Received-SPF: pass (google.com: domain of linux-kernel+bounces-102758-linux.lists.archive=gmail.com@vger.kernel.org designates 139.178.88.99 as permitted sender) client-ip=139.178.88.99;
Precedence: bulk
MIME-Version: 1.0
References: <20240313085817.48892-1-zegao@tencent.com> <d43ed111-085b-432b-ad5b-433d5031fad1@arm.com>
 <CAD8CoPAC0ti2=mAbP5GMN6ZidZnPV-Antf7KrjwPtEnE7gYOdg@mail.gmail.com>
In-Reply-To: <CAD8CoPAC0ti2=mAbP5GMN6ZidZnPV-Antf7KrjwPtEnE7gYOdg@mail.gmail.com>
From: Ze Gao <zegao2021@gmail.com>
Date: Thu, 14 Mar 2024 11:49:15 +0800
Message-ID: <CAD8CoPCAU0OnE9R-usnyggRO5BZe3ey5=JOK8Waf3pvs7uQTtQ@mail.gmail.com>
Subject: Re: [PATCH] sched: Improve rq selection for a blocked task when its
 affinity changes
To: Pierre Gondois <pierre.gondois@arm.com>
Cc: linux-kernel@vger.kernel.org, Peter Zijlstra <peterz@infradead.org>, 
	Mel Gorman <mgorman@suse.de>, Juri Lelli <juri.lelli@redhat.com>, Ingo Molnar <mingo@redhat.com>, 
	Dietmar Eggemann <dietmar.eggemann@arm.com>, Daniel Bristot de Oliveira <bristot@redhat.com>, 
	Ben Segall <bsegall@google.com>, Vincent Guittot <vincent.guittot@linaro.org>, 
	Valentin Schneider <vschneid@redhat.com>, Steven Rostedt <rostedt@goodmis.org>, Ze Gao <zegao@tencent.com>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable

On Thu, Mar 14, 2024 at 10:59=E2=80=AFAM Ze Gao <zegao2021@gmail.com> wrote=
:
>
> On Wed, Mar 13, 2024 at 11:16=E2=80=AFPM Pierre Gondois <pierre.gondois@a=
rm.com> wrote:
> >
> > Hello Ze,
> >
> > I am running stress-ng with the following command:
> >    stress-ng -c 1 -l 10 &
> > and migrating the process with:
> >    taskset -pc [cpus] [pid]
> >
> > The thread seems to be migrated via:
> > sched_setaffinity
> >    \-__sched_setaffinity()
> >      \-__set_cpus_allowed_ptr()
> >        \-__set_cpus_allowed_ptr_locked()
> >          \- [1]
>
>
>
> > [1]
> > /*
> >   * Picking a ~random cpu helps in cases where we are changing affinity
> >   * for groups of tasks (ie. cpuset), so that load balancing is not
> >   * immediately required to distribute the tasks within their new mask.
> >   */
> > dest_cpu =3D cpumask_any_and_distribute(cpu_valid_mask, ctx->new_mask);
> >
> > So it seems the destination CPU chosen among the new CPU affinity mask =
is done
> > here, by picking a random CPU in the mask.
>
> IIUC, this is for running/queued/waking tasks instead of blocked tasks.
>
> Am I missing something obvious here?
>
> > Checking the cpus_ptr in select_idle_sibling() might be useful in other=
 cases,
> > but I think the experiment doesn't show that. Maybe a another small twe=
ak could
>
> The experiment is used to illustrate that the status quo does not do well
> but has to rely on select_fallback_rq() to choose a cpu for a woken task
> which turns out to be a bad choice since it's already monopolized by a
> cpu bound task, that is why a second migration happens with the help
> of the load balancer.

Btw, perf alone does not show obvious results here, you need some
other observability tools to make sure the migration is not initiated by
__set_cpus_allowed_ptr_locked (i.e., for running tasks). I achieve this
by directly adding some tracepoints to both select_fallback_rq() and
select_idle_sibling().

Regards,
        -- Ze

> Actually, we can reuse the same reasons for doing so as in
>
>     commit 46a87b3851f0("sched/core: Distribute tasks within affinity mas=
ks")
>
> > be done at [1] instead ?
>
> As for blocked tasks, check out what is commented on set_task_cpu() and
> select_task_rq(), since we never call set_task_cpu() on blocked tasks whi=
ch
> in turn, we have no way to change p->wake_cpu to dest_cpu being randomly
> chosen here, so when it's woken up, it still needs to go through the
> select_task_rq() process using the outdated p->wake_cpu.
>
>
> Thanks,
>         -- Ze
>
> > Regards,
> > Pierre
> >
> > On 3/13/24 09:58, Ze Gao wrote:
> > > We observered select_idle_sibling() is likely to return the *target* =
cpu
> > > early which is likely to be the previous cpu this task is running on =
even
> > > when it's actually not within the affinity list newly set, from where=
 after
> > > we can only rely on select_fallback_rq() to choose one for us at its =
will
> > > (the first valid mostly for now).
> > >
> > > However, the one chosen by select_fallback_rq() is highly likely not =
a
> > > good enough candidate, sometimes it has to rely on load balancer to k=
ick
> > > in to place itself to a better cpu, which adds one or more unnecessar=
y
> > > migrations in no doubt. For example, this is what I get when I move t=
ask
> > > 3964 to cpu 23-24 where cpu 23 has a cpu bound work pinned already:
> > >
> > >          swapper       0 [013]   959.791829: sched:sched_migrate_task=
: comm=3Dstress-ng-cpu pid=3D3964 prio=3D120 orig_cpu=3D13 dest_cpu=3D23
> > > kworker/24:2-mm    1014 [024]   959.806148: sched:sched_migrate_task:=
 comm=3Dstress-ng-cpu pid=3D3964 prio=3D120 orig_cpu=3D23 dest_cpu=3D24
> > >
> > > The thing is we can actually do better if we do checks early and take=
 more
> > > advantages of the *target* in select_idle_sibling(). That is, we cont=
inue
> > > the idle cpu selection if *target* fails the test of cpumask_test_cpu=
(
> > > *target*, p->cpus_ptr). By doing so, we are likely to pick a good can=
didate,
> > > especially when the newly allowed cpu set shares some cpu resources w=
ith
> > > *target*.
> > >
> > > And with this change, we clearly see the improvement when I move task=
 3964
> > > to cpu 25-26 where cpu 25 has a cpu bound work pinned already.
> > >
> > >          swapper       0 [027]  4249.204658: sched:sched_migrate_task=
: comm=3Dstress-ng-cpu pid=3D3964 prio=3D120 orig_cpu=3D27 dest_cpu=3D26
> > >
> > > Note we do the same check for *prev* in select_idle_sibling() as well=
.
> > >
> > > Signed-off-by: Ze Gao <zegao@tencent.com>
> > > ---
> > >   kernel/sched/fair.c | 13 ++++++++-----
> > >   1 file changed, 8 insertions(+), 5 deletions(-)
> > >
> > > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > > index 533547e3c90a..9ef6e74c6b2a 100644
> > > --- a/kernel/sched/fair.c
> > > +++ b/kernel/sched/fair.c
> > > @@ -7511,16 +7511,19 @@ static int select_idle_sibling(struct task_st=
ruct *p, int prev, int target)
> > >        */
> > >       lockdep_assert_irqs_disabled();
> > >
> > > -     if ((available_idle_cpu(target) || sched_idle_cpu(target)) &&
> > > -         asym_fits_cpu(task_util, util_min, util_max, target))
> > > +     if (cpumask_test_cpu(target, p->cpus_ptr) &&
> > > +             (available_idle_cpu(target) || sched_idle_cpu(target)) =
&&
> > > +             asym_fits_cpu(task_util, util_min, util_max, target))
> > >               return target;
> > >
> > >       /*
> > >        * If the previous CPU is cache affine and idle, don't be stupi=
d:
> > >        */
> > > -     if (prev !=3D target && cpus_share_cache(prev, target) &&
> > > -         (available_idle_cpu(prev) || sched_idle_cpu(prev)) &&
> > > -         asym_fits_cpu(task_util, util_min, util_max, prev)) {
> > > +     if (cpumask_test_cpu(prev, p->cpus_ptr) &&
> > > +             prev !=3D target &&
> > > +             cpus_share_cache(prev, target) &&
> > > +             (available_idle_cpu(prev) || sched_idle_cpu(prev)) &&
> > > +             asym_fits_cpu(task_util, util_min, util_max, prev)) {
> > >
> > >               if (!static_branch_unlikely(&sched_cluster_active) ||
> > >                   cpus_share_resources(prev, target))