 849 163926 10:2 243389 140949 93379 121353071 29585 2 909 140040 11:2 242795 134651 98310 124768957 30895 2 1016 133635 12:2 255234 166622 79843 104696912 26483 4 746 165876 13:2 244944 151595 83855 109808099 27787 3 801 150794 14:2 241301 140982 89935 116954383 30403 6 845 140137 15:2 232271 128564 92821 119185207 31207 4 1416 127148 Signed-off-by: Venkatesh Pallipadi <venki@google.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> LKML-Reference: <1284167957-3675-1-git-send-email-venki@google.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-09-11 01:19:17 +00:00
/*
* Increment the failure counter only on periodic balance.
* We do not want newidle balance, which can be very
* frequent, pollute the failure counter causing
* excessive cache_hot migrations and active balances.
*/
if (idle != CPU_NEWLY_IDLE)
sd->nr_balance_failed++;
if (need_active_balance(&env)) {
unsigned long flags;
raw_spin_lock_irqsave(&busiest->lock, flags);
/*
* Don't kick the active_load_balance_cpu_stop,
* if the curr task on busiest CPU can't be
* moved to this_cpu:
*/
if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) {
raw_spin_unlock_irqrestore(&busiest->lock,
flags);
env.flags |= LBF_ALL_PINNED;
goto out_one_pinned;
}
sched: replace migration_thread with cpu_stop Currently migration_thread is serving three purposes - migration pusher, context to execute active_load_balance() and forced context switcher for expedited RCU synchronize_sched. All three roles are hardcoded into migration_thread() and determining which job is scheduled is slightly messy. This patch kills migration_thread and replaces all three uses with cpu_stop. The three different roles of migration_thread() are splitted into three separate cpu_stop callbacks - migration_cpu_stop(), active_load_balance_cpu_stop() and synchronize_sched_expedited_cpu_stop() - and each use case now simply asks cpu_stop to execute the callback as necessary. synchronize_sched_expedited() was implemented with private preallocated resources and custom multi-cpu queueing and waiting logic, both of which are provided by cpu_stop. synchronize_sched_expedited_count is made atomic and all other shared resources along with the mutex are dropped. synchronize_sched_expedited() also implemented a check to detect cases where not all the callback got executed on their assigned cpus and fall back to synchronize_sched(). If called with cpu hotplug blocked, cpu_stop already guarantees that and the condition cannot happen; otherwise, stop_machine() would break. However, this patch preserves the paranoid check using a cpumask to record on which cpus the stopper ran so that it can serve as a bisection point if something actually goes wrong theree. Because the internal execution state is no longer visible, rcu_expedited_torture_stats() is removed. This patch also renames cpu_stop threads to from "stopper/%d" to "migration/%d". The names of these threads ultimately don't matter and there's no reason to make unnecessary userland visible changes. With this patch applied, stop_machine() and sched now share the same resources. stop_machine() is faster without wasting any resources and sched migration users are much cleaner. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Peter Zijlstra <peterz@infradead.org> Cc: Ingo Molnar <mingo@elte.hu> Cc: Dipankar Sarma <dipankar@in.ibm.com> Cc: Josh Triplett <josh@freedesktop.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Dimitri Sivanich <sivanich@sgi.com>
2010-05-06 16:49:21 +00:00
/*
* ->active_balance synchronizes accesses to
* ->active_balance_work. Once set, it's cleared
* only after active load balance is finished.
*/
if (!busiest->active_balance) {
busiest->active_balance = 1;
busiest->push_cpu = this_cpu;
active_balance = 1;
}
raw_spin_unlock_irqrestore(&busiest->lock, flags);
sched: replace migration_thread with cpu_stop Currently migration_thread is serving three purposes - migration pusher, context to execute active_load_balance() and forced context switcher for expedited RCU synchronize_sched. All three roles are hardcoded into migration_thread() and determining which job is scheduled is slightly messy. This patch kills migration_thread and replaces all three uses with cpu_stop. The three different roles of migration_thread() are splitted into three separate cpu_stop callbacks - migration_cpu_stop(), active_load_balance_cpu_stop() and synchronize_sched_expedited_cpu_stop() - and each use case now simply asks cpu_stop to execute the callback as necessary. synchronize_sched_expedited() was implemented with private preallocated resources and custom multi-cpu queueing and waiting logic, both of which are provided by cpu_stop. synchronize_sched_expedited_count is made atomic and all other shared resources along with the mutex are dropped. synchronize_sched_expedited() also implemented a check to detect cases where not all the callback got executed on their assigned cpus and fall back to synchronize_sched(). If called with cpu hotplug blocked, cpu_stop already guarantees that and the condition cannot happen; otherwise, stop_machine() would break. However, this patch preserves the paranoid check using a cpumask to record on which cpus the stopper ran so that it can serve as a bisection point if something actually goes wrong theree. Because the internal execution state is no longer visible, rcu_expedited_torture_stats() is removed. This patch also renames cpu_stop threads to from "stopper/%d" to "migration/%d". The names of these threads ultimately don't matter and there's no reason to make unnecessary userland visible changes. With this patch applied, stop_machine() and sched now share the same resources. stop_machine() is faster without wasting any resources and sched migration users are much cleaner. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Peter Zijlstra <peterz@infradead.org> Cc: Ingo Molnar <mingo@elte.hu> Cc: Dipankar Sarma <dipankar@in.ibm.com> Cc: Josh Triplett <josh@freedesktop.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Dimitri Sivanich <sivanich@sgi.com>
2010-05-06 16:49:21 +00:00
if (active_balance) {
sched: replace migration_thread with cpu_stop Currently migration_thread is serving three purposes - migration pusher, context to execute active_load_balance() and forced context switcher for expedited RCU synchronize_sched. All three roles are hardcoded into migration_thread() and determining which job is scheduled is slightly messy. This patch kills migration_thread and replaces all three uses with cpu_stop. The three different roles of migration_thread() are splitted into three separate cpu_stop callbacks - migration_cpu_stop(), active_load_balance_cpu_stop() and synchronize_sched_expedited_cpu_stop() - and each use case now simply asks cpu_stop to execute the callback as necessary. synchronize_sched_expedited() was implemented with private preallocated resources and custom multi-cpu queueing and waiting logic, both of which are provided by cpu_stop. synchronize_sched_expedited_count is made atomic and all other shared resources along with the mutex are dropped. synchronize_sched_expedited() also implemented a check to detect cases where not all the callback got executed on their assigned cpus and fall back to synchronize_sched(). If called with cpu hotplug blocked, cpu_stop already guarantees that and the condition cannot happen; otherwise, stop_machine() would break. However, this patch preserves the paranoid check using a cpumask to record on which cpus the stopper ran so that it can serve as a bisection point if something actually goes wrong theree. Because the internal execution state is no longer visible, rcu_expedited_torture_stats() is removed. This patch also renames cpu_stop threads to from "stopper/%d" to "migration/%d". The names of these threads ultimately don't matter and there's no reason to make unnecessary userland visible changes. With this patch applied, stop_machine() and sched now share the same resources. stop_machine() is faster without wasting any resources and sched migration users are much cleaner. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Peter Zijlstra <peterz@infradead.org> Cc: Ingo Molnar <mingo@elte.hu> Cc: Dipankar Sarma <dipankar@in.ibm.com> Cc: Josh Triplett <josh@freedesktop.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Dimitri Sivanich <sivanich@sgi.com>
2010-05-06 16:49:21 +00:00
stop_one_cpu_nowait(cpu_of(busiest),
active_load_balance_cpu_stop, busiest,
&busiest->active_balance_work);
}
/* We've kicked active balancing, force task migration. */
sd->nr_balance_failed = sd->cache_nice_tries+1;
}
} else
sd->nr_balance_failed = 0;
if (likely(!active_balance) || voluntary_active_balance(&env)) {
/* We were unbalanced, so reset the balancing interval */
sd->balance_interval = sd->min_interval;
} else {
/*
* If we've begun active balancing, start to back off. This
* case may not be covered by the all_pinned logic if there
* is only 1 task on the busy runqueue (because we don't call
* detach_tasks).
*/
if (sd->balance_interval < sd->max_interval)
sd->balance_interval *= 2;
}
goto out;
out_balanced:
sched: Fix imbalance flag reset The imbalance flag can stay set whereas there is no imbalance. Let assume that we have 3 tasks that run on a dual cores /dual cluster system. We will have some idle load balance which are triggered during tick. Unfortunately, the tick is also used to queue background work so we can reach the situation where short work has been queued on a CPU which already runs a task. The load balance will detect this imbalance (2 tasks on 1 CPU and an idle CPU) and will try to pull the waiting task on the idle CPU. The waiting task is a worker thread that is pinned on a CPU so an imbalance due to pinned task is detected and the imbalance flag is set. Then, we will not be able to clear the flag because we have at most 1 task on each CPU but the imbalance flag will trig to useless active load balance between the idle CPU and the busy CPU. We need to reset of the imbalance flag as soon as we have reached a balanced state. If all tasks are pinned, we don't consider that as a balanced state and let the imbalance flag set. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Reviewed-by: Preeti U Murthy <preeti@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: riel@redhat.com Cc: Morten.Rasmussen@arm.com Cc: efault@gmx.de Cc: nicolas.pitre@linaro.org Cc: daniel.lezcano@linaro.org Cc: dietmar.eggemann@arm.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/1409051215-16788-2-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-08-26 11:06:44 +00:00
/*
* We reach balance although we may have faced some affinity
sched/fair: Fix imbalance due to CPU affinity The load_balance() has a dedicated mecanism to detect when an imbalance is due to CPU affinity and must be handled at parent level. In this case, the imbalance field of the parent's sched_group is set. The description of sg_imbalanced() gives a typical example of two groups of 4 CPUs each and 4 tasks each with a cpumask covering 1 CPU of the first group and 3 CPUs of the second group. Something like: { 0 1 2 3 } { 4 5 6 7 } * * * * But the load_balance fails to fix this UC on my octo cores system made of 2 clusters of quad cores. Whereas the load_balance is able to detect that the imbalanced is due to CPU affinity, it fails to fix it because the imbalance field is cleared before letting parent level a chance to run. In fact, when the imbalance is detected, the load_balance reruns without the CPU with pinned tasks. But there is no other running tasks in the situation described above and everything looks balanced this time so the imbalance field is immediately cleared. The imbalance field should not be cleared if there is no other task to move when the imbalance is detected. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: https://lkml.kernel.org/r/1561996022-28829-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-07-01 15:47:02 +00:00
* constraints. Clear the imbalance flag only if other tasks got
* a chance to move and fix the imbalance.
sched: Fix imbalance flag reset The imbalance flag can stay set whereas there is no imbalance. Let assume that we have 3 tasks that run on a dual cores /dual cluster system. We will have some idle load balance which are triggered during tick. Unfortunately, the tick is also used to queue background work so we can reach the situation where short work has been queued on a CPU which already runs a task. The load balance will detect this imbalance (2 tasks on 1 CPU and an idle CPU) and will try to pull the waiting task on the idle CPU. The waiting task is a worker thread that is pinned on a CPU so an imbalance due to pinned task is detected and the imbalance flag is set. Then, we will not be able to clear the flag because we have at most 1 task on each CPU but the imbalance flag will trig to useless active load balance between the idle CPU and the busy CPU. We need to reset of the imbalance flag as soon as we have reached a balanced state. If all tasks are pinned, we don't consider that as a balanced state and let the imbalance flag set. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Reviewed-by: Preeti U Murthy <preeti@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: riel@redhat.com Cc: Morten.Rasmussen@arm.com Cc: efault@gmx.de Cc: nicolas.pitre@linaro.org Cc: daniel.lezcano@linaro.org Cc: dietmar.eggemann@arm.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/1409051215-16788-2-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-08-26 11:06:44 +00:00
*/
sched/fair: Fix imbalance due to CPU affinity The load_balance() has a dedicated mecanism to detect when an imbalance is due to CPU affinity and must be handled at parent level. In this case, the imbalance field of the parent's sched_group is set. The description of sg_imbalanced() gives a typical example of two groups of 4 CPUs each and 4 tasks each with a cpumask covering 1 CPU of the first group and 3 CPUs of the second group. Something like: { 0 1 2 3 } { 4 5 6 7 } * * * * But the load_balance fails to fix this UC on my octo cores system made of 2 clusters of quad cores. Whereas the load_balance is able to detect that the imbalanced is due to CPU affinity, it fails to fix it because the imbalance field is cleared before letting parent level a chance to run. In fact, when the imbalance is detected, the load_balance reruns without the CPU with pinned tasks. But there is no other running tasks in the situation described above and everything looks balanced this time so the imbalance field is immediately cleared. The imbalance field should not be cleared if there is no other task to move when the imbalance is detected. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: https://lkml.kernel.org/r/1561996022-28829-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-07-01 15:47:02 +00:00
if (sd_parent && !(env.flags & LBF_ALL_PINNED)) {
sched: Fix imbalance flag reset The imbalance flag can stay set whereas there is no imbalance. Let assume that we have 3 tasks that run on a dual cores /dual cluster system. We will have some idle load balance which are triggered during tick. Unfortunately, the tick is also used to queue background work so we can reach the situation where short work has been queued on a CPU which already runs a task. The load balance will detect this imbalance (2 tasks on 1 CPU and an idle CPU) and will try to pull the waiting task on the idle CPU. The waiting task is a worker thread that is pinned on a CPU so an imbalance due to pinned task is detected and the imbalance flag is set. Then, we will not be able to clear the flag because we have at most 1 task on each CPU but the imbalance flag will trig to useless active load balance between the idle CPU and the busy CPU. We need to reset of the imbalance flag as soon as we have reached a balanced state. If all tasks are pinned, we don't consider that as a balanced state and let the imbalance flag set. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Reviewed-by: Preeti U Murthy <preeti@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: riel@redhat.com Cc: Morten.Rasmussen@arm.com Cc: efault@gmx.de Cc: nicolas.pitre@linaro.org Cc: daniel.lezcano@linaro.org Cc: dietmar.eggemann@arm.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/1409051215-16788-2-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-08-26 11:06:44 +00:00
int *group_imbalance = &sd_parent->groups->sgc->imbalance;
if (*group_imbalance)
*group_imbalance = 0;
}
out_all_pinned:
/*
* We reach balance because all tasks are pinned at this level so
* we can't migrate them. Let the imbalance flag set so parent level
* can try to migrate them.
*/
schedstat_inc(sd->lb_balanced[idle]);
sd->nr_balance_failed = 0;
out_one_pinned:
ld_moved = 0;
/*
* newidle_balance() disregards balance intervals, so we could
* repeatedly reach this code, which would lead to balance_interval
* skyrocketting in a short amount of time. Skip the balance_interval
* increase logic to avoid that.
*/
if (env.idle == CPU_NEWLY_IDLE)
goto out;
/* tune up the balancing interval */
if ((env.flags & LBF_ALL_PINNED &&
sd->balance_interval < MAX_PINNED_INTERVAL) ||
sd->balance_interval < sd->max_interval)
sd->balance_interval *= 2;
out:
return ld_moved;
}
sched: Fix the rq->next_balance logic in rebalance_domains() and idle_balance() Currently, in idle_balance(), we update rq->next_balance when we pull_tasks. However, it is also important to update this in the !pulled_tasks case too. When the CPU is "busy" (the CPU isn't idle), rq->next_balance gets computed using sd->busy_factor (so we increase the balance interval when the CPU is busy). However, when the CPU goes idle, rq->next_balance could still be set to a large value that was computed with the sd->busy_factor. Thus, we need to also update rq->next_balance in idle_balance() in the cases where !pulled_tasks too, so that rq->next_balance gets updated without taking the busy_factor into account when the CPU is about to go idle. This patch makes rq->next_balance get updated independently of whether or not we pulled_task. Also, we add logic to ensure that we always traverse at least 1 of the sched domains to get a proper next_balance value for updating rq->next_balance. Additionally, since load_balance() modifies the sd->balance_interval, we need to re-obtain the sched domain's interval after the call to load_balance() in rebalance_domains() before we update rq->next_balance. This patch adds and uses 2 new helper functions, update_next_balance() and get_sd_balance_interval() to update next_balance and obtain the sched domain's balance_interval. Signed-off-by: Jason Low <jason.low2@hp.com> Reviewed-by: Preeti U Murthy <preeti@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Cc: daniel.lezcano@linaro.org Cc: alex.shi@linaro.org Cc: efault@gmx.de Cc: vincent.guittot@linaro.org Cc: morten.rasmussen@arm.com Cc: aswin@hp.com Link: http://lkml.kernel.org/r/1399596562.2200.7.camel@j-VirtualBox Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-05-09 00:49:22 +00:00
static inline unsigned long
get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
{
unsigned long interval = sd->balance_interval;
if (cpu_busy)
interval *= sd->busy_factor;
/* scale ms to jiffies */
interval = msecs_to_jiffies(interval);
interval = clamp(interval, 1UL, max_load_balance_interval);
return interval;
}
static inline void
update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
sched: Fix the rq->next_balance logic in rebalance_domains() and idle_balance() Currently, in idle_balance(), we update rq->next_balance when we pull_tasks. However, it is also important to update this in the !pulled_tasks case too. When the CPU is "busy" (the CPU isn't idle), rq->next_balance gets computed using sd->busy_factor (so we increase the balance interval when the CPU is busy). However, when the CPU goes idle, rq->next_balance could still be set to a large value that was computed with the sd->busy_factor. Thus, we need to also update rq->next_balance in idle_balance() in the cases where !pulled_tasks too, so that rq->next_balance gets updated without taking the busy_factor into account when the CPU is about to go idle. This patch makes rq->next_balance get updated independently of whether or not we pulled_task. Also, we add logic to ensure that we always traverse at least 1 of the sched domains to get a proper next_balance value for updating rq->next_balance. Additionally, since load_balance() modifies the sd->balance_interval, we need to re-obtain the sched domain's interval after the call to load_balance() in rebalance_domains() before we update rq->next_balance. This patch adds and uses 2 new helper functions, update_next_balance() and get_sd_balance_interval() to update next_balance and obtain the sched domain's balance_interval. Signed-off-by: Jason Low <jason.low2@hp.com> Reviewed-by: Preeti U Murthy <preeti@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Cc: daniel.lezcano@linaro.org Cc: alex.shi@linaro.org Cc: efault@gmx.de Cc: vincent.guittot@linaro.org Cc: morten.rasmussen@arm.com Cc: aswin@hp.com Link: http://lkml.kernel.org/r/1399596562.2200.7.camel@j-VirtualBox Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-05-09 00:49:22 +00:00
{
unsigned long interval, next;
/* used by idle balance, so cpu_busy = 0 */
interval = get_sd_balance_interval(sd, 0);
sched: Fix the rq->next_balance logic in rebalance_domains() and idle_balance() Currently, in idle_balance(), we update rq->next_balance when we pull_tasks. However, it is also important to update this in the !pulled_tasks case too. When the CPU is "busy" (the CPU isn't idle), rq->next_balance gets computed using sd->busy_factor (so we increase the balance interval when the CPU is busy). However, when the CPU goes idle, rq->next_balance could still be set to a large value that was computed with the sd->busy_factor. Thus, we need to also update rq->next_balance in idle_balance() in the cases where !pulled_tasks too, so that rq->next_balance gets updated without taking the busy_factor into account when the CPU is about to go idle. This patch makes rq->next_balance get updated independently of whether or not we pulled_task. Also, we add logic to ensure that we always traverse at least 1 of the sched domains to get a proper next_balance value for updating rq->next_balance. Additionally, since load_balance() modifies the sd->balance_interval, we need to re-obtain the sched domain's interval after the call to load_balance() in rebalance_domains() before we update rq->next_balance. This patch adds and uses 2 new helper functions, update_next_balance() and get_sd_balance_interval() to update next_balance and obtain the sched domain's balance_interval. Signed-off-by: Jason Low <jason.low2@hp.com> Reviewed-by: Preeti U Murthy <preeti@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Cc: daniel.lezcano@linaro.org Cc: alex.shi@linaro.org Cc: efault@gmx.de Cc: vincent.guittot@linaro.org Cc: morten.rasmussen@arm.com Cc: aswin@hp.com Link: http://lkml.kernel.org/r/1399596562.2200.7.camel@j-VirtualBox Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-05-09 00:49:22 +00:00
next = sd->last_balance + interval;
if (time_after(*next_balance, next))
*next_balance = next;
}
/*
* active_load_balance_cpu_stop is run by the CPU stopper. It pushes
sched: replace migration_thread with cpu_stop Currently migration_thread is serving three purposes - migration pusher, context to execute active_load_balance() and forced context switcher for expedited RCU synchronize_sched. All three roles are hardcoded into migration_thread() and determining which job is scheduled is slightly messy. This patch kills migration_thread and replaces all three uses with cpu_stop. The three different roles of migration_thread() are splitted into three separate cpu_stop callbacks - migration_cpu_stop(), active_load_balance_cpu_stop() and synchronize_sched_expedited_cpu_stop() - and each use case now simply asks cpu_stop to execute the callback as necessary. synchronize_sched_expedited() was implemented with private preallocated resources and custom multi-cpu queueing and waiting logic, both of which are provided by cpu_stop. synchronize_sched_expedited_count is made atomic and all other shared resources along with the mutex are dropped. synchronize_sched_expedited() also implemented a check to detect cases where not all the callback got executed on their assigned cpus and fall back to synchronize_sched(). If called with cpu hotplug blocked, cpu_stop already guarantees that and the condition cannot happen; otherwise, stop_machine() would break. However, this patch preserves the paranoid check using a cpumask to record on which cpus the stopper ran so that it can serve as a bisection point if something actually goes wrong theree. Because the internal execution state is no longer visible, rcu_expedited_torture_stats() is removed. This patch also renames cpu_stop threads to from "stopper/%d" to "migration/%d". The names of these threads ultimately don't matter and there's no reason to make unnecessary userland visible changes. With this patch applied, stop_machine() and sched now share the same resources. stop_machine() is faster without wasting any resources and sched migration users are much cleaner. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Peter Zijlstra <peterz@infradead.org> Cc: Ingo Molnar <mingo@elte.hu> Cc: Dipankar Sarma <dipankar@in.ibm.com> Cc: Josh Triplett <josh@freedesktop.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Dimitri Sivanich <sivanich@sgi.com>
2010-05-06 16:49:21 +00:00
* running tasks off the busiest CPU onto idle CPUs. It requires at
* least 1 task to be running on each physical CPU where possible, and
* avoids physical / logical imbalances.
*/
sched: replace migration_thread with cpu_stop Currently migration_thread is serving three purposes - migration pusher, context to execute active_load_balance() and forced context switcher for expedited RCU synchronize_sched. All three roles are hardcoded into migration_thread() and determining which job is scheduled is slightly messy. This patch kills migration_thread and replaces all three uses with cpu_stop. The three different roles of migration_thread() are splitted into three separate cpu_stop callbacks - migration_cpu_stop(), active_load_balance_cpu_stop() and synchronize_sched_expedited_cpu_stop() - and each use case now simply asks cpu_stop to execute the callback as necessary. synchronize_sched_expedited() was implemented with private preallocated resources and custom multi-cpu queueing and waiting logic, both of which are provided by cpu_stop. synchronize_sched_expedited_count is made atomic and all other shared resources along with the mutex are dropped. synchronize_sched_expedited() also implemented a check to detect cases where not all the callback got executed on their assigned cpus and fall back to synchronize_sched(). If called with cpu hotplug blocked, cpu_stop already guarantees that and the condition cannot happen; otherwise, stop_machine() would break. However, this patch preserves the paranoid check using a cpumask to record on which cpus the stopper ran so that it can serve as a bisection point if something actually goes wrong theree. Because the internal execution state is no longer visible, rcu_expedited_torture_stats() is removed. This patch also renames cpu_stop threads to from "stopper/%d" to "migration/%d". The names of these threads ultimately don't matter and there's no reason to make unnecessary userland visible changes. With this patch applied, stop_machine() and sched now share the same resources. stop_machine() is faster without wasting any resources and sched migration users are much cleaner. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Peter Zijlstra <peterz@infradead.org> Cc: Ingo Molnar <mingo@elte.hu> Cc: Dipankar Sarma <dipankar@in.ibm.com> Cc: Josh Triplett <josh@freedesktop.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Dimitri Sivanich <sivanich@sgi.com>
2010-05-06 16:49:21 +00:00
static int active_load_balance_cpu_stop(void *data)
{
sched: replace migration_thread with cpu_stop Currently migration_thread is serving three purposes - migration pusher, context to execute active_load_balance() and forced context switcher for expedited RCU synchronize_sched. All three roles are hardcoded into migration_thread() and determining which job is scheduled is slightly messy. This patch kills migration_thread and replaces all three uses with cpu_stop. The three different roles of migration_thread() are splitted into three separate cpu_stop callbacks - migration_cpu_stop(), active_load_balance_cpu_stop() and synchronize_sched_expedited_cpu_stop() - and each use case now simply asks cpu_stop to execute the callback as necessary. synchronize_sched_expedited() was implemented with private preallocated resources and custom multi-cpu queueing and waiting logic, both of which are provided by cpu_stop. synchronize_sched_expedited_count is made atomic and all other shared resources along with the mutex are dropped. synchronize_sched_expedited() also implemented a check to detect cases where not all the callback got executed on their assigned cpus and fall back to synchronize_sched(). If called with cpu hotplug blocked, cpu_stop already guarantees that and the condition cannot happen; otherwise, stop_machine() would break. However, this patch preserves the paranoid check using a cpumask to record on which cpus the stopper ran so that it can serve as a bisection point if something actually goes wrong theree. Because the internal execution state is no longer visible, rcu_expedited_torture_stats() is removed. This patch also renames cpu_stop threads to from "stopper/%d" to "migration/%d". The names of these threads ultimately don't matter and there's no reason to make unnecessary userland visible changes. With this patch applied, stop_machine() and sched now share the same resources. stop_machine() is faster without wasting any resources and sched migration users are much cleaner. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Peter Zijlstra <peterz@infradead.org> Cc: Ingo Molnar <mingo@elte.hu> Cc: Dipankar Sarma <dipankar@in.ibm.com> Cc: Josh Triplett <josh@freedesktop.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Dimitri Sivanich <sivanich@sgi.com>
2010-05-06 16:49:21 +00:00
struct rq *busiest_rq = data;
int busiest_cpu = cpu_of(busiest_rq);
int target_cpu = busiest_rq->push_cpu;
sched: replace migration_thread with cpu_stop Currently migration_thread is serving three purposes - migration pusher, context to execute active_load_balance() and forced context switcher for expedited RCU synchronize_sched. All three roles are hardcoded into migration_thread() and determining which job is scheduled is slightly messy. This patch kills migration_thread and replaces all three uses with cpu_stop. The three different roles of migration_thread() are splitted into three separate cpu_stop callbacks - migration_cpu_stop(), active_load_balance_cpu_stop() and synchronize_sched_expedited_cpu_stop() - and each use case now simply asks cpu_stop to execute the callback as necessary. synchronize_sched_expedited() was implemented with private preallocated resources and custom multi-cpu queueing and waiting logic, both of which are provided by cpu_stop. synchronize_sched_expedited_count is made atomic and all other shared resources along with the mutex are dropped. synchronize_sched_expedited() also implemented a check to detect cases where not all the callback got executed on their assigned cpus and fall back to synchronize_sched(). If called with cpu hotplug blocked, cpu_stop already guarantees that and the condition cannot happen; otherwise, stop_machine() would break. However, this patch preserves the paranoid check using a cpumask to record on which cpus the stopper ran so that it can serve as a bisection point if something actually goes wrong theree. Because the internal execution state is no longer visible, rcu_expedited_torture_stats() is removed. This patch also renames cpu_stop threads to from "stopper/%d" to "migration/%d". The names of these threads ultimately don't matter and there's no reason to make unnecessary userland visible changes. With this patch applied, stop_machine() and sched now share the same resources. stop_machine() is faster without wasting any resources and sched migration users are much cleaner. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Peter Zijlstra <peterz@infradead.org> Cc: Ingo Molnar <mingo@elte.hu> Cc: Dipankar Sarma <dipankar@in.ibm.com> Cc: Josh Triplett <josh@freedesktop.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Dimitri Sivanich <sivanich@sgi.com>
2010-05-06 16:49:21 +00:00
struct rq *target_rq = cpu_rq(target_cpu);
struct sched_domain *sd;
struct task_struct *p = NULL;
struct rq_flags rf;
sched: replace migration_thread with cpu_stop Currently migration_thread is serving three purposes - migration pusher, context to execute active_load_balance() and forced context switcher for expedited RCU synchronize_sched. All three roles are hardcoded into migration_thread() and determining which job is scheduled is slightly messy. This patch kills migration_thread and replaces all three uses with cpu_stop. The three different roles of migration_thread() are splitted into three separate cpu_stop callbacks - migration_cpu_stop(), active_load_balance_cpu_stop() and synchronize_sched_expedited_cpu_stop() - and each use case now simply asks cpu_stop to execute the callback as necessary. synchronize_sched_expedited() was implemented with private preallocated resources and custom multi-cpu queueing and waiting logic, both of which are provided by cpu_stop. synchronize_sched_expedited_count is made atomic and all other shared resources along with the mutex are dropped. synchronize_sched_expedited() also implemented a check to detect cases where not all the callback got executed on their assigned cpus and fall back to synchronize_sched(). If called with cpu hotplug blocked, cpu_stop already guarantees that and the condition cannot happen; otherwise, stop_machine() would break. However, this patch preserves the paranoid check using a cpumask to record on which cpus the stopper ran so that it can serve as a bisection point if something actually goes wrong theree. Because the internal execution state is no longer visible, rcu_expedited_torture_stats() is removed. This patch also renames cpu_stop threads to from "stopper/%d" to "migration/%d". The names of these threads ultimately don't matter and there's no reason to make unnecessary userland visible changes. With this patch applied, stop_machine() and sched now share the same resources. stop_machine() is faster without wasting any resources and sched migration users are much cleaner. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Peter Zijlstra <peterz@infradead.org> Cc: Ingo Molnar <mingo@elte.hu> Cc: Dipankar Sarma <dipankar@in.ibm.com> Cc: Josh Triplett <josh@freedesktop.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Dimitri Sivanich <sivanich@sgi.com>
2010-05-06 16:49:21 +00:00
rq_lock_irq(busiest_rq, &rf);
/*
* Between queueing the stop-work and running it is a hole in which
* CPUs can become inactive. We should not move tasks from or to
* inactive CPUs.
*/
if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
goto out_unlock;
sched: replace migration_thread with cpu_stop Currently migration_thread is serving three purposes - migration pusher, context to execute active_load_balance() and forced context switcher for expedited RCU synchronize_sched. All three roles are hardcoded into migration_thread() and determining which job is scheduled is slightly messy. This patch kills migration_thread and replaces all three uses with cpu_stop. The three different roles of migration_thread() are splitted into three separate cpu_stop callbacks - migration_cpu_stop(), active_load_balance_cpu_stop() and synchronize_sched_expedited_cpu_stop() - and each use case now simply asks cpu_stop to execute the callback as necessary. synchronize_sched_expedited() was implemented with private preallocated resources and custom multi-cpu queueing and waiting logic, both of which are provided by cpu_stop. synchronize_sched_expedited_count is made atomic and all other shared resources along with the mutex are dropped. synchronize_sched_expedited() also implemented a check to detect cases where not all the callback got executed on their assigned cpus and fall back to synchronize_sched(). If called with cpu hotplug blocked, cpu_stop already guarantees that and the condition cannot happen; otherwise, stop_machine() would break. However, this patch preserves the paranoid check using a cpumask to record on which cpus the stopper ran so that it can serve as a bisection point if something actually goes wrong theree. Because the internal execution state is no longer visible, rcu_expedited_torture_stats() is removed. This patch also renames cpu_stop threads to from "stopper/%d" to "migration/%d". The names of these threads ultimately don't matter and there's no reason to make unnecessary userland visible changes. With this patch applied, stop_machine() and sched now share the same resources. stop_machine() is faster without wasting any resources and sched migration users are much cleaner. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Peter Zijlstra <peterz@infradead.org> Cc: Ingo Molnar <mingo@elte.hu> Cc: Dipankar Sarma <dipankar@in.ibm.com> Cc: Josh Triplett <josh@freedesktop.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Dimitri Sivanich <sivanich@sgi.com>
2010-05-06 16:49:21 +00:00
/* Make sure the requested CPU hasn't gone down in the meantime: */
sched: replace migration_thread with cpu_stop Currently migration_thread is serving three purposes - migration pusher, context to execute active_load_balance() and forced context switcher for expedited RCU synchronize_sched. All three roles are hardcoded into migration_thread() and determining which job is scheduled is slightly messy. This patch kills migration_thread and replaces all three uses with cpu_stop. The three different roles of migration_thread() are splitted into three separate cpu_stop callbacks - migration_cpu_stop(), active_load_balance_cpu_stop() and synchronize_sched_expedited_cpu_stop() - and each use case now simply asks cpu_stop to execute the callback as necessary. synchronize_sched_expedited() was implemented with private preallocated resources and custom multi-cpu queueing and waiting logic, both of which are provided by cpu_stop. synchronize_sched_expedited_count is made atomic and all other shared resources along with the mutex are dropped. synchronize_sched_expedited() also implemented a check to detect cases where not all the callback got executed on their assigned cpus and fall back to synchronize_sched(). If called with cpu hotplug blocked, cpu_stop already guarantees that and the condition cannot happen; otherwise, stop_machine() would break. However, this patch preserves the paranoid check using a cpumask to record on which cpus the stopper ran so that it can serve as a bisection point if something actually goes wrong theree. Because the internal execution state is no longer visible, rcu_expedited_torture_stats() is removed. This patch also renames cpu_stop threads to from "stopper/%d" to "migration/%d". The names of these threads ultimately don't matter and there's no reason to make unnecessary userland visible changes. With this patch applied, stop_machine() and sched now share the same resources. stop_machine() is faster without wasting any resources and sched migration users are much cleaner. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Peter Zijlstra <peterz@infradead.org> Cc: Ingo Molnar <mingo@elte.hu> Cc: Dipankar Sarma <dipankar@in.ibm.com> Cc: Josh Triplett <josh@freedesktop.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Dimitri Sivanich <sivanich@sgi.com>
2010-05-06 16:49:21 +00:00
if (unlikely(busiest_cpu != smp_processor_id() ||
!busiest_rq->active_balance))
goto out_unlock;
/* Is there any task to move? */
if (busiest_rq->nr_running <= 1)
sched: replace migration_thread with cpu_stop Currently migration_thread is serving three purposes - migration pusher, context to execute active_load_balance() and forced context switcher for expedited RCU synchronize_sched. All three roles are hardcoded into migration_thread() and determining which job is scheduled is slightly messy. This patch kills migration_thread and replaces all three uses with cpu_stop. The three different roles of migration_thread() are splitted into three separate cpu_stop callbacks - migration_cpu_stop(), active_load_balance_cpu_stop() and synchronize_sched_expedited_cpu_stop() - and each use case now simply asks cpu_stop to execute the callback as necessary. synchronize_sched_expedited() was implemented with private preallocated resources and custom multi-cpu queueing and waiting logic, both of which are provided by cpu_stop. synchronize_sched_expedited_count is made atomic and all other shared resources along with the mutex are dropped. synchronize_sched_expedited() also implemented a check to detect cases where not all the callback got executed on their assigned cpus and fall back to synchronize_sched(). If called with cpu hotplug blocked, cpu_stop already guarantees that and the condition cannot happen; otherwise, stop_machine() would break. However, this patch preserves the paranoid check using a cpumask to record on which cpus the stopper ran so that it can serve as a bisection point if something actually goes wrong theree. Because the internal execution state is no longer visible, rcu_expedited_torture_stats() is removed. This patch also renames cpu_stop threads to from "stopper/%d" to "migration/%d". The names of these threads ultimately don't matter and there's no reason to make unnecessary userland visible changes. With this patch applied, stop_machine() and sched now share the same resources. stop_machine() is faster without wasting any resources and sched migration users are much cleaner. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Peter Zijlstra <peterz@infradead.org> Cc: Ingo Molnar <mingo@elte.hu> Cc: Dipankar Sarma <dipankar@in.ibm.com> Cc: Josh Triplett <josh@freedesktop.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Dimitri Sivanich <sivanich@sgi.com>
2010-05-06 16:49:21 +00:00
goto out_unlock;
/*
* This condition is "impossible", if it occurs
* we need to fix it. Originally reported by
* Bjorn Helgaas on a 128-CPU setup.
*/
BUG_ON(busiest_rq == target_rq);
/* Search for an sd spanning us and the target CPU. */
sched: Dynamically allocate sched_domain/sched_group data-structures Instead of relying on static allocations for the sched_domain and sched_group trees, dynamically allocate and RCU free them. Allocating this dynamically also allows for some build_sched_groups() simplification since we can now (like with other simplifications) rely on the sched_domain tree instead of hard-coded knowledge. One tricky to note is that detach_destroy_domains() needs to hold rcu_read_lock() over the entire tear-down, per-cpu is not sufficient since that can lead to partial sched_group existance (could possibly be solved by doing the tear-down backwards but this is much more robust). A concequence of the above is that we can no longer print the sched_domain debug stuff from cpu_attach_domain() since that might now run with preemption disabled (due to classic RCU etc.) and sched_domain_debug() does some GFP_KERNEL allocations. Another thing to note is that we now fully rely on normal RCU and not RCU-sched, this is because with the new and exiting RCU flavours we grew over the years BH doesn't necessarily hold off RCU-sched grace periods (-rt is known to break this). This would in fact already cause us grief since we do sched_domain/sched_group iterations from softirq context. This patch is somewhat larger than I would like it to be, but I didn't find any means of shrinking/splitting this. Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> Cc: Nick Piggin <npiggin@kernel.dk> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Link: http://lkml.kernel.org/r/20110407122942.245307941@chello.nl Signed-off-by: Ingo Molnar <mingo@elte.hu>
2011-04-07 12:09:50 +00:00
rcu_read_lock();
for_each_domain(target_cpu, sd) {
if ((sd->flags & SD_LOAD_BALANCE) &&
cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
break;
}
if (likely(sd)) {
struct lb_env env = {
.sd = sd,
.dst_cpu = target_cpu,
.dst_rq = target_rq,
.src_cpu = busiest_rq->cpu,
.src_rq = busiest_rq,
.idle = CPU_IDLE,
sched/fair: Fix load_balance() affinity redo path If load_balance() fails to migrate any tasks because all tasks were affined, load_balance() removes the source CPU from consideration and attempts to redo and balance among the new subset of CPUs. There is a bug in this code path where the algorithm considers all active CPUs in the system (minus the source that was just masked out). This is not valid for two reasons: some active CPUs may not be in the current scheduling domain and one of the active CPUs is dst_cpu. These CPUs should not be considered, as we cannot pull load from them. Instead of failing out of load_balance(), we may end up redoing the search with no valid CPUs and incorrectly concluding the domain is balanced. Additionally, if the group_imbalance flag was just set, it may also be incorrectly unset, thus the flag will not be seen by other CPUs in future load_balance() runs as that algorithm intends. Fix the check by removing CPUs not in the current domain and the dst_cpu from considertation, thus limiting the evaluation to valid remaining CPUs from which load might be migrated. Co-authored-by: Austin Christ <austinwc@codeaurora.org> Co-authored-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Tested-by: Tyler Baicar <tbaicar@codeaurora.org> Signed-off-by: Jeffrey Hugo <jhugo@codeaurora.org> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Austin Christ <austinwc@codeaurora.org> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Timur Tabi <timur@codeaurora.org> Link: http://lkml.kernel.org/r/1496863138-11322-2-git-send-email-jhugo@codeaurora.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-06-07 19:18:57 +00:00
/*
* can_migrate_task() doesn't need to compute new_dst_cpu
* for active balancing. Since we have CPU_IDLE, but no
* @dst_grpmask we need to make that test go away with lying
* about DST_PINNED.
*/
.flags = LBF_DST_PINNED,
};
schedstat_inc(sd->alb_count);
update_rq_clock(busiest_rq);
p = detach_one_task(&env);
if (p) {
schedstat_inc(sd->alb_pushed);
/* Active balancing done, reset the failure counter. */
sd->nr_balance_failed = 0;
} else {
schedstat_inc(sd->alb_failed);
}
}
sched: Dynamically allocate sched_domain/sched_group data-structures Instead of relying on static allocations for the sched_domain and sched_group trees, dynamically allocate and RCU free them. Allocating this dynamically also allows for some build_sched_groups() simplification since we can now (like with other simplifications) rely on the sched_domain tree instead of hard-coded knowledge. One tricky to note is that detach_destroy_domains() needs to hold rcu_read_lock() over the entire tear-down, per-cpu is not sufficient since that can lead to partial sched_group existance (could possibly be solved by doing the tear-down backwards but this is much more robust). A concequence of the above is that we can no longer print the sched_domain debug stuff from cpu_attach_domain() since that might now run with preemption disabled (due to classic RCU etc.) and sched_domain_debug() does some GFP_KERNEL allocations. Another thing to note is that we now fully rely on normal RCU and not RCU-sched, this is because with the new and exiting RCU flavours we grew over the years BH doesn't necessarily hold off RCU-sched grace periods (-rt is known to break this). This would in fact already cause us grief since we do sched_domain/sched_group iterations from softirq context. This patch is somewhat larger than I would like it to be, but I didn't find any means of shrinking/splitting this. Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> Cc: Nick Piggin <npiggin@kernel.dk> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Link: http://lkml.kernel.org/r/20110407122942.245307941@chello.nl Signed-off-by: Ingo Molnar <mingo@elte.hu>
2011-04-07 12:09:50 +00:00
rcu_read_unlock();
sched: replace migration_thread with cpu_stop Currently migration_thread is serving three purposes - migration pusher, context to execute active_load_balance() and forced context switcher for expedited RCU synchronize_sched. All three roles are hardcoded into migration_thread() and determining which job is scheduled is slightly messy. This patch kills migration_thread and replaces all three uses with cpu_stop. The three different roles of migration_thread() are splitted into three separate cpu_stop callbacks - migration_cpu_stop(), active_load_balance_cpu_stop() and synchronize_sched_expedited_cpu_stop() - and each use case now simply asks cpu_stop to execute the callback as necessary. synchronize_sched_expedited() was implemented with private preallocated resources and custom multi-cpu queueing and waiting logic, both of which are provided by cpu_stop. synchronize_sched_expedited_count is made atomic and all other shared resources along with the mutex are dropped. synchronize_sched_expedited() also implemented a check to detect cases where not all the callback got executed on their assigned cpus and fall back to synchronize_sched(). If called with cpu hotplug blocked, cpu_stop already guarantees that and the condition cannot happen; otherwise, stop_machine() would break. However, this patch preserves the paranoid check using a cpumask to record on which cpus the stopper ran so that it can serve as a bisection point if something actually goes wrong theree. Because the internal execution state is no longer visible, rcu_expedited_torture_stats() is removed. This patch also renames cpu_stop threads to from "stopper/%d" to "migration/%d". The names of these threads ultimately don't matter and there's no reason to make unnecessary userland visible changes. With this patch applied, stop_machine() and sched now share the same resources. stop_machine() is faster without wasting any resources and sched migration users are much cleaner. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Peter Zijlstra <peterz@infradead.org> Cc: Ingo Molnar <mingo@elte.hu> Cc: Dipankar Sarma <dipankar@in.ibm.com> Cc: Josh Triplett <josh@freedesktop.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Dimitri Sivanich <sivanich@sgi.com>
2010-05-06 16:49:21 +00:00
out_unlock:
busiest_rq->active_balance = 0;
rq_unlock(busiest_rq, &rf);
if (p)
attach_one_task(target_rq, p);
local_irq_enable();
sched: replace migration_thread with cpu_stop Currently migration_thread is serving three purposes - migration pusher, context to execute active_load_balance() and forced context switcher for expedited RCU synchronize_sched. All three roles are hardcoded into migration_thread() and determining which job is scheduled is slightly messy. This patch kills migration_thread and replaces all three uses with cpu_stop. The three different roles of migration_thread() are splitted into three separate cpu_stop callbacks - migration_cpu_stop(), active_load_balance_cpu_stop() and synchronize_sched_expedited_cpu_stop() - and each use case now simply asks cpu_stop to execute the callback as necessary. synchronize_sched_expedited() was implemented with private preallocated resources and custom multi-cpu queueing and waiting logic, both of which are provided by cpu_stop. synchronize_sched_expedited_count is made atomic and all other shared resources along with the mutex are dropped. synchronize_sched_expedited() also implemented a check to detect cases where not all the callback got executed on their assigned cpus and fall back to synchronize_sched(). If called with cpu hotplug blocked, cpu_stop already guarantees that and the condition cannot happen; otherwise, stop_machine() would break. However, this patch preserves the paranoid check using a cpumask to record on which cpus the stopper ran so that it can serve as a bisection point if something actually goes wrong theree. Because the internal execution state is no longer visible, rcu_expedited_torture_stats() is removed. This patch also renames cpu_stop threads to from "stopper/%d" to "migration/%d". The names of these threads ultimately don't matter and there's no reason to make unnecessary userland visible changes. With this patch applied, stop_machine() and sched now share the same resources. stop_machine() is faster without wasting any resources and sched migration users are much cleaner. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Peter Zijlstra <peterz@infradead.org> Cc: Ingo Molnar <mingo@elte.hu> Cc: Dipankar Sarma <dipankar@in.ibm.com> Cc: Josh Triplett <josh@freedesktop.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Dimitri Sivanich <sivanich@sgi.com>
2010-05-06 16:49:21 +00:00
return 0;
}
static DEFINE_SPINLOCK(balancing);
/*
* Scale the max load_balance interval with the number of CPUs in the system.
* This trades load-balance latency on larger machines for less cross talk.
*/
void update_max_interval(void)
{
max_load_balance_interval = HZ*num_online_cpus()/10;
}
/*
* It checks each scheduling domain to see if it is due to be balanced,
* and initiates a balancing operation if so.
*
* Balancing parameters are set up in init_sched_domains.
*/
static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
{
int continue_balancing = 1;
int cpu = rq->cpu;
sched/fair: Load balance aggressively for SCHED_IDLE CPUs The fair scheduler performs periodic load balance on every CPU to check if it can pull some tasks from other busy CPUs. The duration of this periodic load balance is set to sd->balance_interval for the idle CPUs and is calculated by multiplying the sd->balance_interval with the sd->busy_factor (set to 32 by default) for the busy CPUs. The multiplication is done for busy CPUs to avoid doing load balance too often and rather spend more time executing actual task. While that is the right thing to do for the CPUs busy with SCHED_OTHER or SCHED_BATCH tasks, it may not be the optimal thing for CPUs running only SCHED_IDLE tasks. With the recent enhancements in the fair scheduler around SCHED_IDLE CPUs, we now prefer to enqueue a newly-woken task to a SCHED_IDLE CPU instead of other busy or idle CPUs. The same reasoning should be applied to the load balancer as well to make it migrate tasks more aggressively to a SCHED_IDLE CPU, as that will reduce the scheduling latency of the migrated (SCHED_OTHER) tasks. This patch makes minimal changes to the fair scheduler to do the next load balance soon after the last non SCHED_IDLE task is dequeued from a runqueue, i.e. making the CPU SCHED_IDLE. Also the sd->busy_factor is ignored while calculating the balance_interval for such CPUs. This is done to avoid delaying the periodic load balance by few hundred milliseconds for SCHED_IDLE CPUs. This is tested on ARM64 Hikey620 platform (octa-core) with the help of rt-app and it is verified, using kernel traces, that the newly SCHED_IDLE CPU does load balancing shortly after it becomes SCHED_IDLE and pulls tasks from other busy CPUs. Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org> Link: https://lkml.kernel.org/r/e485827eb8fe7db0943d6f3f6e0f5a4a70272781.1578471925.git.viresh.kumar@linaro.org
2020-01-08 08:27:04 +00:00
int busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
unsigned long interval;
struct sched_domain *sd;
/* Earliest time when we have to do rebalance again */
unsigned long next_balance = jiffies + 60*HZ;
int update_next_balance = 0;
int need_serialize, need_decay = 0;
u64 max_cost = 0;
rcu_read_lock();
for_each_domain(cpu, sd) {
/*
* Decay the newidle max times here because this is a regular
* visit to all the domains. Decay ~1% per second.
*/
if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
sd->max_newidle_lb_cost =
(sd->max_newidle_lb_cost * 253) / 256;
sd->next_decay_max_lb_cost = jiffies + HZ;
need_decay = 1;
}
max_cost += sd->max_newidle_lb_cost;
if (!(sd->flags & SD_LOAD_BALANCE))
continue;
/*
* Stop the load balance at this level. There is another
* CPU in our sched group which is doing load balancing more
* actively.
*/
if (!continue_balancing) {
if (need_decay)
continue;
break;
}
sched/fair: Load balance aggressively for SCHED_IDLE CPUs The fair scheduler performs periodic load balance on every CPU to check if it can pull some tasks from other busy CPUs. The duration of this periodic load balance is set to sd->balance_interval for the idle CPUs and is calculated by multiplying the sd->balance_interval with the sd->busy_factor (set to 32 by default) for the busy CPUs. The multiplication is done for busy CPUs to avoid doing load balance too often and rather spend more time executing actual task. While that is the right thing to do for the CPUs busy with SCHED_OTHER or SCHED_BATCH tasks, it may not be the optimal thing for CPUs running only SCHED_IDLE tasks. With the recent enhancements in the fair scheduler around SCHED_IDLE CPUs, we now prefer to enqueue a newly-woken task to a SCHED_IDLE CPU instead of other busy or idle CPUs. The same reasoning should be applied to the load balancer as well to make it migrate tasks more aggressively to a SCHED_IDLE CPU, as that will reduce the scheduling latency of the migrated (SCHED_OTHER) tasks. This patch makes minimal changes to the fair scheduler to do the next load balance soon after the last non SCHED_IDLE task is dequeued from a runqueue, i.e. making the CPU SCHED_IDLE. Also the sd->busy_factor is ignored while calculating the balance_interval for such CPUs. This is done to avoid delaying the periodic load balance by few hundred milliseconds for SCHED_IDLE CPUs. This is tested on ARM64 Hikey620 platform (octa-core) with the help of rt-app and it is verified, using kernel traces, that the newly SCHED_IDLE CPU does load balancing shortly after it becomes SCHED_IDLE and pulls tasks from other busy CPUs. Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org> Link: https://lkml.kernel.org/r/e485827eb8fe7db0943d6f3f6e0f5a4a70272781.1578471925.git.viresh.kumar@linaro.org
2020-01-08 08:27:04 +00:00
interval = get_sd_balance_interval(sd, busy);
need_serialize = sd->flags & SD_SERIALIZE;
if (need_serialize) {
if (!spin_trylock(&balancing))
goto out;
}
if (time_after_eq(jiffies, sd->last_balance + interval)) {
if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
/*
* The LBF_DST_PINNED logic could have changed
* env->dst_cpu, so we can't know our idle
* state even if we migrated tasks. Update it.
*/
idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
sched/fair: Load balance aggressively for SCHED_IDLE CPUs The fair scheduler performs periodic load balance on every CPU to check if it can pull some tasks from other busy CPUs. The duration of this periodic load balance is set to sd->balance_interval for the idle CPUs and is calculated by multiplying the sd->balance_interval with the sd->busy_factor (set to 32 by default) for the busy CPUs. The multiplication is done for busy CPUs to avoid doing load balance too often and rather spend more time executing actual task. While that is the right thing to do for the CPUs busy with SCHED_OTHER or SCHED_BATCH tasks, it may not be the optimal thing for CPUs running only SCHED_IDLE tasks. With the recent enhancements in the fair scheduler around SCHED_IDLE CPUs, we now prefer to enqueue a newly-woken task to a SCHED_IDLE CPU instead of other busy or idle CPUs. The same reasoning should be applied to the load balancer as well to make it migrate tasks more aggressively to a SCHED_IDLE CPU, as that will reduce the scheduling latency of the migrated (SCHED_OTHER) tasks. This patch makes minimal changes to the fair scheduler to do the next load balance soon after the last non SCHED_IDLE task is dequeued from a runqueue, i.e. making the CPU SCHED_IDLE. Also the sd->busy_factor is ignored while calculating the balance_interval for such CPUs. This is done to avoid delaying the periodic load balance by few hundred milliseconds for SCHED_IDLE CPUs. This is tested on ARM64 Hikey620 platform (octa-core) with the help of rt-app and it is verified, using kernel traces, that the newly SCHED_IDLE CPU does load balancing shortly after it becomes SCHED_IDLE and pulls tasks from other busy CPUs. Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org> Link: https://lkml.kernel.org/r/e485827eb8fe7db0943d6f3f6e0f5a4a70272781.1578471925.git.viresh.kumar@linaro.org
2020-01-08 08:27:04 +00:00
busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
}
sd->last_balance = jiffies;
sched/fair: Load balance aggressively for SCHED_IDLE CPUs The fair scheduler performs periodic load balance on every CPU to check if it can pull some tasks from other busy CPUs. The duration of this periodic load balance is set to sd->balance_interval for the idle CPUs and is calculated by multiplying the sd->balance_interval with the sd->busy_factor (set to 32 by default) for the busy CPUs. The multiplication is done for busy CPUs to avoid doing load balance too often and rather spend more time executing actual task. While that is the right thing to do for the CPUs busy with SCHED_OTHER or SCHED_BATCH tasks, it may not be the optimal thing for CPUs running only SCHED_IDLE tasks. With the recent enhancements in the fair scheduler around SCHED_IDLE CPUs, we now prefer to enqueue a newly-woken task to a SCHED_IDLE CPU instead of other busy or idle CPUs. The same reasoning should be applied to the load balancer as well to make it migrate tasks more aggressively to a SCHED_IDLE CPU, as that will reduce the scheduling latency of the migrated (SCHED_OTHER) tasks. This patch makes minimal changes to the fair scheduler to do the next load balance soon after the last non SCHED_IDLE task is dequeued from a runqueue, i.e. making the CPU SCHED_IDLE. Also the sd->busy_factor is ignored while calculating the balance_interval for such CPUs. This is done to avoid delaying the periodic load balance by few hundred milliseconds for SCHED_IDLE CPUs. This is tested on ARM64 Hikey620 platform (octa-core) with the help of rt-app and it is verified, using kernel traces, that the newly SCHED_IDLE CPU does load balancing shortly after it becomes SCHED_IDLE and pulls tasks from other busy CPUs. Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org> Link: https://lkml.kernel.org/r/e485827eb8fe7db0943d6f3f6e0f5a4a70272781.1578471925.git.viresh.kumar@linaro.org
2020-01-08 08:27:04 +00:00
interval = get_sd_balance_interval(sd, busy);
}
if (need_serialize)
spin_unlock(&balancing);
out:
if (time_after(next_balance, sd->last_balance + interval)) {
next_balance = sd->last_balance + interval;
update_next_balance = 1;
}
}
if (need_decay) {
/*
* Ensure the rq-wide value also decays but keep it at a
* reasonable floor to avoid funnies with rq->avg_idle.
*/
rq->max_idle_balance_cost =
max((u64)sysctl_sched_migration_cost, max_cost);
}
rcu_read_unlock();
/*
* next_balance will be updated only when there is a need.
* When the cpu is attached to null domain for ex, it will not be
* updated.
*/
if (likely(update_next_balance)) {
rq->next_balance = next_balance;
#ifdef CONFIG_NO_HZ_COMMON
/*
* If this CPU has been elected to perform the nohz idle
* balance. Other idle CPUs have already rebalanced with
* nohz_idle_balance() and nohz.next_balance has been
* updated accordingly. This CPU is now running the idle load
* balance for itself and we need to update the
* nohz.next_balance accordingly.
*/
if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
nohz.next_balance = rq->next_balance;
#endif
}
}
static inline int on_null_domain(struct rq *rq)
{
return unlikely(!rcu_dereference_sched(rq->sd));
}
nohz: Rename CONFIG_NO_HZ to CONFIG_NO_HZ_COMMON We are planning to convert the dynticks Kconfig options layout into a choice menu. The user must be able to easily pick any of the following implementations: constant periodic tick, idle dynticks, full dynticks. As this implies a mutual exclusion, the two dynticks implementions need to converge on the selection of a common Kconfig option in order to ease the sharing of a common infrastructure. It would thus seem pretty natural to reuse CONFIG_NO_HZ to that end. It already implements all the idle dynticks code and the full dynticks depends on all that code for now. So ideally the choice menu would propose CONFIG_NO_HZ_IDLE and CONFIG_NO_HZ_EXTENDED then both would select CONFIG_NO_HZ. On the other hand we want to stay backward compatible: if CONFIG_NO_HZ is set in an older config file, we want to enable CONFIG_NO_HZ_IDLE by default. But we can't afford both at the same time or we run into a circular dependency: 1) CONFIG_NO_HZ_IDLE and CONFIG_NO_HZ_EXTENDED both select CONFIG_NO_HZ 2) If CONFIG_NO_HZ is set, we default to CONFIG_NO_HZ_IDLE We might be able to support that from Kconfig/Kbuild but it may not be wise to introduce such a confusing behaviour. So to solve this, create a new CONFIG_NO_HZ_COMMON option which gathers the common code between idle and full dynticks (that common code for now is simply the idle dynticks code) and select it from their referring Kconfig. Then we'll later create CONFIG_NO_HZ_IDLE and map CONFIG_NO_HZ to it for backward compatibility. Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Chris Metcalf <cmetcalf@tilera.com> Cc: Christoph Lameter <cl@linux.com> Cc: Geoff Levand <geoff@infradead.org> Cc: Gilad Ben Yossef <gilad@benyossef.com> Cc: Hakan Akkan <hakanakkan@gmail.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Kevin Hilman <khilman@linaro.org> Cc: Li Zhong <zhong@linux.vnet.ibm.com> Cc: Namhyung Kim <namhyung.kim@lge.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de>
2011-08-10 21:21:01 +00:00
#ifdef CONFIG_NO_HZ_COMMON
sched: Change nohz idle load balancing logic to push model In the new push model, all idle CPUs indeed go into nohz mode. There is still the concept of idle load balancer (performing the load balancing on behalf of all the idle cpu's in the system). Busy CPU kicks the nohz balancer when any of the nohz CPUs need idle load balancing. The kickee CPU does the idle load balancing on behalf of all idle CPUs instead of the normal idle balance. This addresses the below two problems with the current nohz ilb logic: * the idle load balancer continued to have periodic ticks during idle and wokeup frequently, even though it did not have any rebalancing to do on behalf of any of the idle CPUs. * On x86 and CPUs that have APIC timer stoppage on idle CPUs, this periodic wakeup can result in a periodic additional interrupt on a CPU doing the timer broadcast. Also currently we are migrating the unpinned timers from an idle to the cpu doing idle load balancing (when all the cpus in the system are idle, there is no idle load balancing cpu and timers get added to the same idle cpu where the request was made. So the existing optimization works only on semi idle system). And In semi idle system, we no longer have periodic ticks on the idle load balancer CPU. Using that cpu will add more delays to the timers than intended (as that cpu's timer base may not be uptodate wrt jiffies etc). This was causing mysterious slowdowns during boot etc. For now, in the semi idle case, use the nearest busy cpu for migrating timers from an idle cpu. This is good for power-savings anyway. Signed-off-by: Venkatesh Pallipadi <venki@google.com> Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> LKML-Reference: <1274486981.2840.46.camel@sbs-t61.sc.intel.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-05-22 00:09:41 +00:00
/*
* idle load balancing details
* - When one of the busy CPUs notice that there may be an idle rebalancing
* needed, they will kick the idle load balancer, which then does idle
* load balancing for all the idle CPUs.
* - HK_FLAG_MISC CPUs are used for this task, because HK_FLAG_SCHED not set
* anywhere yet.
sched: Change nohz idle load balancing logic to push model In the new push model, all idle CPUs indeed go into nohz mode. There is still the concept of idle load balancer (performing the load balancing on behalf of all the idle cpu's in the system). Busy CPU kicks the nohz balancer when any of the nohz CPUs need idle load balancing. The kickee CPU does the idle load balancing on behalf of all idle CPUs instead of the normal idle balance. This addresses the below two problems with the current nohz ilb logic: * the idle load balancer continued to have periodic ticks during idle and wokeup frequently, even though it did not have any rebalancing to do on behalf of any of the idle CPUs. * On x86 and CPUs that have APIC timer stoppage on idle CPUs, this periodic wakeup can result in a periodic additional interrupt on a CPU doing the timer broadcast. Also currently we are migrating the unpinned timers from an idle to the cpu doing idle load balancing (when all the cpus in the system are idle, there is no idle load balancing cpu and timers get added to the same idle cpu where the request was made. So the existing optimization works only on semi idle system). And In semi idle system, we no longer have periodic ticks on the idle load balancer CPU. Using that cpu will add more delays to the timers than intended (as that cpu's timer base may not be uptodate wrt jiffies etc). This was causing mysterious slowdowns during boot etc. For now, in the semi idle case, use the nearest busy cpu for migrating timers from an idle cpu. This is good for power-savings anyway. Signed-off-by: Venkatesh Pallipadi <venki@google.com> Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> LKML-Reference: <1274486981.2840.46.camel@sbs-t61.sc.intel.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-05-22 00:09:41 +00:00
*/
static inline int find_new_ilb(void)
{
int ilb;
for_each_cpu_and(ilb, nohz.idle_cpus_mask,
housekeeping_cpumask(HK_FLAG_MISC)) {
if (idle_cpu(ilb))
return ilb;
}
return nr_cpu_ids;
}
sched: Change nohz idle load balancing logic to push model In the new push model, all idle CPUs indeed go into nohz mode. There is still the concept of idle load balancer (performing the load balancing on behalf of all the idle cpu's in the system). Busy CPU kicks the nohz balancer when any of the nohz CPUs need idle load balancing. The kickee CPU does the idle load balancing on behalf of all idle CPUs instead of the normal idle balance. This addresses the below two problems with the current nohz ilb logic: * the idle load balancer continued to have periodic ticks during idle and wokeup frequently, even though it did not have any rebalancing to do on behalf of any of the idle CPUs. * On x86 and CPUs that have APIC timer stoppage on idle CPUs, this periodic wakeup can result in a periodic additional interrupt on a CPU doing the timer broadcast. Also currently we are migrating the unpinned timers from an idle to the cpu doing idle load balancing (when all the cpus in the system are idle, there is no idle load balancing cpu and timers get added to the same idle cpu where the request was made. So the existing optimization works only on semi idle system). And In semi idle system, we no longer have periodic ticks on the idle load balancer CPU. Using that cpu will add more delays to the timers than intended (as that cpu's timer base may not be uptodate wrt jiffies etc). This was causing mysterious slowdowns during boot etc. For now, in the semi idle case, use the nearest busy cpu for migrating timers from an idle cpu. This is good for power-savings anyway. Signed-off-by: Venkatesh Pallipadi <venki@google.com> Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> LKML-Reference: <1274486981.2840.46.camel@sbs-t61.sc.intel.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-05-22 00:09:41 +00:00
/*
* Kick a CPU to do the nohz balancing, if it is time for it. We pick any
* idle CPU in the HK_FLAG_MISC housekeeping set (if there is one).
sched: Change nohz idle load balancing logic to push model In the new push model, all idle CPUs indeed go into nohz mode. There is still the concept of idle load balancer (performing the load balancing on behalf of all the idle cpu's in the system). Busy CPU kicks the nohz balancer when any of the nohz CPUs need idle load balancing. The kickee CPU does the idle load balancing on behalf of all idle CPUs instead of the normal idle balance. This addresses the below two problems with the current nohz ilb logic: * the idle load balancer continued to have periodic ticks during idle and wokeup frequently, even though it did not have any rebalancing to do on behalf of any of the idle CPUs. * On x86 and CPUs that have APIC timer stoppage on idle CPUs, this periodic wakeup can result in a periodic additional interrupt on a CPU doing the timer broadcast. Also currently we are migrating the unpinned timers from an idle to the cpu doing idle load balancing (when all the cpus in the system are idle, there is no idle load balancing cpu and timers get added to the same idle cpu where the request was made. So the existing optimization works only on semi idle system). And In semi idle system, we no longer have periodic ticks on the idle load balancer CPU. Using that cpu will add more delays to the timers than intended (as that cpu's timer base may not be uptodate wrt jiffies etc). This was causing mysterious slowdowns during boot etc. For now, in the semi idle case, use the nearest busy cpu for migrating timers from an idle cpu. This is good for power-savings anyway. Signed-off-by: Venkatesh Pallipadi <venki@google.com> Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> LKML-Reference: <1274486981.2840.46.camel@sbs-t61.sc.intel.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-05-22 00:09:41 +00:00
*/
static void kick_ilb(unsigned int flags)
sched: Change nohz idle load balancing logic to push model In the new push model, all idle CPUs indeed go into nohz mode. There is still the concept of idle load balancer (performing the load balancing on behalf of all the idle cpu's in the system). Busy CPU kicks the nohz balancer when any of the nohz CPUs need idle load balancing. The kickee CPU does the idle load balancing on behalf of all idle CPUs instead of the normal idle balance. This addresses the below two problems with the current nohz ilb logic: * the idle load balancer continued to have periodic ticks during idle and wokeup frequently, even though it did not have any rebalancing to do on behalf of any of the idle CPUs. * On x86 and CPUs that have APIC timer stoppage on idle CPUs, this periodic wakeup can result in a periodic additional interrupt on a CPU doing the timer broadcast. Also currently we are migrating the unpinned timers from an idle to the cpu doing idle load balancing (when all the cpus in the system are idle, there is no idle load balancing cpu and timers get added to the same idle cpu where the request was made. So the existing optimization works only on semi idle system). And In semi idle system, we no longer have periodic ticks on the idle load balancer CPU. Using that cpu will add more delays to the timers than intended (as that cpu's timer base may not be uptodate wrt jiffies etc). This was causing mysterious slowdowns during boot etc. For now, in the semi idle case, use the nearest busy cpu for migrating timers from an idle cpu. This is good for power-savings anyway. Signed-off-by: Venkatesh Pallipadi <venki@google.com> Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> LKML-Reference: <1274486981.2840.46.camel@sbs-t61.sc.intel.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-05-22 00:09:41 +00:00
{
int ilb_cpu;
nohz.next_balance++;
ilb_cpu = find_new_ilb();
sched: Change nohz idle load balancing logic to push model In the new push model, all idle CPUs indeed go into nohz mode. There is still the concept of idle load balancer (performing the load balancing on behalf of all the idle cpu's in the system). Busy CPU kicks the nohz balancer when any of the nohz CPUs need idle load balancing. The kickee CPU does the idle load balancing on behalf of all idle CPUs instead of the normal idle balance. This addresses the below two problems with the current nohz ilb logic: * the idle load balancer continued to have periodic ticks during idle and wokeup frequently, even though it did not have any rebalancing to do on behalf of any of the idle CPUs. * On x86 and CPUs that have APIC timer stoppage on idle CPUs, this periodic wakeup can result in a periodic additional interrupt on a CPU doing the timer broadcast. Also currently we are migrating the unpinned timers from an idle to the cpu doing idle load balancing (when all the cpus in the system are idle, there is no idle load balancing cpu and timers get added to the same idle cpu where the request was made. So the existing optimization works only on semi idle system). And In semi idle system, we no longer have periodic ticks on the idle load balancer CPU. Using that cpu will add more delays to the timers than intended (as that cpu's timer base may not be uptodate wrt jiffies etc). This was causing mysterious slowdowns during boot etc. For now, in the semi idle case, use the nearest busy cpu for migrating timers from an idle cpu. This is good for power-savings anyway. Signed-off-by: Venkatesh Pallipadi <venki@google.com> Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> LKML-Reference: <1274486981.2840.46.camel@sbs-t61.sc.intel.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-05-22 00:09:41 +00:00
if (ilb_cpu >= nr_cpu_ids)
return;
sched: Change nohz idle load balancing logic to push model In the new push model, all idle CPUs indeed go into nohz mode. There is still the concept of idle load balancer (performing the load balancing on behalf of all the idle cpu's in the system). Busy CPU kicks the nohz balancer when any of the nohz CPUs need idle load balancing. The kickee CPU does the idle load balancing on behalf of all idle CPUs instead of the normal idle balance. This addresses the below two problems with the current nohz ilb logic: * the idle load balancer continued to have periodic ticks during idle and wokeup frequently, even though it did not have any rebalancing to do on behalf of any of the idle CPUs. * On x86 and CPUs that have APIC timer stoppage on idle CPUs, this periodic wakeup can result in a periodic additional interrupt on a CPU doing the timer broadcast. Also currently we are migrating the unpinned timers from an idle to the cpu doing idle load balancing (when all the cpus in the system are idle, there is no idle load balancing cpu and timers get added to the same idle cpu where the request was made. So the existing optimization works only on semi idle system). And In semi idle system, we no longer have periodic ticks on the idle load balancer CPU. Using that cpu will add more delays to the timers than intended (as that cpu's timer base may not be uptodate wrt jiffies etc). This was causing mysterious slowdowns during boot etc. For now, in the semi idle case, use the nearest busy cpu for migrating timers from an idle cpu. This is good for power-savings anyway. Signed-off-by: Venkatesh Pallipadi <venki@google.com> Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> LKML-Reference: <1274486981.2840.46.camel@sbs-t61.sc.intel.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-05-22 00:09:41 +00:00
flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
if (flags & NOHZ_KICK_MASK)
return;
/*
* Use smp_send_reschedule() instead of resched_cpu().
* This way we generate a sched IPI on the target CPU which
* is idle. And the softirq performing nohz idle load balance
* will be run before returning from the IPI.
*/
smp_send_reschedule(ilb_cpu);
}
/*
* Current decision point for kicking the idle load balancer in the presence
* of idle CPUs in the system.
*/
static void nohz_balancer_kick(struct rq *rq)
{
unsigned long now = jiffies;
struct sched_domain_shared *sds;
struct sched_domain *sd;
int nr_busy, i, cpu = rq->cpu;
unsigned int flags = 0;
if (unlikely(rq->idle_balance))
return;
/*
* We may be recently in ticked or tickless idle mode. At the first
* busy tick after returning from idle, we will update the busy stats.
*/
nohz_balance_exit_idle(rq);
/*
* None are in tickless mode and hence no need for NOHZ idle load
* balancing.
*/
if (likely(!atomic_read(&nohz.nr_cpus)))
return;
if (READ_ONCE(nohz.has_blocked) &&
time_after(now, READ_ONCE(nohz.next_blocked)))
flags = NOHZ_STATS_KICK;
if (time_before(now, nohz.next_balance))
goto out;
sched/fair: Tune down misfit NOHZ kicks In this commit: 3b1baa6496e6 ("sched/fair: Add 'group_misfit_task' load-balance type") we set rq->misfit_task_load whenever the current running task has a utilization greater than 80% of rq->cpu_capacity. A non-zero value in this field enables misfit load balancing. However, if the task being looked at is already running on a CPU of highest capacity, there's nothing more we can do for it. We can currently spot this in update_sd_pick_busiest(), which prevents us from selecting a sched_group of group_type == group_misfit_task as the busiest group, but we don't do any of that in nohz_balancer_kick(). This means that we could repeatedly kick NOHZ CPUs when there's no improvements in terms of load balance to be done. Introduce a check_misfit_status() helper that returns true iff there is a CPU in the system that could give more CPU capacity to a rq's misfit task - IOW, there exists a CPU of higher capacity_orig or the rq's CPU is severely pressured by rt/IRQ. Signed-off-by: Valentin Schneider <valentin.schneider@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andy Lutomirski <luto@kernel.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Dietmar.Eggemann@arm.com Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: morten.rasmussen@arm.com Cc: vincent.guittot@linaro.org Link: https://lkml.kernel.org/r/20190211175946.4961-3-valentin.schneider@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-02-11 17:59:45 +00:00
if (rq->nr_running >= 2) {
flags = NOHZ_KICK_MASK;
goto out;
}
rcu_read_lock();
sd = rcu_dereference(rq->sd);
if (sd) {
/*
* If there's a CFS task and the current CPU has reduced
* capacity; kick the ILB to see if there's a better CPU to run
* on.
*/
if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) {
flags = NOHZ_KICK_MASK;
goto unlock;
}
}
sched/topology: Add lowest CPU asymmetry sched_domain level pointer Add another member to the family of per-cpu sched_domain shortcut pointers. This one, sd_asym_cpucapacity, points to the lowest level at which the SD_ASYM_CPUCAPACITY flag is set. While at it, rename the sd_asym shortcut to sd_asym_packing to avoid confusions. Generally speaking, the largest opportunity to save energy via scheduling comes from a smarter exploitation of heterogeneous platforms (i.e. big.LITTLE). Consequently, the sd_asym_cpucapacity shortcut will be used at first as the lowest domain where Energy-Aware Scheduling (EAS) should be applied. For example, it is possible to apply EAS within a socket on a multi-socket system, as long as each socket has an asymmetric topology. Energy-aware cross-sockets wake-up balancing will only happen when the system is over-utilized, or this_cpu and prev_cpu are in different sockets. Suggested-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Quentin Perret <quentin.perret@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: adharmap@codeaurora.org Cc: chris.redpath@arm.com Cc: currojerez@riseup.net Cc: dietmar.eggemann@arm.com Cc: edubezval@gmail.com Cc: gregkh@linuxfoundation.org Cc: javi.merino@kernel.org Cc: joel@joelfernandes.org Cc: juri.lelli@redhat.com Cc: patrick.bellasi@arm.com Cc: pkondeti@codeaurora.org Cc: rjw@rjwysocki.net Cc: skannan@codeaurora.org Cc: smuckle@google.com Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Cc: viresh.kumar@linaro.org Link: https://lkml.kernel.org/r/20181203095628.11858-7-quentin.perret@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-12-03 09:56:19 +00:00
sd = rcu_dereference(per_cpu(sd_asym_packing, cpu));
if (sd) {
/*
* When ASYM_PACKING; see if there's a more preferred CPU
* currently idle; in which case, kick the ILB to move tasks
* around.
*/
for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) {
if (sched_asym_prefer(i, cpu)) {
flags = NOHZ_KICK_MASK;
goto unlock;
}
}
}
sched/fair: Tune down misfit NOHZ kicks In this commit: 3b1baa6496e6 ("sched/fair: Add 'group_misfit_task' load-balance type") we set rq->misfit_task_load whenever the current running task has a utilization greater than 80% of rq->cpu_capacity. A non-zero value in this field enables misfit load balancing. However, if the task being looked at is already running on a CPU of highest capacity, there's nothing more we can do for it. We can currently spot this in update_sd_pick_busiest(), which prevents us from selecting a sched_group of group_type == group_misfit_task as the busiest group, but we don't do any of that in nohz_balancer_kick(). This means that we could repeatedly kick NOHZ CPUs when there's no improvements in terms of load balance to be done. Introduce a check_misfit_status() helper that returns true iff there is a CPU in the system that could give more CPU capacity to a rq's misfit task - IOW, there exists a CPU of higher capacity_orig or the rq's CPU is severely pressured by rt/IRQ. Signed-off-by: Valentin Schneider <valentin.schneider@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andy Lutomirski <luto@kernel.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Dietmar.Eggemann@arm.com Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: morten.rasmussen@arm.com Cc: vincent.guittot@linaro.org Link: https://lkml.kernel.org/r/20190211175946.4961-3-valentin.schneider@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-02-11 17:59:45 +00:00
sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu));
if (sd) {
/*
* When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
* to run the misfit task on.
*/
if (check_misfit_status(rq, sd)) {
flags = NOHZ_KICK_MASK;
goto unlock;
}
/*
* For asymmetric systems, we do not want to nicely balance
* cache use, instead we want to embrace asymmetry and only
* ensure tasks have enough CPU capacity.
*
* Skip the LLC logic because it's not relevant in that case.
*/
goto unlock;
sched/fair: Tune down misfit NOHZ kicks In this commit: 3b1baa6496e6 ("sched/fair: Add 'group_misfit_task' load-balance type") we set rq->misfit_task_load whenever the current running task has a utilization greater than 80% of rq->cpu_capacity. A non-zero value in this field enables misfit load balancing. However, if the task being looked at is already running on a CPU of highest capacity, there's nothing more we can do for it. We can currently spot this in update_sd_pick_busiest(), which prevents us from selecting a sched_group of group_type == group_misfit_task as the busiest group, but we don't do any of that in nohz_balancer_kick(). This means that we could repeatedly kick NOHZ CPUs when there's no improvements in terms of load balance to be done. Introduce a check_misfit_status() helper that returns true iff there is a CPU in the system that could give more CPU capacity to a rq's misfit task - IOW, there exists a CPU of higher capacity_orig or the rq's CPU is severely pressured by rt/IRQ. Signed-off-by: Valentin Schneider <valentin.schneider@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andy Lutomirski <luto@kernel.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Dietmar.Eggemann@arm.com Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: morten.rasmussen@arm.com Cc: vincent.guittot@linaro.org Link: https://lkml.kernel.org/r/20190211175946.4961-3-valentin.schneider@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-02-11 17:59:45 +00:00
}
sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
if (sds) {
/*
* If there is an imbalance between LLC domains (IOW we could
* increase the overall cache use), we need some less-loaded LLC
* domain to pull some load. Likewise, we may need to spread
* load within the current LLC domain (e.g. packed SMT cores but
* other CPUs are idle). We can't really know from here how busy
* the others are - so just get a nohz balance going if it looks
* like this LLC domain has tasks we could move.
*/
nr_busy = atomic_read(&sds->nr_busy_cpus);
if (nr_busy > 1) {
flags = NOHZ_KICK_MASK;
goto unlock;
}
}
unlock:
rcu_read_unlock();
out:
if (flags)
kick_ilb(flags);
sched: Change nohz idle load balancing logic to push model In the new push model, all idle CPUs indeed go into nohz mode. There is still the concept of idle load balancer (performing the load balancing on behalf of all the idle cpu's in the system). Busy CPU kicks the nohz balancer when any of the nohz CPUs need idle load balancing. The kickee CPU does the idle load balancing on behalf of all idle CPUs instead of the normal idle balance. This addresses the below two problems with the current nohz ilb logic: * the idle load balancer continued to have periodic ticks during idle and wokeup frequently, even though it did not have any rebalancing to do on behalf of any of the idle CPUs. * On x86 and CPUs that have APIC timer stoppage on idle CPUs, this periodic wakeup can result in a periodic additional interrupt on a CPU doing the timer broadcast. Also currently we are migrating the unpinned timers from an idle to the cpu doing idle load balancing (when all the cpus in the system are idle, there is no idle load balancing cpu and timers get added to the same idle cpu where the request was made. So the existing optimization works only on semi idle system). And In semi idle system, we no longer have periodic ticks on the idle load balancer CPU. Using that cpu will add more delays to the timers than intended (as that cpu's timer base may not be uptodate wrt jiffies etc). This was causing mysterious slowdowns during boot etc. For now, in the semi idle case, use the nearest busy cpu for migrating timers from an idle cpu. This is good for power-savings anyway. Signed-off-by: Venkatesh Pallipadi <venki@google.com> Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> LKML-Reference: <1274486981.2840.46.camel@sbs-t61.sc.intel.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-05-22 00:09:41 +00:00
}
static void set_cpu_sd_state_busy(int cpu)
{
struct sched_domain *sd;
rcu_read_lock();
sd = rcu_dereference(per_cpu(sd_llc, cpu));
if (!sd || !sd->nohz_idle)
goto unlock;
sd->nohz_idle = 0;
atomic_inc(&sd->shared->nr_busy_cpus);
unlock:
rcu_read_unlock();
}
void nohz_balance_exit_idle(struct rq *rq)
{
SCHED_WARN_ON(rq != this_rq());
if (likely(!rq->nohz_tick_stopped))
return;
rq->nohz_tick_stopped = 0;
cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
atomic_dec(&nohz.nr_cpus);
set_cpu_sd_state_busy(rq->cpu);
}
static void set_cpu_sd_state_idle(int cpu)
{
struct sched_domain *sd;
rcu_read_lock();
sd = rcu_dereference(per_cpu(sd_llc, cpu));
sched: Fix init NOHZ_IDLE flag On my SMP platform which is made of 5 cores in 2 clusters, I have the nr_busy_cpu field of sched_group_power struct that is not null when the platform is fully idle - which makes the scheduler unhappy. The root cause is: During the boot sequence, some CPUs reach the idle loop and set their NOHZ_IDLE flag while waiting for others CPUs to boot. But the nr_busy_cpus field is initialized later with the assumption that all CPUs are in the busy state whereas some CPUs have already set their NOHZ_IDLE flag. More generally, the NOHZ_IDLE flag must be initialized when new sched_domains are created in order to ensure that NOHZ_IDLE and nr_busy_cpus are aligned. This condition can be ensured by adding a synchronize_rcu() between the destruction of old sched_domains and the creation of new ones so the NOHZ_IDLE flag will not be updated with old sched_domain once it has been initialized. But this solution introduces a additionnal latency in the rebuild sequence that is called during cpu hotplug. As suggested by Frederic Weisbecker, another solution is to have the same rcu lifecycle for both NOHZ_IDLE and sched_domain struct. A new nohz_idle field is added to sched_domain so both status and sched_domain will share the same RCU lifecycle and will be always synchronized. In addition, there is no more need to protect nohz_idle against concurrent access as it is only modified by 2 exclusive functions called by local cpu. This solution has been prefered to the creation of a new struct with an extra pointer indirection for sched_domain. The synchronization is done at the cost of : - An additional indirection and a rcu_dereference for accessing nohz_idle. - We use only the nohz_idle field of the top sched_domain. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: linaro-kernel@lists.linaro.org Cc: peterz@infradead.org Cc: fweisbec@gmail.com Cc: pjt@google.com Cc: rostedt@goodmis.org Cc: efault@gmx.de Link: http://lkml.kernel.org/r/1366729142-14662-1-git-send-email-vincent.guittot@linaro.org [ Fixed !NO_HZ build bug. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-04-23 14:59:02 +00:00
if (!sd || sd->nohz_idle)
goto unlock;
sd->nohz_idle = 1;
atomic_dec(&sd->shared->nr_busy_cpus);
sched: Fix init NOHZ_IDLE flag On my SMP platform which is made of 5 cores in 2 clusters, I have the nr_busy_cpu field of sched_group_power struct that is not null when the platform is fully idle - which makes the scheduler unhappy. The root cause is: During the boot sequence, some CPUs reach the idle loop and set their NOHZ_IDLE flag while waiting for others CPUs to boot. But the nr_busy_cpus field is initialized later with the assumption that all CPUs are in the busy state whereas some CPUs have already set their NOHZ_IDLE flag. More generally, the NOHZ_IDLE flag must be initialized when new sched_domains are created in order to ensure that NOHZ_IDLE and nr_busy_cpus are aligned. This condition can be ensured by adding a synchronize_rcu() between the destruction of old sched_domains and the creation of new ones so the NOHZ_IDLE flag will not be updated with old sched_domain once it has been initialized. But this solution introduces a additionnal latency in the rebuild sequence that is called during cpu hotplug. As suggested by Frederic Weisbecker, another solution is to have the same rcu lifecycle for both NOHZ_IDLE and sched_domain struct. A new nohz_idle field is added to sched_domain so both status and sched_domain will share the same RCU lifecycle and will be always synchronized. In addition, there is no more need to protect nohz_idle against concurrent access as it is only modified by 2 exclusive functions called by local cpu. This solution has been prefered to the creation of a new struct with an extra pointer indirection for sched_domain. The synchronization is done at the cost of : - An additional indirection and a rcu_dereference for accessing nohz_idle. - We use only the nohz_idle field of the top sched_domain. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: linaro-kernel@lists.linaro.org Cc: peterz@infradead.org Cc: fweisbec@gmail.com Cc: pjt@google.com Cc: rostedt@goodmis.org Cc: efault@gmx.de Link: http://lkml.kernel.org/r/1366729142-14662-1-git-send-email-vincent.guittot@linaro.org [ Fixed !NO_HZ build bug. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-04-23 14:59:02 +00:00
unlock:
rcu_read_unlock();
}
/*
* This routine will record that the CPU is going idle with tick stopped.
* This info will be used in performing idle load balancing in the future.
*/
void nohz_balance_enter_idle(int cpu)
{
struct rq *rq = cpu_rq(cpu);
SCHED_WARN_ON(cpu != smp_processor_id());
/* If this CPU is going down, then nothing needs to be done: */
if (!cpu_active(cpu))
return;
/* Spare idle load balancing on CPUs that don't want to be disturbed: */
if (!housekeeping_cpu(cpu, HK_FLAG_SCHED))
return;
/*
* Can be set safely without rq->lock held
* If a clear happens, it will have evaluated last additions because
* rq->lock is held during the check and the clear
*/
rq->has_blocked_load = 1;
/*
* The tick is still stopped but load could have been added in the
* meantime. We set the nohz.has_blocked flag to trig a check of the
* *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
* of nohz.has_blocked can only happen after checking the new load
*/
if (rq->nohz_tick_stopped)
goto out;
/* If we're a completely isolated CPU, we don't play: */
if (on_null_domain(rq))
return;
rq->nohz_tick_stopped = 1;
cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
atomic_inc(&nohz.nr_cpus);
/*
* Ensures that if nohz_idle_balance() fails to observe our
* @idle_cpus_mask store, it must observe the @has_blocked
* store.
*/
smp_mb__after_atomic();
set_cpu_sd_state_idle(cpu);
out:
/*
* Each time a cpu enter idle, we assume that it has blocked load and
* enable the periodic update of the load of idle cpus
*/
WRITE_ONCE(nohz.has_blocked, 1);
}
/*
* Internal function that runs load balance for all idle cpus. The load balance
* can be a simple update of blocked load or a complete load balance with
* tasks movement depending of flags.
* The function returns false if the loop has stopped before running
* through all idle CPUs.
*/
static bool _nohz_idle_balance(struct rq *this_rq, unsigned int flags,
enum cpu_idle_type idle)
sched: Change nohz idle load balancing logic to push model In the new push model, all idle CPUs indeed go into nohz mode. There is still the concept of idle load balancer (performing the load balancing on behalf of all the idle cpu's in the system). Busy CPU kicks the nohz balancer when any of the nohz CPUs need idle load balancing. The kickee CPU does the idle load balancing on behalf of all idle CPUs instead of the normal idle balance. This addresses the below two problems with the current nohz ilb logic: * the idle load balancer continued to have periodic ticks during idle and wokeup frequently, even though it did not have any rebalancing to do on behalf of any of the idle CPUs. * On x86 and CPUs that have APIC timer stoppage on idle CPUs, this periodic wakeup can result in a periodic additional interrupt on a CPU doing the timer broadcast. Also currently we are migrating the unpinned timers from an idle to the cpu doing idle load balancing (when all the cpus in the system are idle, there is no idle load balancing cpu and timers get added to the same idle cpu where the request was made. So the existing optimization works only on semi idle system). And In semi idle system, we no longer have periodic ticks on the idle load balancer CPU. Using that cpu will add more delays to the timers than intended (as that cpu's timer base may not be uptodate wrt jiffies etc). This was causing mysterious slowdowns during boot etc. For now, in the semi idle case, use the nearest busy cpu for migrating timers from an idle cpu. This is good for power-savings anyway. Signed-off-by: Venkatesh Pallipadi <venki@google.com> Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> LKML-Reference: <1274486981.2840.46.camel@sbs-t61.sc.intel.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-05-22 00:09:41 +00:00
{
/* Earliest time when we have to do rebalance again */
unsigned long now = jiffies;
unsigned long next_balance = now + 60*HZ;
bool has_blocked_load = false;
int update_next_balance = 0;
int this_cpu = this_rq->cpu;
int balance_cpu;
int ret = false;
struct rq *rq;
sched: Change nohz idle load balancing logic to push model In the new push model, all idle CPUs indeed go into nohz mode. There is still the concept of idle load balancer (performing the load balancing on behalf of all the idle cpu's in the system). Busy CPU kicks the nohz balancer when any of the nohz CPUs need idle load balancing. The kickee CPU does the idle load balancing on behalf of all idle CPUs instead of the normal idle balance. This addresses the below two problems with the current nohz ilb logic: * the idle load balancer continued to have periodic ticks during idle and wokeup frequently, even though it did not have any rebalancing to do on behalf of any of the idle CPUs. * On x86 and CPUs that have APIC timer stoppage on idle CPUs, this periodic wakeup can result in a periodic additional interrupt on a CPU doing the timer broadcast. Also currently we are migrating the unpinned timers from an idle to the cpu doing idle load balancing (when all the cpus in the system are idle, there is no idle load balancing cpu and timers get added to the same idle cpu where the request was made. So the existing optimization works only on semi idle system). And In semi idle system, we no longer have periodic ticks on the idle load balancer CPU. Using that cpu will add more delays to the timers than intended (as that cpu's timer base may not be uptodate wrt jiffies etc). This was causing mysterious slowdowns during boot etc. For now, in the semi idle case, use the nearest busy cpu for migrating timers from an idle cpu. This is good for power-savings anyway. Signed-off-by: Venkatesh Pallipadi <venki@google.com> Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> LKML-Reference: <1274486981.2840.46.camel@sbs-t61.sc.intel.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-05-22 00:09:41 +00:00
SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
sched: Change nohz idle load balancing logic to push model In the new push model, all idle CPUs indeed go into nohz mode. There is still the concept of idle load balancer (performing the load balancing on behalf of all the idle cpu's in the system). Busy CPU kicks the nohz balancer when any of the nohz CPUs need idle load balancing. The kickee CPU does the idle load balancing on behalf of all idle CPUs instead of the normal idle balance. This addresses the below two problems with the current nohz ilb logic: * the idle load balancer continued to have periodic ticks during idle and wokeup frequently, even though it did not have any rebalancing to do on behalf of any of the idle CPUs. * On x86 and CPUs that have APIC timer stoppage on idle CPUs, this periodic wakeup can result in a periodic additional interrupt on a CPU doing the timer broadcast. Also currently we are migrating the unpinned timers from an idle to the cpu doing idle load balancing (when all the cpus in the system are idle, there is no idle load balancing cpu and timers get added to the same idle cpu where the request was made. So the existing optimization works only on semi idle system). And In semi idle system, we no longer have periodic ticks on the idle load balancer CPU. Using that cpu will add more delays to the timers than intended (as that cpu's timer base may not be uptodate wrt jiffies etc). This was causing mysterious slowdowns during boot etc. For now, in the semi idle case, use the nearest busy cpu for migrating timers from an idle cpu. This is good for power-savings anyway. Signed-off-by: Venkatesh Pallipadi <venki@google.com> Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> LKML-Reference: <1274486981.2840.46.camel@sbs-t61.sc.intel.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-05-22 00:09:41 +00:00
/*
* We assume there will be no idle load after this update and clear
* the has_blocked flag. If a cpu enters idle in the mean time, it will
* set the has_blocked flag and trig another update of idle load.
* Because a cpu that becomes idle, is added to idle_cpus_mask before
* setting the flag, we are sure to not clear the state and not
* check the load of an idle cpu.
*/
WRITE_ONCE(nohz.has_blocked, 0);
/*
* Ensures that if we miss the CPU, we must see the has_blocked
* store from nohz_balance_enter_idle().
*/
smp_mb();
sched: Change nohz idle load balancing logic to push model In the new push model, all idle CPUs indeed go into nohz mode. There is still the concept of idle load balancer (performing the load balancing on behalf of all the idle cpu's in the system). Busy CPU kicks the nohz balancer when any of the nohz CPUs need idle load balancing. The kickee CPU does the idle load balancing on behalf of all idle CPUs instead of the normal idle balance. This addresses the below two problems with the current nohz ilb logic: * the idle load balancer continued to have periodic ticks during idle and wokeup frequently, even though it did not have any rebalancing to do on behalf of any of the idle CPUs. * On x86 and CPUs that have APIC timer stoppage on idle CPUs, this periodic wakeup can result in a periodic additional interrupt on a CPU doing the timer broadcast. Also currently we are migrating the unpinned timers from an idle to the cpu doing idle load balancing (when all the cpus in the system are idle, there is no idle load balancing cpu and timers get added to the same idle cpu where the request was made. So the existing optimization works only on semi idle system). And In semi idle system, we no longer have periodic ticks on the idle load balancer CPU. Using that cpu will add more delays to the timers than intended (as that cpu's timer base may not be uptodate wrt jiffies etc). This was causing mysterious slowdowns during boot etc. For now, in the semi idle case, use the nearest busy cpu for migrating timers from an idle cpu. This is good for power-savings anyway. Signed-off-by: Venkatesh Pallipadi <venki@google.com> Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> LKML-Reference: <1274486981.2840.46.camel@sbs-t61.sc.intel.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-05-22 00:09:41 +00:00
for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
sched: Change nohz idle load balancing logic to push model In the new push model, all idle CPUs indeed go into nohz mode. There is still the concept of idle load balancer (performing the load balancing on behalf of all the idle cpu's in the system). Busy CPU kicks the nohz balancer when any of the nohz CPUs need idle load balancing. The kickee CPU does the idle load balancing on behalf of all idle CPUs instead of the normal idle balance. This addresses the below two problems with the current nohz ilb logic: * the idle load balancer continued to have periodic ticks during idle and wokeup frequently, even though it did not have any rebalancing to do on behalf of any of the idle CPUs. * On x86 and CPUs that have APIC timer stoppage on idle CPUs, this periodic wakeup can result in a periodic additional interrupt on a CPU doing the timer broadcast. Also currently we are migrating the unpinned timers from an idle to the cpu doing idle load balancing (when all the cpus in the system are idle, there is no idle load balancing cpu and timers get added to the same idle cpu where the request was made. So the existing optimization works only on semi idle system). And In semi idle system, we no longer have periodic ticks on the idle load balancer CPU. Using that cpu will add more delays to the timers than intended (as that cpu's timer base may not be uptodate wrt jiffies etc). This was causing mysterious slowdowns during boot etc. For now, in the semi idle case, use the nearest busy cpu for migrating timers from an idle cpu. This is good for power-savings anyway. Signed-off-by: Venkatesh Pallipadi <venki@google.com> Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> LKML-Reference: <1274486981.2840.46.camel@sbs-t61.sc.intel.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-05-22 00:09:41 +00:00
continue;
/*
* If this CPU gets work to do, stop the load balancing
* work being done for other CPUs. Next load
sched: Change nohz idle load balancing logic to push model In the new push model, all idle CPUs indeed go into nohz mode. There is still the concept of idle load balancer (performing the load balancing on behalf of all the idle cpu's in the system). Busy CPU kicks the nohz balancer when any of the nohz CPUs need idle load balancing. The kickee CPU does the idle load balancing on behalf of all idle CPUs instead of the normal idle balance. This addresses the below two problems with the current nohz ilb logic: * the idle load balancer continued to have periodic ticks during idle and wokeup frequently, even though it did not have any rebalancing to do on behalf of any of the idle CPUs. * On x86 and CPUs that have APIC timer stoppage on idle CPUs, this periodic wakeup can result in a periodic additional interrupt on a CPU doing the timer broadcast. Also currently we are migrating the unpinned timers from an idle to the cpu doing idle load balancing (when all the cpus in the system are idle, there is no idle load balancing cpu and timers get added to the same idle cpu where the request was made. So the existing optimization works only on semi idle system). And In semi idle system, we no longer have periodic ticks on the idle load balancer CPU. Using that cpu will add more delays to the timers than intended (as that cpu's timer base may not be uptodate wrt jiffies etc). This was causing mysterious slowdowns during boot etc. For now, in the semi idle case, use the nearest busy cpu for migrating timers from an idle cpu. This is good for power-savings anyway. Signed-off-by: Venkatesh Pallipadi <venki@google.com> Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> LKML-Reference: <1274486981.2840.46.camel@sbs-t61.sc.intel.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-05-22 00:09:41 +00:00
* balancing owner will pick it up.
*/
if (need_resched()) {
has_blocked_load = true;
goto abort;
}
sched: Change nohz idle load balancing logic to push model In the new push model, all idle CPUs indeed go into nohz mode. There is still the concept of idle load balancer (performing the load balancing on behalf of all the idle cpu's in the system). Busy CPU kicks the nohz balancer when any of the nohz CPUs need idle load balancing. The kickee CPU does the idle load balancing on behalf of all idle CPUs instead of the normal idle balance. This addresses the below two problems with the current nohz ilb logic: * the idle load balancer continued to have periodic ticks during idle and wokeup frequently, even though it did not have any rebalancing to do on behalf of any of the idle CPUs. * On x86 and CPUs that have APIC timer stoppage on idle CPUs, this periodic wakeup can result in a periodic additional interrupt on a CPU doing the timer broadcast. Also currently we are migrating the unpinned timers from an idle to the cpu doing idle load balancing (when all the cpus in the system are idle, there is no idle load balancing cpu and timers get added to the same idle cpu where the request was made. So the existing optimization works only on semi idle system). And In semi idle system, we no longer have periodic ticks on the idle load balancer CPU. Using that cpu will add more delays to the timers than intended (as that cpu's timer base may not be uptodate wrt jiffies etc). This was causing mysterious slowdowns during boot etc. For now, in the semi idle case, use the nearest busy cpu for migrating timers from an idle cpu. This is good for power-savings anyway. Signed-off-by: Venkatesh Pallipadi <venki@google.com> Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> LKML-Reference: <1274486981.2840.46.camel@sbs-t61.sc.intel.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-05-22 00:09:41 +00:00
rq = cpu_rq(balance_cpu);
has_blocked_load |= update_nohz_stats(rq, true);
/*
* If time for next balance is due,
* do the balance.
*/
if (time_after_eq(jiffies, rq->next_balance)) {
struct rq_flags rf;
rq_lock_irqsave(rq, &rf);
update_rq_clock(rq);
rq_unlock_irqrestore(rq, &rf);
if (flags & NOHZ_BALANCE_KICK)
rebalance_domains(rq, CPU_IDLE);
}
sched: Change nohz idle load balancing logic to push model In the new push model, all idle CPUs indeed go into nohz mode. There is still the concept of idle load balancer (performing the load balancing on behalf of all the idle cpu's in the system). Busy CPU kicks the nohz balancer when any of the nohz CPUs need idle load balancing. The kickee CPU does the idle load balancing on behalf of all idle CPUs instead of the normal idle balance. This addresses the below two problems with the current nohz ilb logic: * the idle load balancer continued to have periodic ticks during idle and wokeup frequently, even though it did not have any rebalancing to do on behalf of any of the idle CPUs. * On x86 and CPUs that have APIC timer stoppage on idle CPUs, this periodic wakeup can result in a periodic additional interrupt on a CPU doing the timer broadcast. Also currently we are migrating the unpinned timers from an idle to the cpu doing idle load balancing (when all the cpus in the system are idle, there is no idle load balancing cpu and timers get added to the same idle cpu where the request was made. So the existing optimization works only on semi idle system). And In semi idle system, we no longer have periodic ticks on the idle load balancer CPU. Using that cpu will add more delays to the timers than intended (as that cpu's timer base may not be uptodate wrt jiffies etc). This was causing mysterious slowdowns during boot etc. For now, in the semi idle case, use the nearest busy cpu for migrating timers from an idle cpu. This is good for power-savings anyway. Signed-off-by: Venkatesh Pallipadi <venki@google.com> Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> LKML-Reference: <1274486981.2840.46.camel@sbs-t61.sc.intel.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-05-22 00:09:41 +00:00
if (time_after(next_balance, rq->next_balance)) {
next_balance = rq->next_balance;
update_next_balance = 1;
}
sched: Change nohz idle load balancing logic to push model In the new push model, all idle CPUs indeed go into nohz mode. There is still the concept of idle load balancer (performing the load balancing on behalf of all the idle cpu's in the system). Busy CPU kicks the nohz balancer when any of the nohz CPUs need idle load balancing. The kickee CPU does the idle load balancing on behalf of all idle CPUs instead of the normal idle balance. This addresses the below two problems with the current nohz ilb logic: * the idle load balancer continued to have periodic ticks during idle and wokeup frequently, even though it did not have any rebalancing to do on behalf of any of the idle CPUs. * On x86 and CPUs that have APIC timer stoppage on idle CPUs, this periodic wakeup can result in a periodic additional interrupt on a CPU doing the timer broadcast. Also currently we are migrating the unpinned timers from an idle to the cpu doing idle load balancing (when all the cpus in the system are idle, there is no idle load balancing cpu and timers get added to the same idle cpu where the request was made. So the existing optimization works only on semi idle system). And In semi idle system, we no longer have periodic ticks on the idle load balancer CPU. Using that cpu will add more delays to the timers than intended (as that cpu's timer base may not be uptodate wrt jiffies etc). This was causing mysterious slowdowns during boot etc. For now, in the semi idle case, use the nearest busy cpu for migrating timers from an idle cpu. This is good for power-savings anyway. Signed-off-by: Venkatesh Pallipadi <venki@google.com> Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> LKML-Reference: <1274486981.2840.46.camel@sbs-t61.sc.intel.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-05-22 00:09:41 +00:00
}
/* Newly idle CPU doesn't need an update */
if (idle != CPU_NEWLY_IDLE) {
update_blocked_averages(this_cpu);
has_blocked_load |= this_rq->has_blocked_load;
}
if (flags & NOHZ_BALANCE_KICK)
rebalance_domains(this_rq, CPU_IDLE);
WRITE_ONCE(nohz.next_blocked,
now + msecs_to_jiffies(LOAD_AVG_PERIOD));
/* The full idle balance loop has been done */
ret = true;
abort:
/* There is still blocked load, enable periodic update */
if (has_blocked_load)
WRITE_ONCE(nohz.has_blocked, 1);
/*
* next_balance will be updated only when there is a need.
* When the CPU is attached to null domain for ex, it will not be
* updated.
*/
if (likely(update_next_balance))
nohz.next_balance = next_balance;
return ret;
}
/*
* In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
* rebalancing for all the cpus for whom scheduler ticks are stopped.
*/
static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
{
int this_cpu = this_rq->cpu;
unsigned int flags;
if (!(atomic_read(nohz_flags(this_cpu)) & NOHZ_KICK_MASK))
return false;
if (idle != CPU_IDLE) {
atomic_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
return false;
}
/* could be _relaxed() */
flags = atomic_fetch_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
if (!(flags & NOHZ_KICK_MASK))
return false;
_nohz_idle_balance(this_rq, flags, idle);
return true;
sched: Change nohz idle load balancing logic to push model In the new push model, all idle CPUs indeed go into nohz mode. There is still the concept of idle load balancer (performing the load balancing on behalf of all the idle cpu's in the system). Busy CPU kicks the nohz balancer when any of the nohz CPUs need idle load balancing. The kickee CPU does the idle load balancing on behalf of all idle CPUs instead of the normal idle balance. This addresses the below two problems with the current nohz ilb logic: * the idle load balancer continued to have periodic ticks during idle and wokeup frequently, even though it did not have any rebalancing to do on behalf of any of the idle CPUs. * On x86 and CPUs that have APIC timer stoppage on idle CPUs, this periodic wakeup can result in a periodic additional interrupt on a CPU doing the timer broadcast. Also currently we are migrating the unpinned timers from an idle to the cpu doing idle load balancing (when all the cpus in the system are idle, there is no idle load balancing cpu and timers get added to the same idle cpu where the request was made. So the existing optimization works only on semi idle system). And In semi idle system, we no longer have periodic ticks on the idle load balancer CPU. Using that cpu will add more delays to the timers than intended (as that cpu's timer base may not be uptodate wrt jiffies etc). This was causing mysterious slowdowns during boot etc. For now, in the semi idle case, use the nearest busy cpu for migrating timers from an idle cpu. This is good for power-savings anyway. Signed-off-by: Venkatesh Pallipadi <venki@google.com> Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> LKML-Reference: <1274486981.2840.46.camel@sbs-t61.sc.intel.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-05-22 00:09:41 +00:00
}
static void nohz_newidle_balance(struct rq *this_rq)
{
int this_cpu = this_rq->cpu;
/*
* This CPU doesn't want to be disturbed by scheduler
* housekeeping
*/
if (!housekeeping_cpu(this_cpu, HK_FLAG_SCHED))
return;
/* Will wake up very soon. No time for doing anything else*/
if (this_rq->avg_idle < sysctl_sched_migration_cost)
return;
/* Don't need to update blocked load of idle CPUs*/
if (!READ_ONCE(nohz.has_blocked) ||
time_before(jiffies, READ_ONCE(nohz.next_blocked)))
return;
raw_spin_unlock(&this_rq->lock);
/*
* This CPU is going to be idle and blocked load of idle CPUs
* need to be updated. Run the ilb locally as it is a good
* candidate for ilb instead of waking up another idle CPU.
* Kick an normal ilb if we failed to do the update.
*/
if (!_nohz_idle_balance(this_rq, NOHZ_STATS_KICK, CPU_NEWLY_IDLE))
kick_ilb(NOHZ_STATS_KICK);
raw_spin_lock(&this_rq->lock);
}
#else /* !CONFIG_NO_HZ_COMMON */
static inline void nohz_balancer_kick(struct rq *rq) { }
static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
{
return false;
}
static inline void nohz_newidle_balance(struct rq *this_rq) { }
#endif /* CONFIG_NO_HZ_COMMON */
sched: Change nohz idle load balancing logic to push model In the new push model, all idle CPUs indeed go into nohz mode. There is still the concept of idle load balancer (performing the load balancing on behalf of all the idle cpu's in the system). Busy CPU kicks the nohz balancer when any of the nohz CPUs need idle load balancing. The kickee CPU does the idle load balancing on behalf of all idle CPUs instead of the normal idle balance. This addresses the below two problems with the current nohz ilb logic: * the idle load balancer continued to have periodic ticks during idle and wokeup frequently, even though it did not have any rebalancing to do on behalf of any of the idle CPUs. * On x86 and CPUs that have APIC timer stoppage on idle CPUs, this periodic wakeup can result in a periodic additional interrupt on a CPU doing the timer broadcast. Also currently we are migrating the unpinned timers from an idle to the cpu doing idle load balancing (when all the cpus in the system are idle, there is no idle load balancing cpu and timers get added to the same idle cpu where the request was made. So the existing optimization works only on semi idle system). And In semi idle system, we no longer have periodic ticks on the idle load balancer CPU. Using that cpu will add more delays to the timers than intended (as that cpu's timer base may not be uptodate wrt jiffies etc). This was causing mysterious slowdowns during boot etc. For now, in the semi idle case, use the nearest busy cpu for migrating timers from an idle cpu. This is good for power-savings anyway. Signed-off-by: Venkatesh Pallipadi <venki@google.com> Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> LKML-Reference: <1274486981.2840.46.camel@sbs-t61.sc.intel.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-05-22 00:09:41 +00:00
/*
* idle_balance is called by schedule() if this_cpu is about to become
* idle. Attempts to pull tasks from other CPUs.
*
* Returns:
* < 0 - we released the lock and there are !fair tasks present
* 0 - failed, no new tasks
* > 0 - success, new (fair) tasks present
*/
int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
{
unsigned long next_balance = jiffies + HZ;
int this_cpu = this_rq->cpu;
struct sched_domain *sd;
int pulled_task = 0;
u64 curr_cost = 0;
update_misfit_status(NULL, this_rq);
/*
* We must set idle_stamp _before_ calling idle_balance(), such that we
* measure the duration of idle_balance() as idle time.
*/
this_rq->idle_stamp = rq_clock(this_rq);
/*
* Do not pull tasks towards !active CPUs...
*/
if (!cpu_active(this_cpu))
return 0;
/*
* This is OK, because current is on_cpu, which avoids it being picked
* for load-balance and preemption/IRQs are still disabled avoiding
* further scheduler activity on it and we're being very careful to
* re-start the picking loop.
*/
rq_unpin_lock(this_rq, rf);
if (this_rq->avg_idle < sysctl_sched_migration_cost ||
!READ_ONCE(this_rq->rd->overload)) {
rcu_read_lock();
sd = rcu_dereference_check_sched_domain(this_rq->sd);
if (sd)
update_next_balance(sd, &next_balance);
rcu_read_unlock();
nohz_newidle_balance(this_rq);
goto out;
}
raw_spin_unlock(&this_rq->lock);
update_blocked_averages(this_cpu);
rcu_read_lock();
for_each_domain(this_cpu, sd) {
int continue_balancing = 1;
u64 t0, domain_cost;
if (!(sd->flags & SD_LOAD_BALANCE))
continue;
if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
update_next_balance(sd, &next_balance);
break;
}
if (sd->flags & SD_BALANCE_NEWIDLE) {
t0 = sched_clock_cpu(this_cpu);
pulled_task = load_balance(this_cpu, this_rq,
sd, CPU_NEWLY_IDLE,
&continue_balancing);
domain_cost = sched_clock_cpu(this_cpu) - t0;
if (domain_cost > sd->max_newidle_lb_cost)
sd->max_newidle_lb_cost = domain_cost;
curr_cost += domain_cost;
}
update_next_balance(sd, &next_balance);
/*
* Stop searching for tasks to pull if there are
* now runnable tasks on this rq.
*/
if (pulled_task || this_rq->nr_running > 0)
break;
}
rcu_read_unlock();
raw_spin_lock(&this_rq->lock);
if (curr_cost > this_rq->max_idle_balance_cost)
this_rq->max_idle_balance_cost = curr_cost;
out:
/*
* While browsing the domains, we released the rq lock, a task could
* have been enqueued in the meantime. Since we're not going idle,
* pretend we pulled a task.
*/
if (this_rq->cfs.h_nr_running && !pulled_task)
pulled_task = 1;
/* Move the next balance forward */
if (time_after(this_rq->next_balance, next_balance))
this_rq->next_balance = next_balance;
/* Is there a task of a high priority class? */
if (this_rq->nr_running != this_rq->cfs.h_nr_running)
pulled_task = -1;
if (pulled_task)
this_rq->idle_stamp = 0;
rq_repin_lock(this_rq, rf);
return pulled_task;
}
sched: Change nohz idle load balancing logic to push model In the new push model, all idle CPUs indeed go into nohz mode. There is still the concept of idle load balancer (performing the load balancing on behalf of all the idle cpu's in the system). Busy CPU kicks the nohz balancer when any of the nohz CPUs need idle load balancing. The kickee CPU does the idle load balancing on behalf of all idle CPUs instead of the normal idle balance. This addresses the below two problems with the current nohz ilb logic: * the idle load balancer continued to have periodic ticks during idle and wokeup frequently, even though it did not have any rebalancing to do on behalf of any of the idle CPUs. * On x86 and CPUs that have APIC timer stoppage on idle CPUs, this periodic wakeup can result in a periodic additional interrupt on a CPU doing the timer broadcast. Also currently we are migrating the unpinned timers from an idle to the cpu doing idle load balancing (when all the cpus in the system are idle, there is no idle load balancing cpu and timers get added to the same idle cpu where the request was made. So the existing optimization works only on semi idle system). And In semi idle system, we no longer have periodic ticks on the idle load balancer CPU. Using that cpu will add more delays to the timers than intended (as that cpu's timer base may not be uptodate wrt jiffies etc). This was causing mysterious slowdowns during boot etc. For now, in the semi idle case, use the nearest busy cpu for migrating timers from an idle cpu. This is good for power-savings anyway. Signed-off-by: Venkatesh Pallipadi <venki@google.com> Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> LKML-Reference: <1274486981.2840.46.camel@sbs-t61.sc.intel.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-05-22 00:09:41 +00:00
/*
* run_rebalance_domains is triggered when needed from the scheduler tick.
* Also triggered for nohz idle balancing (with nohz_balancing_kick set).
*/
static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
{
struct rq *this_rq = this_rq();
enum cpu_idle_type idle = this_rq->idle_balance ?
CPU_IDLE : CPU_NOT_IDLE;
/*
* If this CPU has a pending nohz_balance_kick, then do the
* balancing on behalf of the other idle CPUs whose ticks are
sched: Improve load balancing in the presence of idle CPUs When a CPU is kicked to do nohz idle balancing, it wakes up to do load balancing on itself, followed by load balancing on behalf of idle CPUs. But it may end up with load after the load balancing attempt on itself. This aborts nohz idle balancing. As a result several idle CPUs are left without tasks till such a time that an ILB CPU finds it unfavorable to pull tasks upon itself. This delays spreading of load across idle CPUs and worse, clutters only a few CPUs with tasks. The effect of the above problem was observed on an SMT8 POWER server with 2 levels of numa domains. Busy loops equal to number of cores were spawned. Since load balancing on fork/exec is discouraged across numa domains, all busy loops would start on one of the numa domains. However it was expected that eventually one busy loop would run per core across all domains due to nohz idle load balancing. But it was observed that it took as long as 10 seconds to spread the load across numa domains. Further investigation showed that this was a consequence of the following: 1. An ILB CPU was chosen from the first numa domain to trigger nohz idle load balancing [Given the experiment, upto 6 CPUs per core could be potentially idle in this domain.] 2. However the ILB CPU would call load_balance() on itself before initiating nohz idle load balancing. 3. Given cores are SMT8, the ILB CPU had enough opportunities to pull tasks from its sibling cores to even out load. 4. Now that the ILB CPU was no longer idle, it would abort nohz idle load balancing As a result the opportunities to spread load across numa domains were lost until such a time that the cores within the first numa domain had equal number of tasks among themselves. This is a pretty bad scenario, since the cores within the first numa domain would have as many as 4 tasks each, while cores in the neighbouring numa domains would all remain idle. Fix this, by checking if a CPU was woken up to do nohz idle load balancing, before it does load balancing upon itself. This way we allow idle CPUs across the system to do load balancing which results in quicker spread of load, instead of performing load balancing within the local sched domain hierarchy of the ILB CPU alone under circumstances such as above. Signed-off-by: Preeti U Murthy <preeti@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Jason Low <jason.low2@hp.com> Cc: benh@kernel.crashing.org Cc: daniel.lezcano@linaro.org Cc: efault@gmx.de Cc: iamjoonsoo.kim@lge.com Cc: morten.rasmussen@arm.com Cc: pjt@google.com Cc: riel@redhat.com Cc: srikar@linux.vnet.ibm.com Cc: svaidy@linux.vnet.ibm.com Cc: tim.c.chen@linux.intel.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/20150326130014.21532.17158.stgit@preeti.in.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-03-26 13:02:44 +00:00
* stopped. Do nohz_idle_balance *before* rebalance_domains to
* give the idle CPUs a chance to load balance. Else we may
sched: Improve load balancing in the presence of idle CPUs When a CPU is kicked to do nohz idle balancing, it wakes up to do load balancing on itself, followed by load balancing on behalf of idle CPUs. But it may end up with load after the load balancing attempt on itself. This aborts nohz idle balancing. As a result several idle CPUs are left without tasks till such a time that an ILB CPU finds it unfavorable to pull tasks upon itself. This delays spreading of load across idle CPUs and worse, clutters only a few CPUs with tasks. The effect of the above problem was observed on an SMT8 POWER server with 2 levels of numa domains. Busy loops equal to number of cores were spawned. Since load balancing on fork/exec is discouraged across numa domains, all busy loops would start on one of the numa domains. However it was expected that eventually one busy loop would run per core across all domains due to nohz idle load balancing. But it was observed that it took as long as 10 seconds to spread the load across numa domains. Further investigation showed that this was a consequence of the following: 1. An ILB CPU was chosen from the first numa domain to trigger nohz idle load balancing [Given the experiment, upto 6 CPUs per core could be potentially idle in this domain.] 2. However the ILB CPU would call load_balance() on itself before initiating nohz idle load balancing. 3. Given cores are SMT8, the ILB CPU had enough opportunities to pull tasks from its sibling cores to even out load. 4. Now that the ILB CPU was no longer idle, it would abort nohz idle load balancing As a result the opportunities to spread load across numa domains were lost until such a time that the cores within the first numa domain had equal number of tasks among themselves. This is a pretty bad scenario, since the cores within the first numa domain would have as many as 4 tasks each, while cores in the neighbouring numa domains would all remain idle. Fix this, by checking if a CPU was woken up to do nohz idle load balancing, before it does load balancing upon itself. This way we allow idle CPUs across the system to do load balancing which results in quicker spread of load, instead of performing load balancing within the local sched domain hierarchy of the ILB CPU alone under circumstances such as above. Signed-off-by: Preeti U Murthy <preeti@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Jason Low <jason.low2@hp.com> Cc: benh@kernel.crashing.org Cc: daniel.lezcano@linaro.org Cc: efault@gmx.de Cc: iamjoonsoo.kim@lge.com Cc: morten.rasmussen@arm.com Cc: pjt@google.com Cc: riel@redhat.com Cc: srikar@linux.vnet.ibm.com Cc: svaidy@linux.vnet.ibm.com Cc: tim.c.chen@linux.intel.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/20150326130014.21532.17158.stgit@preeti.in.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-03-26 13:02:44 +00:00
* load balance only within the local sched_domain hierarchy
* and abort nohz_idle_balance altogether if we pull some load.
*/
if (nohz_idle_balance(this_rq, idle))
return;
/* normal load balance */
update_blocked_averages(this_rq->cpu);
sched: Improve load balancing in the presence of idle CPUs When a CPU is kicked to do nohz idle balancing, it wakes up to do load balancing on itself, followed by load balancing on behalf of idle CPUs. But it may end up with load after the load balancing attempt on itself. This aborts nohz idle balancing. As a result several idle CPUs are left without tasks till such a time that an ILB CPU finds it unfavorable to pull tasks upon itself. This delays spreading of load across idle CPUs and worse, clutters only a few CPUs with tasks. The effect of the above problem was observed on an SMT8 POWER server with 2 levels of numa domains. Busy loops equal to number of cores were spawned. Since load balancing on fork/exec is discouraged across numa domains, all busy loops would start on one of the numa domains. However it was expected that eventually one busy loop would run per core across all domains due to nohz idle load balancing. But it was observed that it took as long as 10 seconds to spread the load across numa domains. Further investigation showed that this was a consequence of the following: 1. An ILB CPU was chosen from the first numa domain to trigger nohz idle load balancing [Given the experiment, upto 6 CPUs per core could be potentially idle in this domain.] 2. However the ILB CPU would call load_balance() on itself before initiating nohz idle load balancing. 3. Given cores are SMT8, the ILB CPU had enough opportunities to pull tasks from its sibling cores to even out load. 4. Now that the ILB CPU was no longer idle, it would abort nohz idle load balancing As a result the opportunities to spread load across numa domains were lost until such a time that the cores within the first numa domain had equal number of tasks among themselves. This is a pretty bad scenario, since the cores within the first numa domain would have as many as 4 tasks each, while cores in the neighbouring numa domains would all remain idle. Fix this, by checking if a CPU was woken up to do nohz idle load balancing, before it does load balancing upon itself. This way we allow idle CPUs across the system to do load balancing which results in quicker spread of load, instead of performing load balancing within the local sched domain hierarchy of the ILB CPU alone under circumstances such as above. Signed-off-by: Preeti U Murthy <preeti@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Jason Low <jason.low2@hp.com> Cc: benh@kernel.crashing.org Cc: daniel.lezcano@linaro.org Cc: efault@gmx.de Cc: iamjoonsoo.kim@lge.com Cc: morten.rasmussen@arm.com Cc: pjt@google.com Cc: riel@redhat.com Cc: srikar@linux.vnet.ibm.com Cc: svaidy@linux.vnet.ibm.com Cc: tim.c.chen@linux.intel.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/20150326130014.21532.17158.stgit@preeti.in.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-03-26 13:02:44 +00:00
rebalance_domains(this_rq, idle);
}
/*
* Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
*/
void trigger_load_balance(struct rq *rq)
{
/* Don't need to rebalance while attached to NULL domain */
if (unlikely(on_null_domain(rq)))
return;
if (time_after_eq(jiffies, rq->next_balance))
raise_softirq(SCHED_SOFTIRQ);
nohz_balancer_kick(rq);
}
static void rq_online_fair(struct rq *rq)
{
update_sysctl();
sched/fair: Disable runtime_enabled on dying rq We kill rq->rd on the CPU_DOWN_PREPARE stage: cpuset_cpu_inactive -> cpuset_update_active_cpus -> partition_sched_domains -> -> cpu_attach_domain -> rq_attach_root -> set_rq_offline This unthrottles all throttled cfs_rqs. But the cpu is still able to call schedule() till take_cpu_down->__cpu_disable() is called from stop_machine. This case the tasks from just unthrottled cfs_rqs are pickable in a standard scheduler way, and they are picked by dying cpu. The cfs_rqs becomes throttled again, and migrate_tasks() in migration_call skips their tasks (one more unthrottle in migrate_tasks()->CPU_DYING does not happen, because rq->rd is already NULL). Patch sets runtime_enabled to zero. This guarantees, the runtime is not accounted, and the cfs_rqs won't exceed given cfs_rq->runtime_remaining = 1, and tasks will be pickable in migrate_tasks(). runtime_enabled is recalculated again when rq becomes online again. Ben Segall also noticed, we always enable runtime in tg_set_cfs_bandwidth(). Actually, we should do that for online cpus only. To prevent races with unthrottle_offline_cfs_rqs() we take get_online_cpus() lock. Reviewed-by: Ben Segall <bsegall@google.com> Reviewed-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Kirill Tkhai <ktkhai@parallels.com> CC: Konstantin Khorenko <khorenko@parallels.com> CC: Paul Turner <pjt@google.com> CC: Mike Galbraith <umgwanakikbuti@gmail.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1403684382.3462.42.camel@tkhai Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-06-25 08:19:42 +00:00
update_runtime_enabled(rq);
}
static void rq_offline_fair(struct rq *rq)
{
update_sysctl();
/* Ensure any throttled groups are reachable by pick_next_task */
unthrottle_offline_cfs_rqs(rq);
}
#endif /* CONFIG_SMP */
/*
2018-02-21 04:17:27 +00:00
* scheduler tick hitting a task of our scheduling class.
*
* NOTE: This function can be called remotely by the tick offload that
* goes along full dynticks. Therefore no local assumption can be made
* and everything must be accessed through the @rq and @curr passed in
* parameters.
*/
static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
{
struct cfs_rq *cfs_rq;
struct sched_entity *se = &curr->se;
for_each_sched_entity(se) {
cfs_rq = cfs_rq_of(se);
entity_tick(cfs_rq, se, queued);
}
if (static_branch_unlikely(&sched_numa_balancing))
task_tick_numa(rq, curr);
sched/fair: Add 'group_misfit_task' load-balance type To maximize throughput in systems with asymmetric CPU capacities (e.g. ARM big.LITTLE) load-balancing has to consider task and CPU utilization as well as per-CPU compute capacity when load-balancing in addition to the current average load based load-balancing policy. Tasks with high utilization that are scheduled on a lower capacity CPU need to be identified and migrated to a higher capacity CPU if possible to maximize throughput. To implement this additional policy an additional group_type (load-balance scenario) is added: 'group_misfit_task'. This represents scenarios where a sched_group has one or more tasks that are not suitable for its per-CPU capacity. 'group_misfit_task' is only considered if the system is not overloaded or imbalanced ('group_imbalanced' or 'group_overloaded'). Identifying misfit tasks requires the rq lock to be held. To avoid taking remote rq locks to examine source sched_groups for misfit tasks, each CPU is responsible for tracking misfit tasks themselves and update the rq->misfit_task flag. This means checking task utilization when tasks are scheduled and on sched_tick. Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dietmar.eggemann@arm.com Cc: gaku.inami.xh@renesas.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1530699470-29808-3-git-send-email-morten.rasmussen@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-07-04 10:17:40 +00:00
update_misfit_status(curr, rq);
sched/fair: Add over-utilization/tipping point indicator Energy-aware scheduling is only meant to be active while the system is _not_ over-utilized. That is, there are spare cycles available to shift tasks around based on their actual utilization to get a more energy-efficient task distribution without depriving any tasks. When above the tipping point task placement is done the traditional way based on load_avg, spreading the tasks across as many cpus as possible based on priority scaled load to preserve smp_nice. Below the tipping point we want to use util_avg instead. We need to define a criteria for when we make the switch. The util_avg for each cpu converges towards 100% regardless of how many additional tasks we may put on it. If we define over-utilized as: sum_{cpus}(rq.cfs.avg.util_avg) + margin > sum_{cpus}(rq.capacity) some individual cpus may be over-utilized running multiple tasks even when the above condition is false. That should be okay as long as we try to spread the tasks out to avoid per-cpu over-utilization as much as possible and if all tasks have the _same_ priority. If the latter isn't true, we have to consider priority to preserve smp_nice. For example, we could have n_cpus nice=-10 util_avg=55% tasks and n_cpus/2 nice=0 util_avg=60% tasks. Balancing based on util_avg we are likely to end up with nice=-10 tasks sharing cpus and nice=0 tasks getting their own as we 1.5*n_cpus tasks in total and 55%+55% is less over-utilized than 55%+60% for those cpus that have to be shared. The system utilization is only 85% of the system capacity, but we are breaking smp_nice. To be sure not to break smp_nice, we have defined over-utilization conservatively as when any cpu in the system is fully utilized at its highest frequency instead: cpu_rq(any).cfs.avg.util_avg + margin > cpu_rq(any).capacity IOW, as soon as one cpu is (nearly) 100% utilized, we switch to load_avg to factor in priority to preserve smp_nice. With this definition, we can skip periodic load-balance as no cpu has an always-running task when the system is not over-utilized. All tasks will be periodic and we can balance them at wake-up. This conservative condition does however mean that some scenarios that could benefit from energy-aware decisions even if one cpu is fully utilized would not get those benefits. For systems where some cpus might have reduced capacity on some cpus (RT-pressure and/or big.LITTLE), we want periodic load-balance checks as soon a just a single cpu is fully utilized as it might one of those with reduced capacity and in that case we want to migrate it. [ peterz: Added a comment explaining why new tasks are not accounted during overutilization detection. ] Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com> Signed-off-by: Quentin Perret <quentin.perret@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: adharmap@codeaurora.org Cc: chris.redpath@arm.com Cc: currojerez@riseup.net Cc: dietmar.eggemann@arm.com Cc: edubezval@gmail.com Cc: gregkh@linuxfoundation.org Cc: javi.merino@kernel.org Cc: joel@joelfernandes.org Cc: juri.lelli@redhat.com Cc: patrick.bellasi@arm.com Cc: pkondeti@codeaurora.org Cc: rjw@rjwysocki.net Cc: skannan@codeaurora.org Cc: smuckle@google.com Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Cc: tkjos@google.com Cc: valentin.schneider@arm.com Cc: vincent.guittot@linaro.org Cc: viresh.kumar@linaro.org Link: https://lkml.kernel.org/r/20181203095628.11858-13-quentin.perret@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-12-03 09:56:25 +00:00
update_overutilized_status(task_rq(curr));
}
/*
* called on fork with the child task as argument from the parent's context
* - child not yet on the tasklist
* - preemption disabled
*/
static void task_fork_fair(struct task_struct *p)
{
struct cfs_rq *cfs_rq;
struct sched_entity *se = &p->se, *curr;
struct rq *rq = this_rq();
struct rq_flags rf;
rq_lock(rq, &rf);
update_rq_clock(rq);
cfs_rq = task_cfs_rq(current);
curr = cfs_rq->curr;
if (curr) {
update_curr(cfs_rq);
se->vruntime = curr->vruntime;
}
place_entity(cfs_rq, se, 1);
if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
/*
* Upon rescheduling, sched_class::put_prev_task() will place
* 'current' within the tree based on its new key value.
*/
swap(curr->vruntime, se->vruntime);
resched_curr(rq);
}
sched: Remove the cfs_rq dependency from set_task_cpu() In order to remove the cfs_rq dependency from set_task_cpu() we need to ensure the task is cfs_rq invariant for all callsites. The simple approach is to substract cfs_rq->min_vruntime from se->vruntime on dequeue, and add cfs_rq->min_vruntime on enqueue. However, this has the downside of breaking FAIR_SLEEPERS since we loose the old vruntime as we only maintain the relative position. To solve this, we observe that we only migrate runnable tasks, we do this using deactivate_task(.sleep=0) and activate_task(.wakeup=0), therefore we can restrain the min_vruntime invariance to that state. The only other case is wakeup balancing, since we want to maintain the old vruntime we cannot make it relative on dequeue, but since we don't migrate inactive tasks, we can do so right before we activate it again. This is where we need the new pre-wakeup hook, we need to call this while still holding the old rq->lock. We could fold it into ->select_task_rq(), but since that has multiple callsites and would obfuscate the locking requirements, that seems like a fudge. This leaves the fork() case, simply make sure that ->task_fork() leaves the ->vruntime in a relative state. This covers all cases where set_task_cpu() gets called, and ensures it sees a relative vruntime. Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> LKML-Reference: <20091216170518.191697025@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-12-16 17:04:41 +00:00
se->vruntime -= cfs_rq->min_vruntime;
rq_unlock(rq, &rf);
}
/*
* Priority of the task has changed. Check to see if we preempt
* the current task.
*/
static void
prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
{
if (!task_on_rq_queued(p))
return;
if (rq->cfs.nr_running == 1)
return;
/*
* Reschedule if we are currently running on this runqueue and
* our priority decreased, or if we are not currently running on
* this runqueue and our priority is higher than the current's
*/
if (rq->curr == p) {
if (p->prio > oldprio)
resched_curr(rq);
} else
check_preempt_curr(rq, p, 0);
}
static inline bool vruntime_normalized(struct task_struct *p)
{
struct sched_entity *se = &p->se;
/*
* In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
* the dequeue_entity(.flags=0) will already have normalized the
* vruntime.
*/
if (p->on_rq)
return true;
/*
* When !on_rq, vruntime of the task has usually NOT been normalized.
* But there are some cases where it has already been normalized:
*
* - A forked child which is waiting for being woken up by
* wake_up_new_task().
* - A task which has been woken up by try_to_wake_up() and
* waiting for actually being woken up by sched_ttwu_pending().
*/
sched/fair: Fix vruntime_normalized() for remote non-migration wakeup When a task which previously ran on a given CPU is remotely queued to wake up on that same CPU, there is a period where the task's state is TASK_WAKING and its vruntime is not normalized. This is not accounted for in vruntime_normalized() which will cause an error in the task's vruntime if it is switched from the fair class during this time. For example if it is boosted to RT priority via rt_mutex_setprio(), rq->min_vruntime will not be subtracted from the task's vruntime but it will be added again when the task returns to the fair class. The task's vruntime will have been erroneously doubled and the effective priority of the task will be reduced. Note this will also lead to inflation of all vruntimes since the doubled vruntime value will become the rq's min_vruntime when other tasks leave the rq. This leads to repeated doubling of the vruntime and priority penalty. Fix this by recognizing a WAKING task's vruntime as normalized only if sched_remote_wakeup is true. This indicates a migration, in which case the vruntime would have been normalized in migrate_task_rq_fair(). Based on a similar patch from John Dias <joaodias@google.com>. Suggested-by: Peter Zijlstra <peterz@infradead.org> Tested-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Signed-off-by: Steve Muckle <smuckle@google.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Chris Redpath <Chris.Redpath@arm.com> Cc: John Dias <joaodias@google.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Miguel de Dios <migueldedios@google.com> Cc: Morten Rasmussen <Morten.Rasmussen@arm.com> Cc: Patrick Bellasi <Patrick.Bellasi@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Quentin Perret <quentin.perret@arm.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Todd Kjos <tkjos@google.com> Cc: kernel-team@android.com Fixes: b5179ac70de8 ("sched/fair: Prepare to fix fairness problems on migration") Link: http://lkml.kernel.org/r/20180831224217.169476-1-smuckle@google.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-08-31 22:42:17 +00:00
if (!se->sum_exec_runtime ||
(p->state == TASK_WAKING && p->sched_remote_wakeup))
return true;
return false;
}
#ifdef CONFIG_FAIR_GROUP_SCHED
/*
* Propagate the changes of the sched_entity across the tg tree to make it
* visible to the root
*/
static void propagate_entity_cfs_rq(struct sched_entity *se)
{
struct cfs_rq *cfs_rq;
/* Start to propagate at parent */
se = se->parent;
for_each_sched_entity(se) {
cfs_rq = cfs_rq_of(se);
if (cfs_rq_throttled(cfs_rq))
break;
update_load_avg(cfs_rq, se, UPDATE_TG);
}
}
#else
static void propagate_entity_cfs_rq(struct sched_entity *se) { }
#endif
static void detach_entity_cfs_rq(struct sched_entity *se)
{
struct cfs_rq *cfs_rq = cfs_rq_of(se);
sched/fair: Rewrite runnable load and utilization average tracking The idea of runnable load average (let runnable time contribute to weight) was proposed by Paul Turner and Ben Segall, and it is still followed by this rewrite. This rewrite aims to solve the following issues: 1. cfs_rq's load average (namely runnable_load_avg and blocked_load_avg) is updated at the granularity of an entity at a time, which results in the cfs_rq's load average is stale or partially updated: at any time, only one entity is up to date, all other entities are effectively lagging behind. This is undesirable. To illustrate, if we have n runnable entities in the cfs_rq, as time elapses, they certainly become outdated: t0: cfs_rq { e1_old, e2_old, ..., en_old } and when we update: t1: update e1, then we have cfs_rq { e1_new, e2_old, ..., en_old } t2: update e2, then we have cfs_rq { e1_old, e2_new, ..., en_old } ... We solve this by combining all runnable entities' load averages together in cfs_rq's avg, and update the cfs_rq's avg as a whole. This is based on the fact that if we regard the update as a function, then: w * update(e) = update(w * e) and update(e1) + update(e2) = update(e1 + e2), then w1 * update(e1) + w2 * update(e2) = update(w1 * e1 + w2 * e2) therefore, by this rewrite, we have an entirely updated cfs_rq at the time we update it: t1: update cfs_rq { e1_new, e2_new, ..., en_new } t2: update cfs_rq { e1_new, e2_new, ..., en_new } ... 2. cfs_rq's load average is different between top rq->cfs_rq and other task_group's per CPU cfs_rqs in whether or not blocked_load_average contributes to the load. The basic idea behind runnable load average (the same for utilization) is that the blocked state is taken into account as opposed to only accounting for the currently runnable state. Therefore, the average should include both the runnable/running and blocked load averages. This rewrite does that. In addition, we also combine runnable/running and blocked averages of all entities into the cfs_rq's average, and update it together at once. This is based on the fact that: update(runnable) + update(blocked) = update(runnable + blocked) This significantly reduces the code as we don't need to separately maintain/update runnable/running load and blocked load. 3. How task_group entities' share is calculated is complex and imprecise. We reduce the complexity in this rewrite to allow a very simple rule: the task_group's load_avg is aggregated from its per CPU cfs_rqs's load_avgs. Then group entity's weight is simply proportional to its own cfs_rq's load_avg / task_group's load_avg. To illustrate, if a task_group has { cfs_rq1, cfs_rq2, ..., cfs_rqn }, then, task_group_avg = cfs_rq1_avg + cfs_rq2_avg + ... + cfs_rqn_avg, then cfs_rqx's entity's share = cfs_rqx_avg / task_group_avg * task_group's share To sum up, this rewrite in principle is equivalent to the current one, but fixes the issues described above. Turns out, it significantly reduces the code complexity and hence increases clarity and efficiency. In addition, the new averages are more smooth/continuous (no spurious spikes and valleys) and updated more consistently and quickly to reflect the load dynamics. As a result, we have less load tracking overhead, better performance, and especially better power efficiency due to more balanced load. Signed-off-by: Yuyang Du <yuyang.du@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: arjan@linux.intel.com Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: fengguang.wu@intel.com Cc: len.brown@intel.com Cc: morten.rasmussen@arm.com Cc: pjt@google.com Cc: rafael.j.wysocki@intel.com Cc: umgwanakikbuti@gmail.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1436918682-4971-3-git-send-email-yuyang.du@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-07-15 00:04:37 +00:00
/* Catch up with the cfs_rq and remove our load when we leave */
update_load_avg(cfs_rq, se, 0);
detach_entity_load_avg(cfs_rq, se);
update_tg_load_avg(cfs_rq, false);
propagate_entity_cfs_rq(se);
}
static void attach_entity_cfs_rq(struct sched_entity *se)
{
struct cfs_rq *cfs_rq = cfs_rq_of(se);
#ifdef CONFIG_FAIR_GROUP_SCHED
/*
* Since the real-depth could have been changed (only FAIR
* class maintain depth value), reset depth properly.
*/
se->depth = se->parent ? se->parent->depth + 1 : 0;
#endif
/* Synchronize entity with its cfs_rq */
update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
attach_entity_load_avg(cfs_rq, se);
update_tg_load_avg(cfs_rq, false);
propagate_entity_cfs_rq(se);
}
static void detach_task_cfs_rq(struct task_struct *p)
{
struct sched_entity *se = &p->se;
struct cfs_rq *cfs_rq = cfs_rq_of(se);
if (!vruntime_normalized(p)) {
/*
* Fix up our vruntime so that the current sleep doesn't
* cause 'unlimited' sleep bonus.
*/
place_entity(cfs_rq, se, 0);
se->vruntime -= cfs_rq->min_vruntime;
}
detach_entity_cfs_rq(se);
}
static void attach_task_cfs_rq(struct task_struct *p)
{
struct sched_entity *se = &p->se;
struct cfs_rq *cfs_rq = cfs_rq_of(se);
attach_entity_cfs_rq(se);
if (!vruntime_normalized(p))
se->vruntime += cfs_rq->min_vruntime;
}
sched/fair: Fix switched_to_fair()'s per entity load tracking Where switched_from_fair() will remove the entity's load from the runqueue, switched_to_fair() does not currently add it back. This means that when a task leaves the fair class for a short duration; say because of PI; we loose its load contribution. This can ripple forward and disturb the load tracking because other operations (enqueue, dequeue) assume its factored in. Only once the runqueue empties will the load tracking recover. When we add it back in, age the per entity average to match up with the runqueue age. This has the obvious problem that if the task leaves the fair class for a significant time, the load will age to 0. Employ the normal migration rule for inter-runqueue moves in task_move_group_fair(). Again, there is the obvious problem of the task migrating while not in the fair class. The alternative solution would be to to omit the chunk in attach_entity_load_avg(), which would effectively reset the timestamp and use whatever avg there was. Signed-off-by: Byungchul Park <byungchul.park@lge.com> [ Rewrote the changelog and comments. ] Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: yuyang.du@intel.com Link: http://lkml.kernel.org/r/1440069720-27038-5-git-send-email-byungchul.park@lge.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-08-20 11:21:59 +00:00
static void switched_from_fair(struct rq *rq, struct task_struct *p)
{
detach_task_cfs_rq(p);
}
static void switched_to_fair(struct rq *rq, struct task_struct *p)
{
attach_task_cfs_rq(p);
if (task_on_rq_queued(p)) {
/*
* We were most likely switched from sched_rt, so
* kick off the schedule if running, otherwise just see
* if we can still preempt the current task.
*/
if (rq->curr == p)
resched_curr(rq);
else
check_preempt_curr(rq, p, 0);
}
}
/* Account for a task changing its policy or group.
*
* This routine is mostly called to set cfs_rq->curr field when a task
* migrates between groups/classes.
*/
static void set_next_task_fair(struct rq *rq, struct task_struct *p, bool first)
{
struct sched_entity *se = &p->se;
#ifdef CONFIG_SMP
if (task_on_rq_queued(p)) {
/*
* Move the next running task to the front of the list, so our
* cfs_tasks list becomes MRU one.
*/
list_move(&se->group_node, &rq->cfs_tasks);
}
#endif
for_each_sched_entity(se) {
struct cfs_rq *cfs_rq = cfs_rq_of(se);
set_next_entity(cfs_rq, se);
/* ensure bandwidth has been allocated on our new cfs_rq */
account_cfs_rq_runtime(cfs_rq, 0);
}
}
void init_cfs_rq(struct cfs_rq *cfs_rq)
{
cfs_rq->tasks_timeline = RB_ROOT_CACHED;
cfs_rq->min_vruntime = (u64)(-(1LL << 20));
#ifndef CONFIG_64BIT
cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
#endif
#ifdef CONFIG_SMP
raw_spin_lock_init(&cfs_rq->removed.lock);
#endif
}
#ifdef CONFIG_FAIR_GROUP_SCHED
static void task_set_group_fair(struct task_struct *p)
{
struct sched_entity *se = &p->se;
set_task_rq(p, task_cpu(p));
se->depth = se->parent ? se->parent->depth + 1 : 0;
}
static void task_move_group_fair(struct task_struct *p)
{
detach_task_cfs_rq(p);
set_task_rq(p, task_cpu(p));
sched/fair: Fix switched_to_fair()'s per entity load tracking Where switched_from_fair() will remove the entity's load from the runqueue, switched_to_fair() does not currently add it back. This means that when a task leaves the fair class for a short duration; say because of PI; we loose its load contribution. This can ripple forward and disturb the load tracking because other operations (enqueue, dequeue) assume its factored in. Only once the runqueue empties will the load tracking recover. When we add it back in, age the per entity average to match up with the runqueue age. This has the obvious problem that if the task leaves the fair class for a significant time, the load will age to 0. Employ the normal migration rule for inter-runqueue moves in task_move_group_fair(). Again, there is the obvious problem of the task migrating while not in the fair class. The alternative solution would be to to omit the chunk in attach_entity_load_avg(), which would effectively reset the timestamp and use whatever avg there was. Signed-off-by: Byungchul Park <byungchul.park@lge.com> [ Rewrote the changelog and comments. ] Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: yuyang.du@intel.com Link: http://lkml.kernel.org/r/1440069720-27038-5-git-send-email-byungchul.park@lge.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-08-20 11:21:59 +00:00
#ifdef CONFIG_SMP
/* Tell se's cfs_rq has been changed -- migrated */
p->se.avg.last_update_time = 0;
#endif
attach_task_cfs_rq(p);
}
static void task_change_group_fair(struct task_struct *p, int type)
{
switch (type) {
case TASK_SET_GROUP:
task_set_group_fair(p);
break;
case TASK_MOVE_GROUP:
task_move_group_fair(p);
break;
}
}
void free_fair_sched_group(struct task_group *tg)
{
int i;
destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
for_each_possible_cpu(i) {
if (tg->cfs_rq)
kfree(tg->cfs_rq[i]);
if (tg->se)
kfree(tg->se[i]);
}
kfree(tg->cfs_rq);
kfree(tg->se);
}
int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
{
struct sched_entity *se;
sched/fair: Fix post_init_entity_util_avg() serialization Chris Wilson reported a divide by 0 at: post_init_entity_util_avg(): > 725 if (cfs_rq->avg.util_avg != 0) { > 726 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight; > -> 727 sa->util_avg /= (cfs_rq->avg.load_avg + 1); > 728 > 729 if (sa->util_avg > cap) > 730 sa->util_avg = cap; > 731 } else { Which given the lack of serialization, and the code generated from update_cfs_rq_load_avg() is entirely possible: if (atomic_long_read(&cfs_rq->removed_load_avg)) { s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0); sa->load_avg = max_t(long, sa->load_avg - r, 0); sa->load_sum = max_t(s64, sa->load_sum - r * LOAD_AVG_MAX, 0); removed_load = 1; } turns into: ffffffff81087064: 49 8b 85 98 00 00 00 mov 0x98(%r13),%rax ffffffff8108706b: 48 85 c0 test %rax,%rax ffffffff8108706e: 74 40 je ffffffff810870b0 ffffffff81087070: 4c 89 f8 mov %r15,%rax ffffffff81087073: 49 87 85 98 00 00 00 xchg %rax,0x98(%r13) ffffffff8108707a: 49 29 45 70 sub %rax,0x70(%r13) ffffffff8108707e: 4c 89 f9 mov %r15,%rcx ffffffff81087081: bb 01 00 00 00 mov $0x1,%ebx ffffffff81087086: 49 83 7d 70 00 cmpq $0x0,0x70(%r13) ffffffff8108708b: 49 0f 49 4d 70 cmovns 0x70(%r13),%rcx Which you'll note ends up with 'sa->load_avg - r' in memory at ffffffff8108707a. By calling post_init_entity_util_avg() under rq->lock we're sure to be fully serialized against PELT updates and cannot observe intermediate state like this. Reported-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Yuyang Du <yuyang.du@intel.com> Cc: bsegall@google.com Cc: morten.rasmussen@arm.com Cc: pjt@google.com Cc: steve.muckle@linaro.org Fixes: 2b8c41daba32 ("sched/fair: Initiate a new task's util avg to a bounded value") Link: http://lkml.kernel.org/r/20160609130750.GQ30909@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-06-09 13:07:50 +00:00
struct cfs_rq *cfs_rq;
int i;
treewide: kzalloc() -> kcalloc() The kzalloc() function has a 2-factor argument form, kcalloc(). This patch replaces cases of: kzalloc(a * b, gfp) with: kcalloc(a * b, gfp) as well as handling cases of: kzalloc(a * b * c, gfp) with: kzalloc(array3_size(a, b, c), gfp) as it's slightly less ugly than: kzalloc_array(array_size(a, b), c, gfp) This does, however, attempt to ignore constant size factors like: kzalloc(4 * 1024, gfp) though any constants defined via macros get caught up in the conversion. Any factors with a sizeof() of "unsigned char", "char", and "u8" were dropped, since they're redundant. The Coccinelle script used for this was: // Fix redundant parens around sizeof(). @@ type TYPE; expression THING, E; @@ ( kzalloc( - (sizeof(TYPE)) * E + sizeof(TYPE) * E , ...) | kzalloc( - (sizeof(THING)) * E + sizeof(THING) * E , ...) ) // Drop single-byte sizes and redundant parens. @@ expression COUNT; typedef u8; typedef __u8; @@ ( kzalloc( - sizeof(u8) * (COUNT) + COUNT , ...) | kzalloc( - sizeof(__u8) * (COUNT) + COUNT , ...) | kzalloc( - sizeof(char) * (COUNT) + COUNT , ...) | kzalloc( - sizeof(unsigned char) * (COUNT) + COUNT , ...) | kzalloc( - sizeof(u8) * COUNT + COUNT , ...) | kzalloc( - sizeof(__u8) * COUNT + COUNT , ...) | kzalloc( - sizeof(char) * COUNT + COUNT , ...) | kzalloc( - sizeof(unsigned char) * COUNT + COUNT , ...) ) // 2-factor product with sizeof(type/expression) and identifier or constant. @@ type TYPE; expression THING; identifier COUNT_ID; constant COUNT_CONST; @@ ( - kzalloc + kcalloc ( - sizeof(TYPE) * (COUNT_ID) + COUNT_ID, sizeof(TYPE) , ...) | - kzalloc + kcalloc ( - sizeof(TYPE) * COUNT_ID + COUNT_ID, sizeof(TYPE) , ...) | - kzalloc + kcalloc ( - sizeof(TYPE) * (COUNT_CONST) + COUNT_CONST, sizeof(TYPE) , ...) | - kzalloc + kcalloc ( - sizeof(TYPE) * COUNT_CONST + COUNT_CONST, sizeof(TYPE) , ...) | - kzalloc + kcalloc ( - sizeof(THING) * (COUNT_ID) + COUNT_ID, sizeof(THING) , ...) | - kzalloc + kcalloc ( - sizeof(THING) * COUNT_ID + COUNT_ID, sizeof(THING) , ...) | - kzalloc + kcalloc ( - sizeof(THING) * (COUNT_CONST) + COUNT_CONST, sizeof(THING) , ...) | - kzalloc + kcalloc ( - sizeof(THING) * COUNT_CONST + COUNT_CONST, sizeof(THING) , ...) ) // 2-factor product, only identifiers. @@ identifier SIZE, COUNT; @@ - kzalloc + kcalloc ( - SIZE * COUNT + COUNT, SIZE , ...) // 3-factor product with 1 sizeof(type) or sizeof(expression), with // redundant parens removed. @@ expression THING; identifier STRIDE, COUNT; type TYPE; @@ ( kzalloc( - sizeof(TYPE) * (COUNT) * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kzalloc( - sizeof(TYPE) * (COUNT) * STRIDE + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kzalloc( - sizeof(TYPE) * COUNT * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kzalloc( - sizeof(TYPE) * COUNT * STRIDE + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kzalloc( - sizeof(THING) * (COUNT) * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) | kzalloc( - sizeof(THING) * (COUNT) * STRIDE + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) | kzalloc( - sizeof(THING) * COUNT * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) | kzalloc( - sizeof(THING) * COUNT * STRIDE + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) ) // 3-factor product with 2 sizeof(variable), with redundant parens removed. @@ expression THING1, THING2; identifier COUNT; type TYPE1, TYPE2; @@ ( kzalloc( - sizeof(TYPE1) * sizeof(TYPE2) * COUNT + array3_size(COUNT, sizeof(TYPE1), sizeof(TYPE2)) , ...) | kzalloc( - sizeof(TYPE1) * sizeof(THING2) * (COUNT) + array3_size(COUNT, sizeof(TYPE1), sizeof(TYPE2)) , ...) | kzalloc( - sizeof(THING1) * sizeof(THING2) * COUNT + array3_size(COUNT, sizeof(THING1), sizeof(THING2)) , ...) | kzalloc( - sizeof(THING1) * sizeof(THING2) * (COUNT) + array3_size(COUNT, sizeof(THING1), sizeof(THING2)) , ...) | kzalloc( - sizeof(TYPE1) * sizeof(THING2) * COUNT + array3_size(COUNT, sizeof(TYPE1), sizeof(THING2)) , ...) | kzalloc( - sizeof(TYPE1) * sizeof(THING2) * (COUNT) + array3_size(COUNT, sizeof(TYPE1), sizeof(THING2)) , ...) ) // 3-factor product, only identifiers, with redundant parens removed. @@ identifier STRIDE, SIZE, COUNT; @@ ( kzalloc( - (COUNT) * STRIDE * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - COUNT * (STRIDE) * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - COUNT * STRIDE * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - (COUNT) * (STRIDE) * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - COUNT * (STRIDE) * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - (COUNT) * STRIDE * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - (COUNT) * (STRIDE) * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - COUNT * STRIDE * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) ) // Any remaining multi-factor products, first at least 3-factor products, // when they're not all constants... @@ expression E1, E2, E3; constant C1, C2, C3; @@ ( kzalloc(C1 * C2 * C3, ...) | kzalloc( - (E1) * E2 * E3 + array3_size(E1, E2, E3) , ...) | kzalloc( - (E1) * (E2) * E3 + array3_size(E1, E2, E3) , ...) | kzalloc( - (E1) * (E2) * (E3) + array3_size(E1, E2, E3) , ...) | kzalloc( - E1 * E2 * E3 + array3_size(E1, E2, E3) , ...) ) // And then all remaining 2 factors products when they're not all constants, // keeping sizeof() as the second factor argument. @@ expression THING, E1, E2; type TYPE; constant C1, C2, C3; @@ ( kzalloc(sizeof(THING) * C2, ...) | kzalloc(sizeof(TYPE) * C2, ...) | kzalloc(C1 * C2 * C3, ...) | kzalloc(C1 * C2, ...) | - kzalloc + kcalloc ( - sizeof(TYPE) * (E2) + E2, sizeof(TYPE) , ...) | - kzalloc + kcalloc ( - sizeof(TYPE) * E2 + E2, sizeof(TYPE) , ...) | - kzalloc + kcalloc ( - sizeof(THING) * (E2) + E2, sizeof(THING) , ...) | - kzalloc + kcalloc ( - sizeof(THING) * E2 + E2, sizeof(THING) , ...) | - kzalloc + kcalloc ( - (E1) * E2 + E1, E2 , ...) | - kzalloc + kcalloc ( - (E1) * (E2) + E1, E2 , ...) | - kzalloc + kcalloc ( - E1 * E2 + E1, E2 , ...) ) Signed-off-by: Kees Cook <keescook@chromium.org>
2018-06-12 21:03:40 +00:00
tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
if (!tg->cfs_rq)
goto err;
treewide: kzalloc() -> kcalloc() The kzalloc() function has a 2-factor argument form, kcalloc(). This patch replaces cases of: kzalloc(a * b, gfp) with: kcalloc(a * b, gfp) as well as handling cases of: kzalloc(a * b * c, gfp) with: kzalloc(array3_size(a, b, c), gfp) as it's slightly less ugly than: kzalloc_array(array_size(a, b), c, gfp) This does, however, attempt to ignore constant size factors like: kzalloc(4 * 1024, gfp) though any constants defined via macros get caught up in the conversion. Any factors with a sizeof() of "unsigned char", "char", and "u8" were dropped, since they're redundant. The Coccinelle script used for this was: // Fix redundant parens around sizeof(). @@ type TYPE; expression THING, E; @@ ( kzalloc( - (sizeof(TYPE)) * E + sizeof(TYPE) * E , ...) | kzalloc( - (sizeof(THING)) * E + sizeof(THING) * E , ...) ) // Drop single-byte sizes and redundant parens. @@ expression COUNT; typedef u8; typedef __u8; @@ ( kzalloc( - sizeof(u8) * (COUNT) + COUNT , ...) | kzalloc( - sizeof(__u8) * (COUNT) + COUNT , ...) | kzalloc( - sizeof(char) * (COUNT) + COUNT , ...) | kzalloc( - sizeof(unsigned char) * (COUNT) + COUNT , ...) | kzalloc( - sizeof(u8) * COUNT + COUNT , ...) | kzalloc( - sizeof(__u8) * COUNT + COUNT , ...) | kzalloc( - sizeof(char) * COUNT + COUNT , ...) | kzalloc( - sizeof(unsigned char) * COUNT + COUNT , ...) ) // 2-factor product with sizeof(type/expression) and identifier or constant. @@ type TYPE; expression THING; identifier COUNT_ID; constant COUNT_CONST; @@ ( - kzalloc + kcalloc ( - sizeof(TYPE) * (COUNT_ID) + COUNT_ID, sizeof(TYPE) , ...) | - kzalloc + kcalloc ( - sizeof(TYPE) * COUNT_ID + COUNT_ID, sizeof(TYPE) , ...) | - kzalloc + kcalloc ( - sizeof(TYPE) * (COUNT_CONST) + COUNT_CONST, sizeof(TYPE) , ...) | - kzalloc + kcalloc ( - sizeof(TYPE) * COUNT_CONST + COUNT_CONST, sizeof(TYPE) , ...) | - kzalloc + kcalloc ( - sizeof(THING) * (COUNT_ID) + COUNT_ID, sizeof(THING) , ...) | - kzalloc + kcalloc ( - sizeof(THING) * COUNT_ID + COUNT_ID, sizeof(THING) , ...) | - kzalloc + kcalloc ( - sizeof(THING) * (COUNT_CONST) + COUNT_CONST, sizeof(THING) , ...) | - kzalloc + kcalloc ( - sizeof(THING) * COUNT_CONST + COUNT_CONST, sizeof(THING) , ...) ) // 2-factor product, only identifiers. @@ identifier SIZE, COUNT; @@ - kzalloc + kcalloc ( - SIZE * COUNT + COUNT, SIZE , ...) // 3-factor product with 1 sizeof(type) or sizeof(expression), with // redundant parens removed. @@ expression THING; identifier STRIDE, COUNT; type TYPE; @@ ( kzalloc( - sizeof(TYPE) * (COUNT) * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kzalloc( - sizeof(TYPE) * (COUNT) * STRIDE + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kzalloc( - sizeof(TYPE) * COUNT * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kzalloc( - sizeof(TYPE) * COUNT * STRIDE + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kzalloc( - sizeof(THING) * (COUNT) * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) | kzalloc( - sizeof(THING) * (COUNT) * STRIDE + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) | kzalloc( - sizeof(THING) * COUNT * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) | kzalloc( - sizeof(THING) * COUNT * STRIDE + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) ) // 3-factor product with 2 sizeof(variable), with redundant parens removed. @@ expression THING1, THING2; identifier COUNT; type TYPE1, TYPE2; @@ ( kzalloc( - sizeof(TYPE1) * sizeof(TYPE2) * COUNT + array3_size(COUNT, sizeof(TYPE1), sizeof(TYPE2)) , ...) | kzalloc( - sizeof(TYPE1) * sizeof(THING2) * (COUNT) + array3_size(COUNT, sizeof(TYPE1), sizeof(TYPE2)) , ...) | kzalloc( - sizeof(THING1) * sizeof(THING2) * COUNT + array3_size(COUNT, sizeof(THING1), sizeof(THING2)) , ...) | kzalloc( - sizeof(THING1) * sizeof(THING2) * (COUNT) + array3_size(COUNT, sizeof(THING1), sizeof(THING2)) , ...) | kzalloc( - sizeof(TYPE1) * sizeof(THING2) * COUNT + array3_size(COUNT, sizeof(TYPE1), sizeof(THING2)) , ...) | kzalloc( - sizeof(TYPE1) * sizeof(THING2) * (COUNT) + array3_size(COUNT, sizeof(TYPE1), sizeof(THING2)) , ...) ) // 3-factor product, only identifiers, with redundant parens removed. @@ identifier STRIDE, SIZE, COUNT; @@ ( kzalloc( - (COUNT) * STRIDE * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - COUNT * (STRIDE) * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - COUNT * STRIDE * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - (COUNT) * (STRIDE) * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - COUNT * (STRIDE) * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - (COUNT) * STRIDE * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - (COUNT) * (STRIDE) * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - COUNT * STRIDE * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) ) // Any remaining multi-factor products, first at least 3-factor products, // when they're not all constants... @@ expression E1, E2, E3; constant C1, C2, C3; @@ ( kzalloc(C1 * C2 * C3, ...) | kzalloc( - (E1) * E2 * E3 + array3_size(E1, E2, E3) , ...) | kzalloc( - (E1) * (E2) * E3 + array3_size(E1, E2, E3) , ...) | kzalloc( - (E1) * (E2) * (E3) + array3_size(E1, E2, E3) , ...) | kzalloc( - E1 * E2 * E3 + array3_size(E1, E2, E3) , ...) ) // And then all remaining 2 factors products when they're not all constants, // keeping sizeof() as the second factor argument. @@ expression THING, E1, E2; type TYPE; constant C1, C2, C3; @@ ( kzalloc(sizeof(THING) * C2, ...) | kzalloc(sizeof(TYPE) * C2, ...) | kzalloc(C1 * C2 * C3, ...) | kzalloc(C1 * C2, ...) | - kzalloc + kcalloc ( - sizeof(TYPE) * (E2) + E2, sizeof(TYPE) , ...) | - kzalloc + kcalloc ( - sizeof(TYPE) * E2 + E2, sizeof(TYPE) , ...) | - kzalloc + kcalloc ( - sizeof(THING) * (E2) + E2, sizeof(THING) , ...) | - kzalloc + kcalloc ( - sizeof(THING) * E2 + E2, sizeof(THING) , ...) | - kzalloc + kcalloc ( - (E1) * E2 + E1, E2 , ...) | - kzalloc + kcalloc ( - (E1) * (E2) + E1, E2 , ...) | - kzalloc + kcalloc ( - E1 * E2 + E1, E2 , ...) ) Signed-off-by: Kees Cook <keescook@chromium.org>
2018-06-12 21:03:40 +00:00
tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
if (!tg->se)
goto err;
tg->shares = NICE_0_LOAD;
init_cfs_bandwidth(tg_cfs_bandwidth(tg));
for_each_possible_cpu(i) {
cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
GFP_KERNEL, cpu_to_node(i));
if (!cfs_rq)
goto err;
se = kzalloc_node(sizeof(struct sched_entity),
GFP_KERNEL, cpu_to_node(i));
if (!se)
goto err_free_rq;
init_cfs_rq(cfs_rq);
init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
init_entity_runnable_average(se);
}
return 1;
err_free_rq:
kfree(cfs_rq);
err:
return 0;
}
void online_fair_sched_group(struct task_group *tg)
{
struct sched_entity *se;
struct rq_flags rf;
struct rq *rq;
int i;
for_each_possible_cpu(i) {
rq = cpu_rq(i);
se = tg->se[i];
rq_lock_irq(rq, &rf);
update_rq_clock(rq);
attach_entity_cfs_rq(se);
sync_throttle(tg, i);
rq_unlock_irq(rq, &rf);
}
}
void unregister_fair_sched_group(struct task_group *tg)
{
unsigned long flags;
struct rq *rq;
int cpu;
for_each_possible_cpu(cpu) {
if (tg->se[cpu])
remove_entity_load_avg(tg->se[cpu]);
/*
* Only empty task groups can be destroyed; so we can speculatively
* check on_list without danger of it being re-added.
*/
if (!tg->cfs_rq[cpu]->on_list)
continue;
rq = cpu_rq(cpu);
raw_spin_lock_irqsave(&rq->lock, flags);
list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
raw_spin_unlock_irqrestore(&rq->lock, flags);
}
}
void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
struct sched_entity *se, int cpu,
struct sched_entity *parent)
{
struct rq *rq = cpu_rq(cpu);
cfs_rq->tg = tg;
cfs_rq->rq = rq;
init_cfs_rq_runtime(cfs_rq);
tg->cfs_rq[cpu] = cfs_rq;
tg->se[cpu] = se;
/* se could be NULL for root_task_group */
if (!se)
return;
if (!parent) {
se->cfs_rq = &rq->cfs;
se->depth = 0;
} else {
se->cfs_rq = parent->my_q;
se->depth = parent->depth + 1;
}
se->my_q = cfs_rq;
/* guarantee group entities always have weight */
update_load_set(&se->load, NICE_0_LOAD);
se->parent = parent;
}
static DEFINE_MUTEX(shares_mutex);
int sched_group_set_shares(struct task_group *tg, unsigned long shares)
{
int i;
/*
* We can't change the weight of the root cgroup.
*/
if (!tg->se[0])
return -EINVAL;
shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
mutex_lock(&shares_mutex);
if (tg->shares == shares)
goto done;
tg->shares = shares;
for_each_possible_cpu(i) {
struct rq *rq = cpu_rq(i);
struct sched_entity *se = tg->se[i];
struct rq_flags rf;
/* Propagate contribution to hierarchy */
rq_lock_irqsave(rq, &rf);
update_rq_clock(rq);
sched/core: Fix group_entity's share update The update of the share of a cfs_rq is done when its load_avg is updated but before the group_entity's load_avg has been updated for the past time slot. This generates wrong load_avg accounting which can be significant when small tasks are involved in the scheduling. Let take the example of a task a that is dequeued of its task group A: root (cfs_rq) \ (se) A (cfs_rq) \ (se) a Task "a" was the only task in task group A which becomes idle when a is dequeued. We have the sequence: - dequeue_entity a->se - update_load_avg(a->se) - dequeue_entity_load_avg(A->cfs_rq, a->se) - update_cfs_shares(A->cfs_rq) A->cfs_rq->load.weight == 0 A->se->load.weight is updated with the new share (0 in this case) - dequeue_entity A->se - update_load_avg(A->se) but its weight is now null so the last time slot (up to a tick) will be accounted with a weight of 0 instead of its real weight during the time slot. The last time slot will be accounted as an idle one whereas it was a running one. If the running time of task a is short enough that no tick happens when it runs, all running time of group entity A->se will be accounted as idle time. Instead, we should update the share of a cfs_rq (in fact the weight of its group entity) only after having updated the load_avg of the group_entity. update_cfs_shares() now takes the sched_entity as a parameter instead of the cfs_rq, and the weight of the group_entity is updated only once its load_avg has been synced with current time. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: pjt@google.com Link: http://lkml.kernel.org/r/1482335426-7664-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-12-21 15:50:26 +00:00
for_each_sched_entity(se) {
update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
update_cfs_group(se);
sched/core: Fix group_entity's share update The update of the share of a cfs_rq is done when its load_avg is updated but before the group_entity's load_avg has been updated for the past time slot. This generates wrong load_avg accounting which can be significant when small tasks are involved in the scheduling. Let take the example of a task a that is dequeued of its task group A: root (cfs_rq) \ (se) A (cfs_rq) \ (se) a Task "a" was the only task in task group A which becomes idle when a is dequeued. We have the sequence: - dequeue_entity a->se - update_load_avg(a->se) - dequeue_entity_load_avg(A->cfs_rq, a->se) - update_cfs_shares(A->cfs_rq) A->cfs_rq->load.weight == 0 A->se->load.weight is updated with the new share (0 in this case) - dequeue_entity A->se - update_load_avg(A->se) but its weight is now null so the last time slot (up to a tick) will be accounted with a weight of 0 instead of its real weight during the time slot. The last time slot will be accounted as an idle one whereas it was a running one. If the running time of task a is short enough that no tick happens when it runs, all running time of group entity A->se will be accounted as idle time. Instead, we should update the share of a cfs_rq (in fact the weight of its group entity) only after having updated the load_avg of the group_entity. update_cfs_shares() now takes the sched_entity as a parameter instead of the cfs_rq, and the weight of the group_entity is updated only once its load_avg has been synced with current time. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: pjt@google.com Link: http://lkml.kernel.org/r/1482335426-7664-1-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-12-21 15:50:26 +00:00
}
rq_unlock_irqrestore(rq, &rf);
}
done:
mutex_unlock(&shares_mutex);
return 0;
}
#else /* CONFIG_FAIR_GROUP_SCHED */
void free_fair_sched_group(struct task_group *tg) { }
int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
{
return 1;
}
void online_fair_sched_group(struct task_group *tg) { }
void unregister_fair_sched_group(struct task_group *tg) { }
#endif /* CONFIG_FAIR_GROUP_SCHED */
static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
{
struct sched_entity *se = &task->se;
unsigned int rr_interval = 0;
/*
* Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
* idle runqueue:
*/
if (rq->cfs.load.weight)
rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
return rr_interval;
}
/*
* All the scheduling class methods:
*/
const struct sched_class fair_sched_class = {
.next = &idle_sched_class,
.enqueue_task = enqueue_task_fair,
.dequeue_task = dequeue_task_fair,
.yield_task = yield_task_fair,
.yield_to_task = yield_to_task_fair,
.check_preempt_curr = check_preempt_wakeup,
.pick_next_task = __pick_next_task_fair,
.put_prev_task = put_prev_task_fair,
.set_next_task = set_next_task_fair,
#ifdef CONFIG_SMP
sched: Fix pick_next_task() vs 'change' pattern race Commit 67692435c411 ("sched: Rework pick_next_task() slow-path") inadvertly introduced a race because it changed a previously unexplored dependency between dropping the rq->lock and sched_class::put_prev_task(). The comments about dropping rq->lock, in for example newidle_balance(), only mentions the task being current and ->on_cpu being set. But when we look at the 'change' pattern (in for example sched_setnuma()): queued = task_on_rq_queued(p); /* p->on_rq == TASK_ON_RQ_QUEUED */ running = task_current(rq, p); /* rq->curr == p */ if (queued) dequeue_task(...); if (running) put_prev_task(...); /* change task properties */ if (queued) enqueue_task(...); if (running) set_next_task(...); It becomes obvious that if we do this after put_prev_task() has already been called on @p, things go sideways. This is exactly what the commit in question allows to happen when it does: prev->sched_class->put_prev_task(rq, prev, rf); if (!rq->nr_running) newidle_balance(rq, rf); The newidle_balance() call will drop rq->lock after we've called put_prev_task() and that allows the above 'change' pattern to interleave and mess up the state. Furthermore, it turns out we lost the RT-pull when we put the last DL task. Fix both problems by extracting the balancing from put_prev_task() and doing a multi-class balance() pass before put_prev_task(). Fixes: 67692435c411 ("sched: Rework pick_next_task() slow-path") Reported-by: Quentin Perret <qperret@google.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Tested-by: Quentin Perret <qperret@google.com> Tested-by: Valentin Schneider <valentin.schneider@arm.com>
2019-11-08 10:11:52 +00:00
.balance = balance_fair,
.select_task_rq = select_task_rq_fair,
.migrate_task_rq = migrate_task_rq_fair,
.rq_online = rq_online_fair,
.rq_offline = rq_offline_fair,
sched: Remove the cfs_rq dependency from set_task_cpu() In order to remove the cfs_rq dependency from set_task_cpu() we need to ensure the task is cfs_rq invariant for all callsites. The simple approach is to substract cfs_rq->min_vruntime from se->vruntime on dequeue, and add cfs_rq->min_vruntime on enqueue. However, this has the downside of breaking FAIR_SLEEPERS since we loose the old vruntime as we only maintain the relative position. To solve this, we observe that we only migrate runnable tasks, we do this using deactivate_task(.sleep=0) and activate_task(.wakeup=0), therefore we can restrain the min_vruntime invariance to that state. The only other case is wakeup balancing, since we want to maintain the old vruntime we cannot make it relative on dequeue, but since we don't migrate inactive tasks, we can do so right before we activate it again. This is where we need the new pre-wakeup hook, we need to call this while still holding the old rq->lock. We could fold it into ->select_task_rq(), but since that has multiple callsites and would obfuscate the locking requirements, that seems like a fudge. This leaves the fork() case, simply make sure that ->task_fork() leaves the ->vruntime in a relative state. This covers all cases where set_task_cpu() gets called, and ensures it sees a relative vruntime. Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> LKML-Reference: <20091216170518.191697025@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-12-16 17:04:41 +00:00
.task_dead = task_dead_fair,
.set_cpus_allowed = set_cpus_allowed_common,
#endif
.task_tick = task_tick_fair,
.task_fork = task_fork_fair,
.prio_changed = prio_changed_fair,
.switched_from = switched_from_fair,
.switched_to = switched_to_fair,
.get_rr_interval = get_rr_interval_fair,
sched/cputime: Fix clock_nanosleep()/clock_gettime() inconsistency Commit d670ec13178d0 "posix-cpu-timers: Cure SMP wobbles" fixes one glibc test case in cost of breaking another one. After that commit, calling clock_nanosleep(TIMER_ABSTIME, X) and then clock_gettime(&Y) can result of Y time being smaller than X time. Reproducer/tester can be found further below, it can be compiled and ran by: gcc -o tst-cpuclock2 tst-cpuclock2.c -pthread while ./tst-cpuclock2 ; do : ; done This reproducer, when running on a buggy kernel, will complain about "clock_gettime difference too small". Issue happens because on start in thread_group_cputimer() we initialize sum_exec_runtime of cputimer with threads runtime not yet accounted and then add the threads runtime to running cputimer again on scheduler tick, making it's sum_exec_runtime bigger than actual threads runtime. KOSAKI Motohiro posted a fix for this problem, but that patch was never applied: https://lkml.org/lkml/2013/5/26/191 . This patch takes different approach to cure the problem. It calls update_curr() when cputimer starts, that assure we will have updated stats of running threads and on the next schedule tick we will account only the runtime that elapsed from cputimer start. That also assure we have consistent state between cpu times of individual threads and cpu time of the process consisted by those threads. Full reproducer (tst-cpuclock2.c): #define _GNU_SOURCE #include <unistd.h> #include <sys/syscall.h> #include <stdio.h> #include <time.h> #include <pthread.h> #include <stdint.h> #include <inttypes.h> /* Parameters for the Linux kernel ABI for CPU clocks. */ #define CPUCLOCK_SCHED 2 #define MAKE_PROCESS_CPUCLOCK(pid, clock) \ ((~(clockid_t) (pid) << 3) | (clockid_t) (clock)) static pthread_barrier_t barrier; /* Help advance the clock. */ static void *chew_cpu(void *arg) { pthread_barrier_wait(&barrier); while (1) ; return NULL; } /* Don't use the glibc wrapper. */ static int do_nanosleep(int flags, const struct timespec *req) { clockid_t clock_id = MAKE_PROCESS_CPUCLOCK(0, CPUCLOCK_SCHED); return syscall(SYS_clock_nanosleep, clock_id, flags, req, NULL); } static int64_t tsdiff(const struct timespec *before, const struct timespec *after) { int64_t before_i = before->tv_sec * 1000000000ULL + before->tv_nsec; int64_t after_i = after->tv_sec * 1000000000ULL + after->tv_nsec; return after_i - before_i; } int main(void) { int result = 0; pthread_t th; pthread_barrier_init(&barrier, NULL, 2); if (pthread_create(&th, NULL, chew_cpu, NULL) != 0) { perror("pthread_create"); return 1; } pthread_barrier_wait(&barrier); /* The test. */ struct timespec before, after, sleeptimeabs; int64_t sleepdiff, diffabs; const struct timespec sleeptime = {.tv_sec = 0,.tv_nsec = 100000000 }; /* The relative nanosleep. Not sure why this is needed, but its presence seems to make it easier to reproduce the problem. */ if (do_nanosleep(0, &sleeptime) != 0) { perror("clock_nanosleep"); return 1; } /* Get the current time. */ if (clock_gettime(CLOCK_PROCESS_CPUTIME_ID, &before) < 0) { perror("clock_gettime[2]"); return 1; } /* Compute the absolute sleep time based on the current time. */ uint64_t nsec = before.tv_nsec + sleeptime.tv_nsec; sleeptimeabs.tv_sec = before.tv_sec + nsec / 1000000000; sleeptimeabs.tv_nsec = nsec % 1000000000; /* Sleep for the computed time. */ if (do_nanosleep(TIMER_ABSTIME, &sleeptimeabs) != 0) { perror("absolute clock_nanosleep"); return 1; } /* Get the time after the sleep. */ if (clock_gettime(CLOCK_PROCESS_CPUTIME_ID, &after) < 0) { perror("clock_gettime[3]"); return 1; } /* The time after sleep should always be equal to or after the absolute sleep time passed to clock_nanosleep. */ sleepdiff = tsdiff(&sleeptimeabs, &after); if (sleepdiff < 0) { printf("absolute clock_nanosleep woke too early: %" PRId64 "\n", sleepdiff); result = 1; printf("Before %llu.%09llu\n", before.tv_sec, before.tv_nsec); printf("After %llu.%09llu\n", after.tv_sec, after.tv_nsec); printf("Sleep %llu.%09llu\n", sleeptimeabs.tv_sec, sleeptimeabs.tv_nsec); } /* The difference between the timestamps taken before and after the clock_nanosleep call should be equal to or more than the duration of the sleep. */ diffabs = tsdiff(&before, &after); if (diffabs < sleeptime.tv_nsec) { printf("clock_gettime difference too small: %" PRId64 "\n", diffabs); result = 1; } pthread_cancel(th); return result; } Signed-off-by: Stanislaw Gruszka <sgruszka@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Frederic Weisbecker <fweisbec@gmail.com> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/20141112155843.GA24803@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-11-12 15:58:44 +00:00
.update_curr = update_curr_fair,
#ifdef CONFIG_FAIR_GROUP_SCHED
.task_change_group = task_change_group_fair,
#endif
sched/cpufreq, sched/uclamp: Add clamps for FAIR and RT tasks Each time a frequency update is required via schedutil, a frequency is selected to (possibly) satisfy the utilization reported by each scheduling class and irqs. However, when utilization clamping is in use, the frequency selection should consider userspace utilization clamping hints. This will allow, for example, to: - boost tasks which are directly affecting the user experience by running them at least at a minimum "requested" frequency - cap low priority tasks not directly affecting the user experience by running them only up to a maximum "allowed" frequency These constraints are meant to support a per-task based tuning of the frequency selection thus supporting a fine grained definition of performance boosting vs energy saving strategies in kernel space. Add support to clamp the utilization of RUNNABLE FAIR and RT tasks within the boundaries defined by their aggregated utilization clamp constraints. Do that by considering the max(min_util, max_util) to give boosted tasks the performance they need even when they happen to be co-scheduled with other capped tasks. Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alessio Balsini <balsini@android.com> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Joel Fernandes <joelaf@google.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Quentin Perret <quentin.perret@arm.com> Cc: Rafael J . Wysocki <rafael.j.wysocki@intel.com> Cc: Steve Muckle <smuckle@google.com> Cc: Suren Baghdasaryan <surenb@google.com> Cc: Tejun Heo <tj@kernel.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Todd Kjos <tkjos@google.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Link: https://lkml.kernel.org/r/20190621084217.8167-10-patrick.bellasi@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-06-21 08:42:10 +00:00
#ifdef CONFIG_UCLAMP_TASK
.uclamp_enabled = 1,
#endif
};
#ifdef CONFIG_SCHED_DEBUG
void print_cfs_stats(struct seq_file *m, int cpu)
{
struct cfs_rq *cfs_rq, *pos;
rcu_read_lock();
for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
print_cfs_rq(m, cpu, cfs_rq);
rcu_read_unlock();
}
sched/numa: Fix numa balancing stats in /proc/pid/sched Commit 44dba3d5d6a1 ("sched: Refactor task_struct to use numa_faults instead of numa_* pointers") modified the way tsk->numa_faults stats are accounted. However that commit never touched show_numa_stats() that is displayed in /proc/pid/sched and thus the numbers displayed in /proc/pid/sched don't match the actual numbers. Fix it by making sure that /proc/pid/sched reflects the task fault numbers. Also add group fault stats too. Also couple of more modifications are added here: 1. Format changes: - Previously we would list two entries per node, one for private and one for shared. Also the home node info was listed in each entry. - Now preferred node, total_faults and current node are displayed separately. - Now there is one entry per node, that lists private,shared task and group faults. 2. Unit changes: - p->numa_pages_migrated was getting reset after every read of /proc/pid/sched. It's more useful to have absolute numbers since differential migrations between two accesses can be more easily calculated. Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Acked-by: Rik van Riel <riel@redhat.com> Cc: Iulia Manda <iulia.manda21@gmail.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1435252903-1081-4-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-06-25 17:21:43 +00:00
#ifdef CONFIG_NUMA_BALANCING
void show_numa_stats(struct task_struct *p, struct seq_file *m)
{
int node;
unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
struct numa_group *ng;
sched/numa: Fix numa balancing stats in /proc/pid/sched Commit 44dba3d5d6a1 ("sched: Refactor task_struct to use numa_faults instead of numa_* pointers") modified the way tsk->numa_faults stats are accounted. However that commit never touched show_numa_stats() that is displayed in /proc/pid/sched and thus the numbers displayed in /proc/pid/sched don't match the actual numbers. Fix it by making sure that /proc/pid/sched reflects the task fault numbers. Also add group fault stats too. Also couple of more modifications are added here: 1. Format changes: - Previously we would list two entries per node, one for private and one for shared. Also the home node info was listed in each entry. - Now preferred node, total_faults and current node are displayed separately. - Now there is one entry per node, that lists private,shared task and group faults. 2. Unit changes: - p->numa_pages_migrated was getting reset after every read of /proc/pid/sched. It's more useful to have absolute numbers since differential migrations between two accesses can be more easily calculated. Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Acked-by: Rik van Riel <riel@redhat.com> Cc: Iulia Manda <iulia.manda21@gmail.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1435252903-1081-4-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-06-25 17:21:43 +00:00
rcu_read_lock();
ng = rcu_dereference(p->numa_group);
sched/numa: Fix numa balancing stats in /proc/pid/sched Commit 44dba3d5d6a1 ("sched: Refactor task_struct to use numa_faults instead of numa_* pointers") modified the way tsk->numa_faults stats are accounted. However that commit never touched show_numa_stats() that is displayed in /proc/pid/sched and thus the numbers displayed in /proc/pid/sched don't match the actual numbers. Fix it by making sure that /proc/pid/sched reflects the task fault numbers. Also add group fault stats too. Also couple of more modifications are added here: 1. Format changes: - Previously we would list two entries per node, one for private and one for shared. Also the home node info was listed in each entry. - Now preferred node, total_faults and current node are displayed separately. - Now there is one entry per node, that lists private,shared task and group faults. 2. Unit changes: - p->numa_pages_migrated was getting reset after every read of /proc/pid/sched. It's more useful to have absolute numbers since differential migrations between two accesses can be more easily calculated. Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Acked-by: Rik van Riel <riel@redhat.com> Cc: Iulia Manda <iulia.manda21@gmail.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1435252903-1081-4-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-06-25 17:21:43 +00:00
for_each_online_node(node) {
if (p->numa_faults) {
tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
}
if (ng) {
gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)],
gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
sched/numa: Fix numa balancing stats in /proc/pid/sched Commit 44dba3d5d6a1 ("sched: Refactor task_struct to use numa_faults instead of numa_* pointers") modified the way tsk->numa_faults stats are accounted. However that commit never touched show_numa_stats() that is displayed in /proc/pid/sched and thus the numbers displayed in /proc/pid/sched don't match the actual numbers. Fix it by making sure that /proc/pid/sched reflects the task fault numbers. Also add group fault stats too. Also couple of more modifications are added here: 1. Format changes: - Previously we would list two entries per node, one for private and one for shared. Also the home node info was listed in each entry. - Now preferred node, total_faults and current node are displayed separately. - Now there is one entry per node, that lists private,shared task and group faults. 2. Unit changes: - p->numa_pages_migrated was getting reset after every read of /proc/pid/sched. It's more useful to have absolute numbers since differential migrations between two accesses can be more easily calculated. Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Acked-by: Rik van Riel <riel@redhat.com> Cc: Iulia Manda <iulia.manda21@gmail.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1435252903-1081-4-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-06-25 17:21:43 +00:00
}
print_numa_stats(m, node, tsf, tpf, gsf, gpf);
}
rcu_read_unlock();
sched/numa: Fix numa balancing stats in /proc/pid/sched Commit 44dba3d5d6a1 ("sched: Refactor task_struct to use numa_faults instead of numa_* pointers") modified the way tsk->numa_faults stats are accounted. However that commit never touched show_numa_stats() that is displayed in /proc/pid/sched and thus the numbers displayed in /proc/pid/sched don't match the actual numbers. Fix it by making sure that /proc/pid/sched reflects the task fault numbers. Also add group fault stats too. Also couple of more modifications are added here: 1. Format changes: - Previously we would list two entries per node, one for private and one for shared. Also the home node info was listed in each entry. - Now preferred node, total_faults and current node are displayed separately. - Now there is one entry per node, that lists private,shared task and group faults. 2. Unit changes: - p->numa_pages_migrated was getting reset after every read of /proc/pid/sched. It's more useful to have absolute numbers since differential migrations between two accesses can be more easily calculated. Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Acked-by: Rik van Riel <riel@redhat.com> Cc: Iulia Manda <iulia.manda21@gmail.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1435252903-1081-4-git-send-email-srikar@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-06-25 17:21:43 +00:00
}
#endif /* CONFIG_NUMA_BALANCING */
#endif /* CONFIG_SCHED_DEBUG */
__init void init_sched_fair_class(void)
{
#ifdef CONFIG_SMP
open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
nohz: Rename CONFIG_NO_HZ to CONFIG_NO_HZ_COMMON We are planning to convert the dynticks Kconfig options layout into a choice menu. The user must be able to easily pick any of the following implementations: constant periodic tick, idle dynticks, full dynticks. As this implies a mutual exclusion, the two dynticks implementions need to converge on the selection of a common Kconfig option in order to ease the sharing of a common infrastructure. It would thus seem pretty natural to reuse CONFIG_NO_HZ to that end. It already implements all the idle dynticks code and the full dynticks depends on all that code for now. So ideally the choice menu would propose CONFIG_NO_HZ_IDLE and CONFIG_NO_HZ_EXTENDED then both would select CONFIG_NO_HZ. On the other hand we want to stay backward compatible: if CONFIG_NO_HZ is set in an older config file, we want to enable CONFIG_NO_HZ_IDLE by default. But we can't afford both at the same time or we run into a circular dependency: 1) CONFIG_NO_HZ_IDLE and CONFIG_NO_HZ_EXTENDED both select CONFIG_NO_HZ 2) If CONFIG_NO_HZ is set, we default to CONFIG_NO_HZ_IDLE We might be able to support that from Kconfig/Kbuild but it may not be wise to introduce such a confusing behaviour. So to solve this, create a new CONFIG_NO_HZ_COMMON option which gathers the common code between idle and full dynticks (that common code for now is simply the idle dynticks code) and select it from their referring Kconfig. Then we'll later create CONFIG_NO_HZ_IDLE and map CONFIG_NO_HZ to it for backward compatibility. Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Chris Metcalf <cmetcalf@tilera.com> Cc: Christoph Lameter <cl@linux.com> Cc: Geoff Levand <geoff@infradead.org> Cc: Gilad Ben Yossef <gilad@benyossef.com> Cc: Hakan Akkan <hakanakkan@gmail.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Kevin Hilman <khilman@linaro.org> Cc: Li Zhong <zhong@linux.vnet.ibm.com> Cc: Namhyung Kim <namhyung.kim@lge.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de>
2011-08-10 21:21:01 +00:00
#ifdef CONFIG_NO_HZ_COMMON
nohz.next_balance = jiffies;
nohz.next_blocked = jiffies;
zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
#endif
#endif /* SMP */
}
/*
* Helper functions to facilitate extracting info from tracepoints.
*/
const struct sched_avg *sched_trace_cfs_rq_avg(struct cfs_rq *cfs_rq)
{
#ifdef CONFIG_SMP
return cfs_rq ? &cfs_rq->avg : NULL;
#else
return NULL;
#endif
}
EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_avg);
char *sched_trace_cfs_rq_path(struct cfs_rq *cfs_rq, char *str, int len)
{
if (!cfs_rq) {
if (str)
strlcpy(str, "(null)", len);
else
return NULL;
}
cfs_rq_tg_path(cfs_rq, str, len);
return str;
}
EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_path);
int sched_trace_cfs_rq_cpu(struct cfs_rq *cfs_rq)
{
return cfs_rq ? cpu_of(rq_of(cfs_rq)) : -1;
}
EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_cpu);
const struct sched_avg *sched_trace_rq_avg_rt(struct rq *rq)
{
#ifdef CONFIG_SMP
return rq ? &rq->avg_rt : NULL;
#else
return NULL;
#endif
}
EXPORT_SYMBOL_GPL(sched_trace_rq_avg_rt);
const struct sched_avg *sched_trace_rq_avg_dl(struct rq *rq)
{
#ifdef CONFIG_SMP
return rq ? &rq->avg_dl : NULL;
#else
return NULL;
#endif
}
EXPORT_SYMBOL_GPL(sched_trace_rq_avg_dl);
const struct sched_avg *sched_trace_rq_avg_irq(struct rq *rq)
{
#if defined(CONFIG_SMP) && defined(CONFIG_HAVE_SCHED_AVG_IRQ)
return rq ? &rq->avg_irq : NULL;
#else
return NULL;
#endif
}
EXPORT_SYMBOL_GPL(sched_trace_rq_avg_irq);
int sched_trace_rq_cpu(struct rq *rq)
{
return rq ? cpu_of(rq) : -1;
}
EXPORT_SYMBOL_GPL(sched_trace_rq_cpu);
const struct cpumask *sched_trace_rd_span(struct root_domain *rd)
{
#ifdef CONFIG_SMP
return rd ? rd->span : NULL;
#else
return NULL;
#endif
}
EXPORT_SYMBOL_GPL(sched_trace_rd_span);