Re: [RFC PATCH v3 0/6] sched/cpufreq: Make schedutil energy aware

From: Peter Zijlstra <peterz@infradead.org>
To: Douglas Raillard <douglas.raillard@arm.com>
Cc: linux-kernel@vger.kernel.org, linux-pm@vger.kernel.org,
	mingo@redhat.com, rjw@rjwysocki.net, viresh.kumar@linaro.org,
	juri.lelli@redhat.com, vincent.guittot@linaro.org,
	dietmar.eggemann@arm.com, qperret@qperret.net,
	patrick.bellasi@matbug.net, dh.han@samsung.com
Subject: Re: [RFC PATCH v3 0/6] sched/cpufreq: Make schedutil energy aware
Date: Thu, 17 Oct 2019 16:53:22 +0200	[thread overview]
Message-ID: <20191017145322.GK2311@hirez.programming.kicks-ass.net> (raw)
In-Reply-To: <7edb1b73-54e7-5729-db5d-6b3b1b616064@arm.com>

On Thu, Oct 17, 2019 at 03:23:04PM +0100, Douglas Raillard wrote:
> On 10/17/19 10:50 AM, Peter Zijlstra wrote:
> > Now, the thing is, we use map_util_freq() in more places, and should we
> > not reflect this increase in C for all of them? That is, why is this
> > patch changing get_next_freq() and not map_util_freq().
> > 
> > I don't think that question is answered in the Changelogs.
> > 
> > Exactly because it does change the energy consumption (it must) should
> > that not also be reflected in the EAS logic?
> 
> map_util_freq() is only used in schedutil and in EAS by compute_energy(), so
> I retarget the question at compute_energy(). It is supposed to compute
> the energy consumed by a given CPU if a given task was migrated on it.
> This implies some assumptions on util signals:
> 1) util(_est.enqueued) of the task is representative of the task's
>    duty cycle (the semantic of util is somehow blurry for aperiodic tasks
>    AFAIK).
> 2) util of the task is CPU-invariant

( we know this to be false, but do indeed make this assumption because
simplicity, taking IPC differences into account would just complicate
things more )

> 3) task util + target CPU util = new target CPU util for the
>    foreseeable future, i.e. the amount of future we can predict with
>    reasonable accuracy. Otherwise we would end up moving things around
>    all the time.
> 
> Since ramp boost is designed to be transient, integrating it
> (indirectly) in "target CPU util" will add some noise to the placement
> decision, potentially rendering it obsolete as soon as the boosting
> stops. Basing a costly decision on that does not sound like a good
> idea IMHO.

Well, we _hope_ recent past is a reasonable predictor for the near
future. We of course know it'll be complete crap every so often, but
hope that on average it is true more than false :-)

Anyway, the above seems like a sensible enough argument, and seems
worthy of being part of the Changelog. Also maybe a comment in schedutil
as to why there is a map_util_freq() modifier there.