Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1753110AbaFGU4R (ORCPT ); Sat, 7 Jun 2014 16:56:17 -0400 Received: from v094114.home.net.pl ([79.96.170.134]:56252 "HELO v094114.home.net.pl" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with SMTP id S1753003AbaFGU4P (ORCPT ); Sat, 7 Jun 2014 16:56:15 -0400 From: "Rafael J. Wysocki" To: "Srivatsa S. Bhat" Cc: viresh.kumar@linaro.org, pavel@ucw.cz, svaidy@linux.vnet.ibm.com, ego@linux.vnet.ibm.com, linux-pm@vger.kernel.org, linux-kernel@vger.kernel.org Subject: Re: [PATCH v3] cpufreq: governor: Be friendly towards latency-sensitive bursty workloads Date: Sat, 07 Jun 2014 23:13:35 +0200 Message-ID: <2855981.L5ysAMpFml@vostro.rjw.lan> User-Agent: KMail/4.11.5 (Linux/3.15.0-rc5+; KDE/4.11.5; x86_64; ; ) In-Reply-To: <20140607204114.25673.87194.stgit@srivatsabhat.in.ibm.com> References: <20140607204114.25673.87194.stgit@srivatsabhat.in.ibm.com> MIME-Version: 1.0 Content-Transfer-Encoding: 7Bit Content-Type: text/plain; charset="utf-8" Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On Sunday, June 08, 2014 02:11:43 AM Srivatsa S. Bhat wrote: > Cpufreq governors like the ondemand governor calculate the load on the CPU > periodically by employing deferrable timers. A deferrable timer won't fire > if the CPU is completely idle (and there are no other timers to be run), in > order to avoid unnecessary wakeups and thus save CPU power. > > However, the load calculation logic is agnostic to all this, and this can > lead to the problem described below. > > > Time (ms) CPU 1 > > 100 Task-A running > > 110 Governor's timer fires, finds load as 100% in the last > 10ms interval and increases the CPU frequency. > > 110.5 Task-A running > > 120 Governor's timer fires, finds load as 100% in the last > 10ms interval and increases the CPU frequency. > > 125 Task-A went to sleep. With nothing else to do, CPU 1 > went completely idle. > > 200 Task-A woke up and started running again. > > 200.5 Governor's deferred timer (which was originally programmed > to fire at time 130) fires now. It calculates load for the > time period 120 to 200.5, and finds the load is almost zero. > Hence it decreases the CPU frequency to the minimum. > > 210 Governor's timer fires, finds load as 100% in the last > 10ms interval and increases the CPU frequency. > > > So, after the workload woke up and started running, the frequency was suddenly > dropped to absolute minimum, and after that, there was an unnecessary delay of > 10ms (sampling period) to increase the CPU frequency back to a reasonable value. > And this pattern repeats for every wake-up-from-cpu-idle for that workload. > This can be quite undesirable for latency- or response-time sensitive bursty > workloads. So we need to fix the governor's logic to detect such wake-up-from- > cpu-idle scenarios and start the workload at a reasonably high CPU frequency. > > One extreme solution would be to fake a load of 100% in such scenarios. But > that might lead to undesirable side-effects such as frequency spikes (which > might also need voltage changes) especially if the previous frequency happened > to be very low. > > We just want to avoid the stupidity of dropping down the frequency to a minimum > and then enduring a needless (and long) delay before ramping it up back again. > So, let us simply carry forward the previous load - that is, let us just pretend > that the 'load' for the current time-window is the same as the load for the > previous window. That way, the frequency and voltage will continue to be set > to whatever values they were set at previously. This means that bursty workloads > will get a chance to influence the CPU frequency at which they wake up from > cpu-idle, based on their past execution history. Thus, they might be able to > avoid suffering from slow wakeups and long response-times. > > However, we should take care not to over-do this. For example, such a "copy > previous load" logic will benefit cases like this: (where # represents busy > and . represents idle) > > ##########.........#########.........###########...........##########........ > > but it will be detrimental in cases like the one shown below, because it will > retain the high frequency (copied from the previous interval) even in a mostly > idle system: > > ##########.........#.................#.....................#............... > > (i.e., the workload finished and the remaining tasks are such that their busy > periods are smaller than the sampling interval, which causes the timer to > always get deferred. So, this will make the copy-previous-load logic copy > the initial high load to subsequent idle periods over and over again, thus > keeping the frequency high unnecessarily). > > So, we modify this copy-previous-load logic such that it is used only once > upon every wakeup-from-idle. Thus if we have 2 consecutive idle periods, the > previous load won't get blindly copied over; cpufreq will freshly evaluate the > load in the second idle interval, thus ensuring that the system comes back to > its normal state. > > [ The right way to solve this whole problem is to teach the CPU frequency > governors to also track load on a per-task basis, not just a per-CPU basis, > and then use both the data sources intelligently to set the appropriate > frequency on the CPUs. But that involves redesigning the cpufreq subsystem, > so this patch should make the situation bearable until then. ] > > Experimental results: > +-------------------+ > > I ran a modified version of ebizzy (called 'sleeping-ebizzy') that sleeps in > between its execution such that its total utilization can be a user-defined > value, say 10% or 20% (higher the utilization specified, lesser the amount of > sleeps injected). This ebizzy was run with a single-thread, tied to CPU 8. > > Behavior observed with tracing (sample taken from 40% utilization runs): > ------------------------------------------------------------------------ > > Without patch: > ~~~~~~~~~~~~~~ > kworker/8:2-12137 416.335742: cpu_frequency: state=2061000 cpu_id=8 > kworker/8:2-12137 416.335744: sched_switch: prev_comm=kworker/8:2 ==> next_comm=ebizzy > <...>-40753 416.345741: sched_switch: prev_comm=ebizzy ==> next_comm=kworker/8:2 > kworker/8:2-12137 416.345744: cpu_frequency: state=4123000 cpu_id=8 > kworker/8:2-12137 416.345746: sched_switch: prev_comm=kworker/8:2 ==> next_comm=ebizzy > <...>-40753 416.355738: sched_switch: prev_comm=ebizzy ==> next_comm=kworker/8:2 > --------------------------------------------------------------------- > <...>-40753 416.402202: sched_switch: prev_comm=ebizzy ==> next_comm=swapper/8 > -0 416.502130: sched_switch: prev_comm=swapper/8 ==> next_comm=ebizzy > <...>-40753 416.505738: sched_switch: prev_comm=ebizzy ==> next_comm=kworker/8:2 > kworker/8:2-12137 416.505739: cpu_frequency: state=2061000 cpu_id=8 > kworker/8:2-12137 416.505741: sched_switch: prev_comm=kworker/8:2 ==> next_comm=ebizzy > <...>-40753 416.515739: sched_switch: prev_comm=ebizzy ==> next_comm=kworker/8:2 > kworker/8:2-12137 416.515742: cpu_frequency: state=4123000 cpu_id=8 > kworker/8:2-12137 416.515744: sched_switch: prev_comm=kworker/8:2 ==> next_comm=ebizzy > > Observation: Ebizzy went idle at 416.402202, and started running again at > 416.502130. But cpufreq noticed the long idle period, and dropped the frequency > at 416.505739, only to increase it back again at 416.515742, realizing that the > workload is in-fact CPU bound. Thus ebizzy needlessly ran at the lowest frequency > for almost 13 milliseconds (almost 1 full sample period), and this pattern > repeats on every sleep-wakeup. This could hurt latency-sensitive workloads quite > a lot. > > With patch: > ~~~~~~~~~~~ > > kworker/8:2-29802 464.832535: cpu_frequency: state=2061000 cpu_id=8 > --------------------------------------------------------------------- > kworker/8:2-29802 464.962538: sched_switch: prev_comm=kworker/8:2 ==> next_comm=ebizzy > <...>-40738 464.972533: sched_switch: prev_comm=ebizzy ==> next_comm=kworker/8:2 > kworker/8:2-29802 464.972536: cpu_frequency: state=4123000 cpu_id=8 > kworker/8:2-29802 464.972538: sched_switch: prev_comm=kworker/8:2 ==> next_comm=ebizzy > <...>-40738 464.982531: sched_switch: prev_comm=ebizzy ==> next_comm=kworker/8:2 > --------------------------------------------------------------------- > kworker/8:2-29802 465.022533: sched_switch: prev_comm=kworker/8:2 ==> next_comm=ebizzy > <...>-40738 465.032531: sched_switch: prev_comm=ebizzy ==> next_comm=kworker/8:2 > kworker/8:2-29802 465.032532: sched_switch: prev_comm=kworker/8:2 ==> next_comm=ebizzy > <...>-40738 465.035797: sched_switch: prev_comm=ebizzy ==> next_comm=swapper/8 > -0 465.240178: sched_switch: prev_comm=swapper/8 ==> next_comm=ebizzy > <...>-40738 465.242533: sched_switch: prev_comm=ebizzy ==> next_comm=kworker/8:2 > kworker/8:2-29802 465.242535: sched_switch: prev_comm=kworker/8:2 ==> next_comm=ebizzy > <...>-40738 465.252531: sched_switch: prev_comm=ebizzy ==> next_comm=kworker/8:2 > > Observation: Ebizzy went idle at 465.035797, and started running again at > 465.240178. Since ebizzy was the only real workload running on this CPU, > cpufreq retained the frequency at 4.1Ghz throughout the run of ebizzy, no > matter how many times ebizzy slept and woke-up in-between. Thus, ebizzy > got the 10ms worth of 4.1 Ghz benefit during every sleep-wakeup (as compared > to the run without the patch) and this boost gave a modest improvement in total > throughput, as shown below. > > Sleeping-ebizzy records-per-second: > ----------------------------------- > > Utilization Without patch With patch Difference (Absolute and % values) > 10% 274767 277046 + 2279 (+0.829%) > 20% 543429 553484 + 10055 (+1.850%) > 40% 1090744 1107959 + 17215 (+1.578%) > 60% 1634908 1662018 + 27110 (+1.658%) > > A rudimentary and somewhat approximately latency-sensitive workload such as > sleeping-ebizzy itself showed a consistent, noticeable performance improvement > with this patch. Hence, workloads that are truly latency-sensitive will benefit > quite a bit from this change. Moreover, this is an overall win-win since this > patch does not hurt power-savings at all (because, this patch does not reduce > the idle time or idle residency; and the high frequency of the CPU when it goes > to cpu-idle does not affect/hurt the power-savings of deep idle states). > > Signed-off-by: Srivatsa S. Bhat > Reviewed-by: Gautham R. Shenoy > Acked-by: Viresh Kumar > --- > > Changes in v3: > * Modified the "copy-previous-load" logic to copy only once, upon the first > wakeup from idle, to fix the flaw pointed out by Pavel Machek. > > * Fixed the 64 bit division issue reported by Fengguang Wu's build robot. Applied to bleeding-edge, thanks! -- I speak only for myself. Rafael J. Wysocki, Intel Open Source Technology Center. -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/