mirror of
https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git
synced 2024-08-21 16:29:59 +00:00
dccp ccid-3: Preventing Oscillations
This implements [RFC 3448, 4.5], which performs congestion avoidance behaviour by reducing the transmit rate as the queueing delay (measured in terms of long-term RTT) increases. Oscillation can be turned on/off via a module option (do_osc_prev) and via sysfs (using mode 0644), the default is off. Overflow analysis: ------------------ * oscillation prevention is done after update_x(), so that t_ipi <= 64000; * hence the multiplication "t_ipi * sqrt(R_sample)" needs 64 bits; * done using u64 for sqrt_sample and explicit typecast of t_ipi; * the divisor, R_sqmean, is non-zero because oscillation prevention is first called when receiving the second feedback packet, and tfrc_scaled_rtt() > 0. A detailed discussion of the algorithm (with plots) is on http://www.erg.abdn.ac.uk/users/gerrit/dccp/notes/ccid3/sender_notes/oscillation_prevention/ The algorithm has negative side effects: * when allowing to decrease t_ipi (leads to a large RTT) and * when using it during slow-start; both uses are therefore disabled. Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
This commit is contained in:
parent
53ac9570c8
commit
a3cbdde8e9
3 changed files with 59 additions and 2 deletions
|
@ -49,6 +49,8 @@ static int ccid3_debug;
|
|||
/*
|
||||
* Transmitter Half-Connection Routines
|
||||
*/
|
||||
/* Oscillation Prevention/Reduction: recommended by rfc3448bis, on by default */
|
||||
static int do_osc_prev = true;
|
||||
|
||||
/*
|
||||
* Compute the initial sending rate X_init in the manner of RFC 3390:
|
||||
|
@ -296,6 +298,9 @@ static int ccid3_hc_tx_send_packet(struct sock *sk, struct sk_buff *skb)
|
|||
hctx->s = ccid3_hc_tx_measure_packet_size(sk, skb->len);
|
||||
ccid3_update_send_interval(hctx);
|
||||
|
||||
/* Seed value for Oscillation Prevention (sec. 4.5) */
|
||||
hctx->r_sqmean = tfrc_scaled_sqrt(hctx->rtt);
|
||||
|
||||
} else {
|
||||
delay = ktime_us_delta(hctx->t_nom, now);
|
||||
ccid3_pr_debug("delay=%ld\n", (long)delay);
|
||||
|
@ -400,6 +405,38 @@ static void ccid3_hc_tx_packet_recv(struct sock *sk, struct sk_buff *skb)
|
|||
hctx->s, hctx->p, hctx->x_calc,
|
||||
(unsigned)(hctx->x_recv >> 6),
|
||||
(unsigned)(hctx->x >> 6));
|
||||
/*
|
||||
* Oscillation Reduction (RFC 3448, 4.5) - modifying t_ipi according to
|
||||
* RTT changes, multiplying by X/X_inst = sqrt(R_sample)/R_sqmean. This
|
||||
* can be useful if few connections share a link, avoiding that buffer
|
||||
* fill levels (RTT) oscillate as a result of frequent adjustments to X.
|
||||
* A useful presentation with background information is in
|
||||
* Joerg Widmer, "Equation-Based Congestion Control",
|
||||
* MSc Thesis, University of Mannheim, Germany, 2000
|
||||
* (sec. 3.6.4), who calls this ISM ("Inter-packet Space Modulation").
|
||||
*/
|
||||
if (do_osc_prev) {
|
||||
r_sample = tfrc_scaled_sqrt(r_sample);
|
||||
/*
|
||||
* The modulation can work in both ways: increase/decrease t_ipi
|
||||
* according to long-term increases/decreases of the RTT. The
|
||||
* former is a useful measure, since it works against queue
|
||||
* build-up. The latter temporarily increases the sending rate,
|
||||
* so that buffers fill up more quickly. This in turn causes
|
||||
* the RTT to increase, so that either later reduction becomes
|
||||
* necessary or the RTT stays at a very high level. Decreasing
|
||||
* t_ipi is therefore not supported.
|
||||
* Furthermore, during the initial slow-start phase the RTT
|
||||
* naturally increases, where using the algorithm would cause
|
||||
* delays. Hence it is disabled during the initial slow-start.
|
||||
*/
|
||||
if (r_sample > hctx->r_sqmean && hctx->p > 0)
|
||||
hctx->t_ipi = div_u64((u64)hctx->t_ipi * (u64)r_sample,
|
||||
hctx->r_sqmean);
|
||||
hctx->t_ipi = min_t(u32, hctx->t_ipi, TFRC_T_MBI);
|
||||
/* update R_sqmean _after_ computing the modulation factor */
|
||||
hctx->r_sqmean = tfrc_ewma(hctx->r_sqmean, r_sample, 9);
|
||||
}
|
||||
|
||||
/* unschedule no feedback timer */
|
||||
sk_stop_timer(sk, &hctx->no_feedback_timer);
|
||||
|
@ -749,6 +786,9 @@ static struct ccid_operations ccid3 = {
|
|||
.ccid_hc_tx_getsockopt = ccid3_hc_tx_getsockopt,
|
||||
};
|
||||
|
||||
module_param(do_osc_prev, bool, 0644);
|
||||
MODULE_PARM_DESC(do_osc_prev, "Use Oscillation Prevention (RFC 3448, 4.5)");
|
||||
|
||||
#ifdef CONFIG_IP_DCCP_CCID3_DEBUG
|
||||
module_param(ccid3_debug, bool, 0644);
|
||||
MODULE_PARM_DESC(ccid3_debug, "Enable debug messages");
|
||||
|
|
|
@ -47,8 +47,8 @@
|
|||
/* Two seconds as per RFC 3448 4.2 */
|
||||
#define TFRC_INITIAL_TIMEOUT (2 * USEC_PER_SEC)
|
||||
|
||||
/* Parameter t_mbi from [RFC 3448, 4.3]: backoff interval in seconds */
|
||||
#define TFRC_T_MBI 64
|
||||
/* Maximum backoff interval t_mbi (RFC 3448, 4.3) */
|
||||
#define TFRC_T_MBI (64 * USEC_PER_SEC)
|
||||
|
||||
/*
|
||||
* The t_delta parameter (RFC 3448, 4.6): delays of less than %USEC_PER_MSEC are
|
||||
|
@ -76,6 +76,7 @@ enum ccid3_options {
|
|||
* @x_recv - Receive rate in 64 * bytes per second
|
||||
* @x_calc - Calculated rate in bytes per second
|
||||
* @rtt - Estimate of current round trip time in usecs
|
||||
* @r_sqmean - Estimate of long-term RTT (RFC 3448, 4.5)
|
||||
* @p - Current loss event rate (0-1) scaled by 1000000
|
||||
* @s - Packet size in bytes
|
||||
* @t_rto - Nofeedback Timer setting in usecs
|
||||
|
@ -94,6 +95,7 @@ struct ccid3_hc_tx_sock {
|
|||
u64 x_recv;
|
||||
u32 x_calc;
|
||||
u32 rtt;
|
||||
u16 r_sqmean;
|
||||
u32 p;
|
||||
u32 t_rto;
|
||||
u32 t_ipi;
|
||||
|
|
|
@ -47,6 +47,21 @@ static inline u32 scaled_div32(u64 a, u64 b)
|
|||
return result;
|
||||
}
|
||||
|
||||
/**
|
||||
* tfrc_scaled_sqrt - Compute scaled integer sqrt(x) for 0 < x < 2^22-1
|
||||
* Uses scaling to improve accuracy of the integer approximation of sqrt(). The
|
||||
* scaling factor of 2^10 limits the maximum @sample to 4e6; this is okay for
|
||||
* clamped RTT samples (dccp_sample_rtt).
|
||||
* Should best be used for expressions of type sqrt(x)/sqrt(y), since then the
|
||||
* scaling factor is neutralised. For this purpose, it avoids returning zero.
|
||||
*/
|
||||
static inline u16 tfrc_scaled_sqrt(const u32 sample)
|
||||
{
|
||||
const unsigned long non_zero_sample = sample ? : 1;
|
||||
|
||||
return int_sqrt(non_zero_sample << 10);
|
||||
}
|
||||
|
||||
/**
|
||||
* tfrc_ewma - Exponentially weighted moving average
|
||||
* @weight: Weight to be used as damping factor, in units of 1/10
|
||||
|
|
Loading…
Reference in a new issue