Message-ID: <CAJZ5v0g7sz6V328byKpZOcieDjU14E1=sRfO88XvB9DGa5ZnDA@mail.gmail.com>
Date:   Thu, 16 Mar 2017 00:32:10 +0100
From:   "Rafael J. Wysocki" <rafael@...nel.org>
To:     Patrick Bellasi <patrick.bellasi@....com>
Cc:     "Rafael J. Wysocki" <rjw@...ysocki.net>,
        Viresh Kumar <viresh.kumar@...aro.org>,
        Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
        Linux PM <linux-pm@...r.kernel.org>,
        Steven Rostedt <rostedt@...dmis.org>,
        Thomas Gleixner <tglx@...utronix.de>,
        Vincent Guittot <vincent.guittot@...aro.org>,
        John Stultz <john.stultz@...aro.org>,
        Juri Lelli <juri.lelli@....com>, Todd Kjos <tkjos@...roid.com>,
        Tim Murray <timmurray@...gle.com>,
        Andres Oportus <andresoportus@...gle.com>,
        Joel Fernandes <joelaf@...gle.com>,
        Morten Rasmussen <morten.rasmussen@....com>,
        Dietmar Eggemann <dietmar.eggemann@....com>,
        Chris Redpath <chris.redpath@....com>,
        Ingo Molnar <mingo@...hat.com>,
        Peter Zijlstra <peterz@...radead.org>,
        "Rafael J . Wysocki" <rafael.j.wysocki@...el.com>
Subject: Re: [PATCH 3/6] cpufreq: schedutil: ensure max frequency while
 running RT/DL tasks

On Wed, Mar 15, 2017 at 3:40 PM, Patrick Bellasi
<patrick.bellasi@....com> wrote:
> On 15-Mar 12:52, Rafael J. Wysocki wrote:
>> On Friday, March 03, 2017 12:38:30 PM Patrick Bellasi wrote:
>> > On 03-Mar 14:01, Viresh Kumar wrote:
>> > > On 02-03-17, 15:45, Patrick Bellasi wrote:
>> > > > diff --git a/kernel/sched/cpufreq_schedutil.c b/kernel/sched/cpufreq_schedutil.c
>> > > > @@ -293,15 +305,29 @@ static void sugov_update_shared(struct update_util_data *hook, u64 time,
>> > > >         if (curr == sg_policy->thread)
>> > > >                 goto done;
>> > > >
>> > > > +       /*
>> > > > +        * While RT/DL tasks are running we do not want FAIR tasks to
>> > > > +        * overwrite this CPU's flags, still we can update utilization and
>> > > > +        * frequency (if required/possible) to be fair with these tasks.
>> > > > +        */
>> > > > +       rt_mode = task_has_dl_policy(curr) ||
>> > > > +                 task_has_rt_policy(curr) ||
>> > > > +                 (flags & SCHED_CPUFREQ_RT_DL);
>> > > > +       if (rt_mode)
>> > > > +               sg_cpu->flags |= flags;
>> > > > +       else
>> > > > +               sg_cpu->flags = flags;
>> > >
>> > > This looks so hacked up :)
>> >
>> > It is... a bit... :)
>> >
>> > > Wouldn't it be better to let the scheduler tell us what all kind of tasks it has
>> > > in the rq of a CPU and pass a mask of flags?
>> >
>> > That would definitively report a more consistent view of what's going
>> > on on each CPU.
>> >
>> > > I think it wouldn't be difficult (or time consuming) for the
>> > > scheduler to know that, but I am not 100% sure.
>> >
>> > Main issue perhaps is that cpufreq_update_{util,this_cpu} are
>> > currently called by the scheduling classes codes and not from the core
>> > scheduler. However I agree that it should be possible to build up such
>> > information and make it available to the scheduling classes code.
>> >
>> > I'll have a look at that.
>> >
>> > > IOW, the flags field in cpufreq_update_util() will represent all tasks in the
>> > > rq, instead of just the task that is getting enqueued/dequeued..
>> > >
>> > > And obviously we need to get some utilization numbers for the RT and DL tasks
>> > > going forward, switching to max isn't going to work for ever :)
>> >
>> > Regarding this last point, there are WIP patches Juri is working on to
>> > feed DL demands to schedutil, his presentation at last ELC partially
>> > covers these developments:
>> >   https://www.youtube.com/watch?v=wzrcWNIneWY&index=37&list=PLbzoR-pLrL6pSlkQDW7RpnNLuxPq6WVUR
>> >
>> > Instead, RT tasks are currently covered by an rt_avg metric which we
>> > already know is not fitting for most purposes.
>> > It seems that the main goal is twofold: move people to DL whenever
>> > possible otherwise live with the go-to-max policy which is the only
>> > sensible solution to satisfy the RT's class main goal, i.e. latency
>> > reduction.
>> >
>> > Of course such a go-to-max policy for all RT tasks we already know
>> > that is going to destroy energy on many different mobile scenarios.
>> >
>> > As a possible mitigation for that, while still being compliant with
>> > the main RT's class goal, we recently posted the SchedTune v3
>> > proposal:
>> >   https://lkml.org/lkml/2017/2/28/355
>> >
>> > In that proposal, the simple usage of CGroups and the new capacity_max
>> > attribute of the (existing) CPU controller should allow to define what
>> > is the "max" value which is just enough to match the latency
>> > constraints of a mobile application without sacrificing too much
>> > energy.
>
> Given the following interesting question, let's add Thomas Gleixner to
> the discussion, which can be interested in sharing his viewpoint.
>
>> And who's going to figure out what "max" value is most suitable?  And how?
>
> That's usually up to the system profiling which is part of the
> platform optimizations and tunings.
> For example it's possible to run  experiments to measure the bandwidth
> and (completion) latency requirements from the RT workloads.
>
> It's something which people developing embedded/mobile systems are
> quite aware of.

Well, I was expecting an answer like this to be honest and let me say
that it is not too convincing from my perspective.

At least throwing embedded and mobile into one bag seems to be a
stretch, because the usage patterns of those two groups are quite
different, even though they may be similar from the hardware POV.

Mobile are mostly used as general-purpose computers nowadays (and I
guess we're essentially talking about anything running Android, not
just phones, aren't we?) with applications installed by users rather
than by system integrators, so I'm doubtful about the viability of the
"system integrators should take care of it" assumption in this
particular case.

> I'm also quite confident on saying that most of
> them can agree that just going to the max OPP, each and every time a
> RT task becomes RUNNABLE, it is something which is more likely to hurt
> than to give benefits.

It would be good to be able to estimate that likelihood somehow ...

> AFAIK the current policy (i.e. "go to max") has been adopted for the
> following main reasons, which I'm reporting with some observations.
>
>
> .:: Missing of a suitable utilization metric for RT tasks
>
>  There is actually a utilization signal (rq->rt_avg) but it has been
>  verified to be "too slow" for the practical usage of driving OPP
>  selection.
>  Other possibilities are perhaps under exploration but they are not
>  yet there.
>
>
> .:: Promote the migration from RT to DEADLINE
>
>  Which makes a lot of sens for many existing use-cases, starting from
>  Android as well. However, it's also true that we cannot (at least yet)
>  split the world in DEALINE vs FAIR.
>  There is still, and there will be, a fair amount of RT tasks which
>  just it makes sense to serve at best both from the performance as
>  well as the power/energy standpoint.
>
>
> .:: Because RT is all about "reducing latencies"
>
>  Running at the maximum OPP is certainly the best way to aim for the
>  minimum latencies but... RT is about doing things "in time", which
>  does not imply "as fast as possible".
>  There can be many different workloads where a lower OPP is just good
>  enough to meet the expected soft RT behaviors provided by the Linux
>  RT scheduler.

AFAICS, RT is about determinism and meeting deadlines is one aspect of that.

Essentially, the problem here is to know which OPP is sufficient to do
things "on time" and there simply is no good way to figure that out
for RT tasks, again, AFAICS.

And since we can't say which OPP is sufficient, we pretty much have no
choice but to use the top-most one.

> All that considered, the modifications proposed in this series,
> combined with other bits which are for discussion in this [1] other
> posting, can work together to provide a better and more tunable OPP
> selection policy for RT tasks.

OK, but "more tunable" need not imply "better".

Thanks,
Rafael

Powered by blists - more mailing lists