LKML Archive on lore.kernel.org
 help / color / Atom feed
From: Vincent Guittot <vincent.guittot@linaro.org>
To: Mel Gorman <mgorman@suse.de>
Cc: Patrick Bellasi <patrick.bellasi@matbug.net>,
	Dietmar Eggemann <dietmar.eggemann@arm.com>,
	Peter Zijlstra <peterz@infradead.org>,
	Qais Yousef <qais.yousef@arm.com>, Ingo Molnar <mingo@redhat.com>,
	Randy Dunlap <rdunlap@infradead.org>,
	Jonathan Corbet <corbet@lwn.net>,
	Juri Lelli <juri.lelli@redhat.com>,
	Steven Rostedt <rostedt@goodmis.org>,
	Ben Segall <bsegall@google.com>,
	Luis Chamberlain <mcgrof@kernel.org>,
	Kees Cook <keescook@chromium.org>,
	Iurii Zaikin <yzaikin@google.com>,
	Quentin Perret <qperret@google.com>,
	Valentin Schneider <valentin.schneider@arm.com>,
	Pavan Kondeti <pkondeti@codeaurora.org>,
	linux-doc@vger.kernel.org,
	linux-kernel <linux-kernel@vger.kernel.org>,
	linux-fs <linux-fsdevel@vger.kernel.org>
Subject: Re: [PATCH 1/2] sched/uclamp: Add a new sysctl to control RT default boost value
Date: Wed, 3 Jun 2020 16:59:00 +0200
Message-ID: <CAKfTPtAvMvPk5Ea2kaxXE8GzQ+Nc_PS+EKB1jAa03iJwQORSqA@mail.gmail.com> (raw)
In-Reply-To: <20200603101022.GG3070@suse.de>

On Wed, 3 Jun 2020 at 12:10, Mel Gorman <mgorman@suse.de> wrote:
>
> On Wed, Jun 03, 2020 at 10:29:22AM +0200, Patrick Bellasi wrote:
> >
> > Hi Dietmar,
> > thanks for sharing these numbers.
> >
> > On Tue, Jun 02, 2020 at 18:46:00 +0200, Dietmar Eggemann <dietmar.eggemann@arm.com> wrote...
> >
> > [...]
> >
> > > I ran these tests on 'Ubuntu 18.04 Desktop' on Intel E5-2690 v2
> > > (2 sockets * 10 cores * 2 threads) with powersave governor as:
> > >
> > > $ numactl -N 0 ./run-mmtests.sh XXX
> >
> > Great setup, it's worth to rule out all possible noise source (freq
> > scaling, thermal throttling, NUMA scheduler, etc.).
>
> config-network-netperf-cross-socket will do the binding of the server
> and client to two CPUs that are on one socket. However, it does not take
> care to avoid HT siblings although that could be implemented. The same
> configuration should limit the CPU to C1. It does not change the governor
> but all that would take is adding "cpupower frequency-set -g performance"
> to the end of the configuration.
>
> > Wondering if disabling HT can also help here in reducing results "noise"?
> >
> > > w/ config-network-netperf-unbound.
> > >
> > > Running w/o 'numactl -N 0' gives slightly worse results.
> > >
> > > without-clamp      : CONFIG_UCLAMP_TASK is not set
> > > with-clamp         : CONFIG_UCLAMP_TASK=y,
> > >                      CONFIG_UCLAMP_TASK_GROUP is not set
> > > with-clamp-tskgrp  : CONFIG_UCLAMP_TASK=y,
> > >                      CONFIG_UCLAMP_TASK_GROUP=y
> > >
> > >
> > > netperf-udp
> > >                                 ./5.7.0-rc7            ./5.7.0-rc7            ./5.7.0-rc7
> > >                               without-clamp             with-clamp      with-clamp-tskgrp
> >
> > Can you please specify how to read the following scores? I give it a run
> > to my local netperf and it reports Throughput, thous I would expect the
> > higher the better... but... this seems something different.
> >
> > > Hmean     send-64         153.62 (   0.00%)      151.80 *  -1.19%*      155.60 *   1.28%*
> > > Hmean     send-128        306.77 (   0.00%)      306.27 *  -0.16%*      309.39 *   0.85%*
> > > Hmean     send-256        608.54 (   0.00%)      604.28 *  -0.70%*      613.42 *   0.80%*
> > > Hmean     send-1024      2395.80 (   0.00%)     2365.67 *  -1.26%*     2409.50 *   0.57%*
> > > Hmean     send-2048      4608.70 (   0.00%)     4544.02 *  -1.40%*     4665.96 *   1.24%*
> > > Hmean     send-3312      7223.97 (   0.00%)     7158.88 *  -0.90%*     7331.23 *   1.48%*
> > > Hmean     send-4096      8729.53 (   0.00%)     8598.78 *  -1.50%*     8860.47 *   1.50%*
> > > Hmean     send-8192     14961.77 (   0.00%)    14418.92 *  -3.63%*    14908.36 *  -0.36%*
> > > Hmean     send-16384    25799.50 (   0.00%)    25025.64 *  -3.00%*    25831.20 *   0.12%*
> >
> > If I read it as the lower the score the better, all the above results
> > tell us that with-clamp is even better, while with-clamp-tskgrp
> > is not that much worst.
> >
>
> The figures are throughput to taking the first line
>
> without-clamp           153.62
> with-clamp              151.80 (worse, so the percentage difference is negative)
> with-clamp-tskgrp       155.60 (better so the percentage different is positive)
>
> > The other way around (the higher the score the better) would look odd
> > since we definitively add in more code and complexity when uclamp has
> > the TG support enabled we would not expect better scores.
> >
>
> Netperf for small differences is very fickle as small differences in timing
> or code layout can make a difference. Boot-to-boot variance can also be
> an issue and bisection is generally unreliable. In this case, I relied on
> the perf annotation and differences in ftrace function_graph to determine
> that uclamp was introducing enough overhead to be considered a problem.

When I want to stress the fast path i usually use "perf bench sched pipe -T "
The tip/sched/core on my arm octo core gives the following results for
20 iterations of perf bench sched pipe -T -l 50000

all uclamp config disabled  50035.4(+/- 0.334%)
all uclamp config enabled  48749.8(+/- 0.339%)   -2.64%

It's quite easy to reproduce and probably easier to study the impact

>
> > > Hmean     recv-64         153.62 (   0.00%)      151.80 *  -1.19%*      155.60 *   1.28%*
> > > Hmean     recv-128        306.77 (   0.00%)      306.27 *  -0.16%*      309.39 *   0.85%*
> > > Hmean     recv-256        608.54 (   0.00%)      604.28 *  -0.70%*      613.42 *   0.80%*
> > > Hmean     recv-1024      2395.80 (   0.00%)     2365.67 *  -1.26%*     2409.50 *   0.57%*
> > > Hmean     recv-2048      4608.70 (   0.00%)     4544.02 *  -1.40%*     4665.95 *   1.24%*
> > > Hmean     recv-3312      7223.97 (   0.00%)     7158.88 *  -0.90%*     7331.23 *   1.48%*
> > > Hmean     recv-4096      8729.53 (   0.00%)     8598.78 *  -1.50%*     8860.47 *   1.50%*
> > > Hmean     recv-8192     14961.61 (   0.00%)    14418.88 *  -3.63%*    14908.30 *  -0.36%*
> > > Hmean     recv-16384    25799.39 (   0.00%)    25025.49 *  -3.00%*    25831.00 *   0.12%*
> > >
> > > netperf-tcp
> > >
> > > Hmean     64              818.65 (   0.00%)      812.98 *  -0.69%*      826.17 *   0.92%*
> > > Hmean     128            1569.55 (   0.00%)     1555.79 *  -0.88%*     1586.94 *   1.11%*
> > > Hmean     256            2952.86 (   0.00%)     2915.07 *  -1.28%*     2968.15 *   0.52%*
> > > Hmean     1024          10425.91 (   0.00%)    10296.68 *  -1.24%*    10418.38 *  -0.07%*
> > > Hmean     2048          17454.51 (   0.00%)    17369.57 *  -0.49%*    17419.24 *  -0.20%*
> > > Hmean     3312          22509.95 (   0.00%)    22229.69 *  -1.25%*    22373.32 *  -0.61%*
> > > Hmean     4096          25033.23 (   0.00%)    24859.59 *  -0.69%*    24912.50 *  -0.48%*
> > > Hmean     8192          32080.51 (   0.00%)    31744.51 *  -1.05%*    31800.45 *  -0.87%*
> > > Hmean     16384         36531.86 (   0.00%)    37064.68 *   1.46%*    37397.71 *   2.37%*
> > >
> > > The diffs are smaller than on openSUSE Leap 15.1 and some of the
> > > uclamp taskgroup results are better?
> > >
> > > With this test setup we now can play with the uclamp code in
> > > enqueue_task() and dequeue_task().
> > >
> > > ---
> > >
> > > W/ config-network-netperf-unbound (only netperf-udp and buffer size 64):
> > >
> > > $ perf diff 5.7.0-rc7_without-clamp/perf.data 5.7.0-rc7_with-clamp/perf.data | grep activate_task
> > >
> > > # Event 'cycles:ppp'
> > > #
> > > # Baseline  Delta Abs  Shared Object            Symbol
> > >
> > >      0.02%     +0.54%  [kernel.vmlinux]         [k] activate_task
> > >      0.02%     +0.38%  [kernel.vmlinux]         [k] deactivate_task
> > >
> > > $ perf diff 5.7.0-rc7_without-clamp/perf.data 5.7.0-rc7_with-clamp-tskgrp/perf.data | grep activate_task
> > >
> > >      0.02%     +0.35%  [kernel.vmlinux]         [k] activate_task
> > >      0.02%     +0.34%  [kernel.vmlinux]         [k] deactivate_task
> >
> > These data makes more sense to me, AFAIR we measured <1% impact in the
> > wakeup path using cycletest.
> >
>
> 1% doesn't sound like a lot but UDP_STREAM is an example of a load with
> a *lot* of wakeups so even though the impact on each individual wakeup
> is small, it builds up.
>
> > I would also suggest to always report the overheads for
> >   __update_load_avg_cfs_rq()
> > as a reference point. We use that code quite a lot in the wakeup path
> > and it's a good proxy for relative comparisons.
> >
> >
> > > I still see 20 out of 90 tests with the warning message that the
> > > desired confidence was not achieved though.
> >
> > Where the 90 comes from? From the above table we run 9 sizes for
> > {udp-send, udp-recv, tcp} and 3 kernels. Should not give us 81 results?
> >
> > Maybe the Warning are generated only when a test has to be repeated?
>
> The warning is issued when it could not get a reliable result within the
> iterations allowed.
>
> > > "
> > > !!! WARNING
> > > !!! Desired confidence was not achieved within the specified iterations.
> > > !!! This implies that there was variability in the test environment that
> > > !!! must be investigated before going further.
> > > !!! Confidence intervals: Throughput      : 6.727% <-- more than 5% !!!
> > > !!!                       Local CPU util  : 0.000%
> > > !!!                       Remote CPU util : 0.000%
> > > "
> > >
> > > mmtests seems to run netperf with the following '-I' and 'i' parameter
> > > hardcoded: 'netperf -t UDP_STREAM -i 3,3 -I 95,5'
> >
> > This means that we compute a score's (average +-2.5%) with a 95% confidence.
> >
> > Does not that means that every +-2.5% difference in the results
> > above should be considered in the noise?
> >
>
> Usually yes but the impact is small enough to be within noise but
> still detectable. Where we get hurt is when there are multiple problems
> introduced where each contribute overhead that is within the noise but when
> all added together there is a regression outside the noise. Uclamp is not
> special in this respect, it just happens to be the current focus.  We met
> this type of problem before with PSI that was resolved by e0c274472d5d
> ("psi: make disabling/enabling easier for vendor kernels").
>
> > I would say that it could be useful to run with more iterations
> > and, given the small numbers we are looking at (apparently we are
> > scared by a 1% overhead), we should better use a more aggressive CI.
> >
> > What about something like:
> >
> >    netperf -t UDP_STREAM -i 3,30 -I 99,1
> >
> > ?
> >
>
> You could but the runtime of netperf will be variable, it will not be
> guaranteed to give consistent results and it may mask the true variability
> of the workload. While we could debate which is a valid approach, I
> think it makes sense to minimise the overhead of uclamp when it's not
> configured even if that means putting it behind a static branch that is
> enabled via a command-line parameter or a Kconfig that specifies whether
> it's on or off by default.
>
> --
> Mel Gorman
> SUSE Labs

  reply index

Thread overview: 68+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2020-05-11 15:40 Qais Yousef
2020-05-11 15:40 ` [PATCH 2/2] Documentation/sysctl: Document uclamp sysctl knobs Qais Yousef
2020-05-11 17:18 ` [PATCH 1/2] sched/uclamp: Add a new sysctl to control RT default boost value Qais Yousef
2020-05-12  2:10 ` Pavan Kondeti
2020-05-12 11:46   ` Qais Yousef
2020-05-15 11:08 ` Patrick Bellasi
2020-05-18  8:31 ` Dietmar Eggemann
2020-05-18 16:49   ` Qais Yousef
2020-05-28 13:23 ` Peter Zijlstra
2020-05-28 15:58   ` Qais Yousef
2020-05-28 16:11     ` Peter Zijlstra
2020-05-28 16:51       ` Qais Yousef
2020-05-28 18:29         ` Peter Zijlstra
2020-05-28 19:08           ` Patrick Bellasi
2020-05-28 19:20           ` Dietmar Eggemann
2020-05-29  9:11           ` Qais Yousef
2020-05-29 10:21         ` Mel Gorman
2020-05-29 15:11           ` Qais Yousef
2020-05-29 16:02             ` Mel Gorman
2020-05-29 16:05               ` Qais Yousef
2020-05-29 10:08       ` Mel Gorman
2020-05-29 16:04         ` Qais Yousef
2020-05-29 16:57           ` Mel Gorman
2020-06-02 16:46         ` Dietmar Eggemann
2020-06-03  8:29           ` Patrick Bellasi
2020-06-03 10:10             ` Mel Gorman
2020-06-03 14:59               ` Vincent Guittot [this message]
2020-06-03 16:52                 ` Qais Yousef
2020-06-04 12:14                   ` Vincent Guittot
2020-06-05 10:45                     ` Qais Yousef
2020-06-09 15:29                       ` Vincent Guittot
2020-06-08 12:31                     ` Qais Yousef
2020-06-08 13:06                       ` Valentin Schneider
2020-06-08 14:44                       ` Steven Rostedt
2020-06-11 10:13                         ` Qais Yousef
2020-06-09 17:10                       ` Vincent Guittot
2020-06-11 10:24                         ` Qais Yousef
2020-06-11 12:01                           ` Vincent Guittot
2020-06-23 15:44                             ` Qais Yousef
2020-06-24  8:45                               ` Vincent Guittot
2020-06-05  7:55                   ` Patrick Bellasi
2020-06-05 11:32                     ` Qais Yousef
2020-06-05 13:27                       ` Patrick Bellasi
2020-06-03  9:40           ` Mel Gorman
2020-06-03 12:41             ` Qais Yousef
2020-06-04 13:40               ` Mel Gorman
2020-06-05 10:58                 ` Qais Yousef
2020-06-11 10:58                 ` Qais Yousef
2020-06-16 11:08                   ` Qais Yousef
2020-06-16 13:56                     ` Lukasz Luba
  -- strict thread matches above, loose matches on Subject: below --
2020-04-03 12:30 Qais Yousef
2020-04-14 18:21 ` Patrick Bellasi
2020-04-15  7:46   ` Patrick Bellasi
2020-04-20 15:04     ` Qais Yousef
2020-04-20  8:24   ` Dietmar Eggemann
2020-04-20 15:19     ` Qais Yousef
2020-04-21  0:52       ` Steven Rostedt
2020-04-21 11:16         ` Dietmar Eggemann
2020-04-21 11:23           ` Qais Yousef
2020-04-20 14:50   ` Qais Yousef
2020-04-15 10:11 ` Quentin Perret
2020-04-20 15:08   ` Qais Yousef
2020-04-20  8:29 ` Dietmar Eggemann
2020-04-20 15:13   ` Qais Yousef
2020-04-21 11:18     ` Dietmar Eggemann
2020-04-21 11:27       ` Qais Yousef
2020-04-22 10:59         ` Dietmar Eggemann
2020-04-22 13:13           ` Qais Yousef

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=CAKfTPtAvMvPk5Ea2kaxXE8GzQ+Nc_PS+EKB1jAa03iJwQORSqA@mail.gmail.com \
    --to=vincent.guittot@linaro.org \
    --cc=bsegall@google.com \
    --cc=corbet@lwn.net \
    --cc=dietmar.eggemann@arm.com \
    --cc=juri.lelli@redhat.com \
    --cc=keescook@chromium.org \
    --cc=linux-doc@vger.kernel.org \
    --cc=linux-fsdevel@vger.kernel.org \
    --cc=linux-kernel@vger.kernel.org \
    --cc=mcgrof@kernel.org \
    --cc=mgorman@suse.de \
    --cc=mingo@redhat.com \
    --cc=patrick.bellasi@matbug.net \
    --cc=peterz@infradead.org \
    --cc=pkondeti@codeaurora.org \
    --cc=qais.yousef@arm.com \
    --cc=qperret@google.com \
    --cc=rdunlap@infradead.org \
    --cc=rostedt@goodmis.org \
    --cc=valentin.schneider@arm.com \
    --cc=yzaikin@google.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link

LKML Archive on lore.kernel.org

Archives are clonable:
	git clone --mirror https://lore.kernel.org/lkml/0 lkml/git/0.git
	git clone --mirror https://lore.kernel.org/lkml/1 lkml/git/1.git
	git clone --mirror https://lore.kernel.org/lkml/2 lkml/git/2.git
	git clone --mirror https://lore.kernel.org/lkml/3 lkml/git/3.git
	git clone --mirror https://lore.kernel.org/lkml/4 lkml/git/4.git
	git clone --mirror https://lore.kernel.org/lkml/5 lkml/git/5.git
	git clone --mirror https://lore.kernel.org/lkml/6 lkml/git/6.git
	git clone --mirror https://lore.kernel.org/lkml/7 lkml/git/7.git
	git clone --mirror https://lore.kernel.org/lkml/8 lkml/git/8.git
	git clone --mirror https://lore.kernel.org/lkml/9 lkml/git/9.git

	# If you have public-inbox 1.1+ installed, you may
	# initialize and index your mirror using the following commands:
	public-inbox-init -V2 lkml lkml/ https://lore.kernel.org/lkml \
		linux-kernel@vger.kernel.org
	public-inbox-index lkml

Example config snippet for mirrors

Newsgroup available over NNTP:
	nntp://nntp.lore.kernel.org/org.kernel.vger.linux-kernel


AGPL code for this site: git clone https://public-inbox.org/public-inbox.git