Message-ID: <CAKfTPtA-F9RaYJ=mMS84KZfrXnvNru2j=tki+e2Vs9ndgow4NA@mail.gmail.com>
Date:   Thu, 29 Oct 2020 15:30:35 +0100
From:   Vincent Guittot <vincent.guittot@...aro.org>
To:     Valentin Schneider <valentin.schneider@....com>
Cc:     Ingo Molnar <mingo@...hat.com>,
        Peter Zijlstra <peterz@...radead.org>,
        Juri Lelli <juri.lelli@...hat.com>,
        Dietmar Eggemann <dietmar.eggemann@....com>,
        Steven Rostedt <rostedt@...dmis.org>,
        Ben Segall <bsegall@...gle.com>, Mel Gorman <mgorman@...e.de>,
        linux-kernel <linux-kernel@...r.kernel.org>,
        Morten Rasmussen <morten.rasmussen@....com>,
        Tao Zhou <ouwen210@...mail.com>
Subject: Re: [PATCH v2] sched/fair: prefer prev cpu in asymmetric wakeup path

On Thu, 29 Oct 2020 at 15:19, Vincent Guittot
<vincent.guittot@...aro.org> wrote:
>
> On Thu, 29 Oct 2020 at 12:16, Valentin Schneider
> <valentin.schneider@....com> wrote:
> >
> >
> > Hi Vincent,
> >
> > On 28/10/20 17:44, Vincent Guittot wrote:
> > > During fast wakeup path, scheduler always check whether local or prev cpus
> > > are good candidates for the task before looking for other cpus in the
> > > domain. With
> > >   commit b7a331615d25 ("sched/fair: Add asymmetric CPU capacity wakeup scan")
> > > the heterogenous system gains a dedicated path but doesn't try to reuse
> > > prev cpu whenever possible. If the previous cpu is idle and belong to the
> > > LLC domain, we should check it 1st before looking for another cpu because
> > > it stays one of the best candidate and this also stabilizes task placement
> > > on the system.
> > >
> > > This change aligns asymmetric path behavior with symmetric one and reduces
> > > cases where the task migrates across all cpus of the sd_asym_cpucapacity
> > > domains at wakeup.
> > >
> > > This change does not impact normal EAS mode but only the overloaded case or
> > > when EAS is not used.
> > >
> > > - On hikey960 with performance governor (EAS disable)
> > >
> > > ./perf bench sched pipe -T -l 50000
> > >              mainline           w/ patch
> > > # migrations   999364                  0
> > > ops/sec        149313(+/-0.28%)   182587(+/- 0.40) +22%
> > >
> > > - On hikey with performance governor
> > >
> > > ./perf bench sched pipe -T -l 50000
> > >              mainline           w/ patch
> > > # migrations        0                  0
> > > ops/sec         47721(+/-0.76%)    47899(+/- 0.56) +0.4%
> > >
> > > According to test on hikey, the patch doesn't impact symmetric system
> > > compared to current implementation (only tested on arm64)
> > >
> > > Also read the uclamped value of task's utilization at most twice instead
> > > instead each time we compare task's utilization with cpu's capacity.
> > >
> > > Fixes: b7a331615d25 ("sched/fair: Add asymmetric CPU capacity wakeup scan")
> > > Signed-off-by: Vincent Guittot <vincent.guittot@...aro.org>
> >
> > Other than the below, I quite like this!
> >
> > > ---
> > > Changes in v2:
> > > - merge asymmetric and symmetric path instead of duplicating tests on target,
> > >   prev and other special cases.
> > >
> > > - factorize call to uclamp_task_util(p) and use fits_capacity(). This could
> > >   explain part of the additionnal improvement compared to v1 (+22% instead of
> > >   +17% on v1).
> > >
> > > - Keep using LLC instead of asym domain for early check of target, prev and
> > >   recent_used_cpu to ensure cache sharing between the task. This doesn't
> > >   change anything for dynamiQ but will ensure same cache for legacy big.LITTLE
> > >   and also simply the changes.
> > >
> >
> > On legacy big.LITTLE systems, sd_asym_cpucapacity spans all CPUs, so we
> > would iterate over those in select_idle_capacity() anyway - the policy
> > we've been going for is that capacity fitness trumps cache use.
> >
> > This does require the system to have a decent interconnect, cache snooping
> > & co, but that is IMO a requirement of any sane asymmetric system.
> >
> > To put words into code, this is the kind of check I would see:
> >
> >   if (static_branch_unlikely(&sched_asym_cpucapacity))
> >         return fits_capacity(task_util, capacity_of(cpu));
> >   else
>
> You can't make the shortcut that prev will always belong to the domain
> so you have to check that prev belongs to the sd_asym_cpucapacity.
> Even if it's true with current mobile Soc, This code is generic core
> code and must handle any kind of funny topology than HW guys could
> imagine

We would have something like that

static inline bool cpus_share_domain(int this_cpu, int that_cpu)
{
    if (static_branch_unlikely(&sched_asym_cpucapacity))
        return per_cpu(sd_asym_cpucapacity, this_cpu) ==
per_cpu(sd_asym_cpucapacity, that_cpu);

    return cpus_share_cache(this_cpu, that_cpu);
}

>
> >         return cpus_share_cache(cpu, other);
> >
> > > - don't check capacity for the per-cpu kthread UC because the assumption is
> > >   that the wakee queued work for the per-cpu kthread that is now complete and
> > >   the task was already on this cpu.
> > >
> > > - On an asymmetric system where an exclusive cpuset defines a symmetric island,
> > >   task's load is synced and tested although it's not needed. But taking care of
> > >   this special case by testing if sd_asym_cpucapacity is not null impacts by
> > >   more than 4% the performance of default sched_asym_cpucapacity path.
> > >
> > > - The huge increase of the number of migration for hikey960 mainline comes from
> > >   teh fact that the ftrace buffer was overloaded by events in the tests done
> > >   with v1.

Powered by blists - more mailing lists