LKML Archive on lore.kernel.org
help / color / mirror / Atom feed
From: Morten Rasmussen <morten.rasmussen@arm.com>
To: Daniel Lezcano <daniel.lezcano@linaro.org>
Cc: Vincent Guittot <vincent.guittot@linaro.org>,
	Peter Zijlstra <peterz@infradead.org>,
	Ingo Molnar <mingo@redhat.com>,
	Juri Lelli <juri.lelli@redhat.com>,
	Dietmar Eggemann <dietmar.eggemann@arm.com>,
	Steven Rostedt <rostedt@goodmis.org>,
	Ben Segall <bsegall@google.com>,
	linux-kernel <linux-kernel@vger.kernel.org>,
	Qais Yousef <qais.yousef@arm.com>,
	Valentin Schneider <valentin.schneider@arm.com>
Subject: Re: [PATCH V2] sched: fair: Use the earliest break even
Date: Wed, 18 Mar 2020 15:38:23 +0100	[thread overview]
Message-ID: <20200318143823.GC6103@e123083-lin> (raw)
In-Reply-To: <798e9bde-a207-3a0e-0f13-0e27d60fd286@linaro.org>

On Wed, Mar 18, 2020 at 11:17:49AM +0100, Daniel Lezcano wrote:
> On 18/03/2020 09:24, Morten Rasmussen wrote:
> > On Tue, Mar 17, 2020 at 06:07:43PM +0100, Daniel Lezcano wrote:
> >> On 17/03/2020 15:30, Morten Rasmussen wrote:
> >>> On Tue, Mar 17, 2020 at 02:48:51PM +0100, Daniel Lezcano wrote:
> >>>> On 17/03/2020 08:56, Morten Rasmussen wrote:
> >>>>> On Thu, Mar 12, 2020 at 11:04:19AM +0100, Daniel Lezcano wrote:
> >>>>>>>> In order to be more energy efficient but without impacting the
> >>>>>>>> performances, let's use another criteria: the break even deadline.
> >>>>>>>>
> >>>>>>>> At idle time, when we store the idle state the CPU is entering in, we
> >>>>>>>> compute the next deadline where the CPU could be woken up without
> >>>>>>>> spending more energy to sleep.
> >>>>>
> >>>>> I don't follow the argument that sleeping longer should improve energy
> >>>>> consumption. 
> >>>>
> >>>> May be it is not explained correctly.
> >>>>
> >>>> The patch is about selecting a CPU with the smallest break even deadline
> >>>> value. In a group of idle CPUs in the same idle state, we will pick the
> >>>> one with the smallest break even dead line which is the one with the
> >>>> highest probability it already reached its target residency.
> >>>>
> >>>> It is best effort.
> >>>
> >>> Indeed. I get what the patch does, I just don't see how the patch
> >>> improves energy efficiency.
> >>
> >> If the CPU is woken up before it reached the break even, the idle state
> >> cost in energy is greater than the energy it saved.
> >>
> >> Am I misunderstanding your point?
> > 
> > Considering just the waking then yes, it reaches energy break-even.
> > However, considering all the CPUs in the system, it just moves the idle
> > entry/exit energy cost to a different CPU, it doesn't go away.
> > 
> > Whether you have:
> > 
> >                |-BE-|
> >            ____    ____
> > CPU0:  ___/    \__/    \___
> > 
> > CPU1:  ____________________
> > 
> > Or:
> >                |-BE-|
> >            ____
> > CPU0:  ___/    \___________
> >                    ____
> > CPU1:  ___________/    \___
> > 
> > _
> >   = CPU busy = P_{busy}
> > _ = CPU idle = P_{idle}
> > / = CPU idle exit = P_{exit}
> > \ = CPU idle entry = P_{entry}
> > 
> > The sum of areas under the curves is the same, i.e. the total energy is
> > unchanged.
> 
> It is a counter-intuitive comment, now I get it, thanks for the
> clarification. It is a good point.

No problem.

> Taking into consideration the dynamic, in the case #1, the break even is
> not reached, the idle duration is smaller and that leads the governor to
> choose shallower idle states after and consequently CPU0 will be used in
> priority. We end up with CPU0 in a shallow state and CPU1 in a deep state.

Indeed. I was speculating earlier if the opposite could happen too. If
we extended the second case to form a repeating pattern, could we
prevent somehow prevent CPU1 from reaching a deeper state? Could we have
pattern that would keep both CPUs in shallow state where it would have
been more efficient to consolidate the wake-ups on CPU0 and let CPU1
enter deeper states?

> 
> With the case #2, we can have the CPUs in both deep state and the
> governor should be keeping choosing the same idle state.

Ideally yes. However it depends on the break-even times of the deeper
states and when the next wake-ups happen.

> I don't know what is more energy/perf efficient. IMO this is very
> workload dependent. The only way to check is to test. Hopefully I can
> find a platform for that.

Moving the wake-up shouldn't impact energy directly, although it have a
positive latency impact as you are more likely to avoid waking up CPUs
that haven't finished the idle entry sequence. However, changing the
wake-up pattern could have an indirect energy impact, positive or
negative. It isn't clear to me either what outcome to expect.

Morten

  reply	other threads:[~2020-03-18 14:38 UTC|newest]

Thread overview: 19+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2020-03-11 20:26 Daniel Lezcano
2020-03-12  8:36 ` Vincent Guittot
2020-03-12 10:04   ` Daniel Lezcano
2020-03-12 12:27     ` Vincent Guittot
2020-03-13 12:15       ` Daniel Lezcano
2020-03-13 13:15         ` Vincent Guittot
2020-03-13 13:17           ` Daniel Lezcano
2020-03-13 13:21             ` Vincent Guittot
2020-03-17  7:56     ` Morten Rasmussen
2020-03-17 13:48       ` Daniel Lezcano
2020-03-17 14:30         ` Morten Rasmussen
2020-03-17 17:07           ` Daniel Lezcano
2020-03-18  8:24             ` Morten Rasmussen
2020-03-18 10:17               ` Daniel Lezcano
2020-03-18 14:38                 ` Morten Rasmussen [this message]
2020-03-18  7:51           ` Vincent Guittot
2020-03-18  8:33             ` Morten Rasmussen
2020-03-17 10:56 ` Valentin Schneider
2020-03-17 13:59   ` Daniel Lezcano

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20200318143823.GC6103@e123083-lin \
    --to=morten.rasmussen@arm.com \
    --cc=bsegall@google.com \
    --cc=daniel.lezcano@linaro.org \
    --cc=dietmar.eggemann@arm.com \
    --cc=juri.lelli@redhat.com \
    --cc=linux-kernel@vger.kernel.org \
    --cc=mingo@redhat.com \
    --cc=peterz@infradead.org \
    --cc=qais.yousef@arm.com \
    --cc=rostedt@goodmis.org \
    --cc=valentin.schneider@arm.com \
    --cc=vincent.guittot@linaro.org \
    --subject='Re: [PATCH V2] sched: fair: Use the earliest break even' \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).