Re: [RFC PATCH 0/1] sched/pelt: Change PELT halflife at runtime

From: Vincent Guittot <vincent.guittot@linaro.org>
To: Peter Zijlstra <peterz@infradead.org>
Cc: Dietmar Eggemann <dietmar.eggemann@arm.com>,
	Qais Yousef <qyousef@layalina.io>,
	Kajetan Puchalski <kajetan.puchalski@arm.com>,
	Jian-Min Liu <jian-min.liu@mediatek.com>,
	Ingo Molnar <mingo@kernel.org>,
	Morten Rasmussen <morten.rasmussen@arm.com>,
	Vincent Donnefort <vdonnefort@google.com>,
	Quentin Perret <qperret@google.com>,
	Patrick Bellasi <patrick.bellasi@matbug.net>,
	Abhijeet Dharmapurikar <adharmap@quicinc.com>,
	Qais Yousef <qais.yousef@arm.com>,
	linux-kernel@vger.kernel.org,
	Jonathan JMChen <jonathan.jmchen@mediatek.com>
Subject: Re: [RFC PATCH 0/1] sched/pelt: Change PELT halflife at runtime
Date: Mon, 20 Feb 2023 14:39:33 +0100	[thread overview]
Message-ID: <CAKfTPtAFFZaGSyjNL+9LfDBBR0b9TWw3DwFKJZXURgcf-_W6NA@mail.gmail.com> (raw)
In-Reply-To: <Y/NHvvcgOov6DoLc@hirez.programming.kicks-ass.net>

On Mon, 20 Feb 2023 at 11:13, Peter Zijlstra <peterz@infradead.org> wrote:
>
> On Thu, Feb 09, 2023 at 05:16:46PM +0100, Vincent Guittot wrote:
>
> > > The results is very similar to PELT halflife reduction. The advantage is
> > > that 'util_est_faster' is only activated selectively when the runtime of
> > > the current task in its current activation is long enough to create this
> > > CPU util boost.
> >
> > IIUC how util_est_faster works, it removes the waiting time when
> > sharing cpu time with other tasks. So as long as there is no (runnable
> > but not running time), the result is the same as current util_est.
>
> Uh.. it's double the speed, no? Even if there is no contention, the
> fake/in-situ pelt sum runs at double time and thus will ramp up faster
> than normal.

Ah yes. I haven't noticed it was (delta * 2) and not delta

>
> > util_est_faster makes a difference only when the task alternates
> > between runnable and running slices.
>
> UTIL_EST was supposed to help mitigate some of that, but yes. Also note
> that _FASTER sorta sucks here because it starts from 0 every time, if it
> were to start from the state saved by util_est_dequeue(), it would ramp
> up faster still.

Yes.

>
> Patch has a comment along those lines I think.
>
> > Have you considered using runnable_avg metrics in the increase of cpu
> > freq ? This takes into the runnable slice and not only the running
> > time and increase faster than util_avg when tasks compete for the same
> > CPU
>
> Interesting! Indeed, that's boosting the DVFS for contention. And as
> deggeman's reply shows, it seems to work well.
>
> I wonder if that one place where it regresses is exactly the case
> without contention.

Yes that might be the case indeed. I would expect uclamp_min to help
for ensuring a min frequency such scenario