Re: [PATCH v2 1/9] sched/uclamp: Fix relationship between uclamp and migration margin

From: Qais Yousef <qyousef@layalina.io>
To: Valentin Schneider <vschneid@redhat.com>
Cc: Qais Yousef <qais.yousef@arm.com>, Ingo Molnar <mingo@kernel.org>,
	"Peter Zijlstra (Intel)" <peterz@infradead.org>,
	Vincent Guittot <vincent.guittot@linaro.org>,
	Dietmar Eggemann <dietmar.eggemann@arm.com>,
	linux-kernel@vger.kernel.org, Xuewen Yan <xuewen.yan94@gmail.com>,
	Lukasz Luba <lukasz.luba@arm.com>, Wei Wang <wvw@google.com>,
	Jonathan JMChen <Jonathan.JMChen@mediatek.com>,
	Hank <han.lin@mediatek.com>
Subject: Re: [PATCH v2 1/9] sched/uclamp: Fix relationship between uclamp and migration margin
Date: Tue, 8 Nov 2022 11:33:51 +0000	[thread overview]
Message-ID: <20221108113351.ne4kobarl5ooqffp@airbuntu> (raw)
In-Reply-To: <xhsmhpmdymw1n.mognet@vschneid.remote.csb>

On 11/07/22 18:58, Valentin Schneider wrote:
> On 05/11/22 19:24, Qais Yousef wrote:
> > On 11/04/22 17:35, Valentin Schneider wrote:
> >
> >> > +	/*
> >> > +	 * We must use capacity_orig_of() for comparing against uclamp_min and
> >> > +	 * uclamp_max. We only care about capacity pressure (by using
> >> > +	 * capacity_of()) for comparing against the real util.
> >> > +	 *
> >> > +	 * If a task is boosted to 1024 for example, we don't want a tiny
> >> > +	 * pressure to skew the check whether it fits a CPU or not.
> >> > +	 *
> >> > +	 * Similarly if a task is capped to capacity_orig_of(little_cpu), it
> >> > +	 * should fit a little cpu even if there's some pressure.
> >> > +	 *
> >> > +	 * Only exception is for thermal pressure since it has a direct impact
> >> > +	 * on available OPP of the system.
> >> > +	 *
> >> > +	 * We honour it for uclamp_min only as a drop in performance level
> >> > +	 * could result in not getting the requested minimum performance level.
> >> > +	 *
> >>
> >> Why specifically care about OPPs here? Per our CPU capacity model, a task
> >> alone on a CPUx throttled to f=fmax/2 and a task coscheduled on a CPUy with
> >> RT/DL tasks and/or IRQs such that cpu_capacity(CPUy) = 50% are both getting
> >> (roughly) the same performance level.
> >
> > Depends how you define performance level. What you call performance level,
> > I think is better called bandwidth. Uclamp is a performance and not a bandwidth
> > hint.
> >
> > If a 10% task:
> >
> >       p->util_avg = 10% * 1024
> >
> > is requesting max performance level
> >
> >       p->uclamp_min = 1024
> >
> > This will translate to running at highest frequency and in case of big.LITTLE
> > system, the biggest CPU too.
> >
> > RT/DL pressure has no impact in the task being able to achieve this; that is
> > running at max frequency and biggest cpu.
> >
> > If the cpu has no bandwidth to fit this task, then our usual comparison of
> > util_avg with capacity_of() should fail as usual.
> >
> 
> Ok so we *do* have this with how the fitting criteria are combined (I
> didn't get that when I first scanned through the code); thanks for
> elaborating on that.

Oh yeah, this hasn't changed.

> 
> > In the example above, the RT/DL pressure has to be pretty high for the 10% task
> > not to fit from bandwidth point of view. Which has nothing to do with
> > uclamp_min.  Only thermal pressure which drops OPPs can actually affect the
> > uclamp_min hint/request.
> >
> > That is, when the task runs it will run at maximum frequency regardless of the
> > RT/DL pressure. The fact that the bandwidth of the CPU can be stolen has
> > nothing to do with uclamp_min hint.
> >
> >
> > Thanks!
> >
> > --
> > Qais Yousef
> 