Message-ID: <20200317143053.GF10914@e105550-lin.cambridge.arm.com>
Date:   Tue, 17 Mar 2020 14:30:53 +0000
From:   Morten Rasmussen <morten.rasmussen@....com>
To:     Daniel Lezcano <daniel.lezcano@...aro.org>
Cc:     Vincent Guittot <vincent.guittot@...aro.org>,
        Peter Zijlstra <peterz@...radead.org>,
        Ingo Molnar <mingo@...hat.com>,
        Juri Lelli <juri.lelli@...hat.com>,
        Dietmar Eggemann <dietmar.eggemann@....com>,
        Steven Rostedt <rostedt@...dmis.org>,
        Ben Segall <bsegall@...gle.com>,
        linux-kernel <linux-kernel@...r.kernel.org>,
        Qais Yousef <qais.yousef@....com>,
        Valentin Schneider <valentin.schneider@....com>
Subject: Re: [PATCH V2] sched: fair: Use the earliest break even

On Tue, Mar 17, 2020 at 02:48:51PM +0100, Daniel Lezcano wrote:
> 
> Hi Morten,
> 
> On 17/03/2020 08:56, Morten Rasmussen wrote:
> > Hi Daniel,
> > 
> > First, I think letting the scheduler know about desired minimum idle
> > times is an interesting optimization if the overhead can be kept at a
> > minimum. I do have a few comments about the patch though.
> > 
> > On Thu, Mar 12, 2020 at 11:04:19AM +0100, Daniel Lezcano wrote:
> >> On 12/03/2020 09:36, Vincent Guittot wrote:
> >>> Hi Daniel,
> >>>
> >>> On Wed, 11 Mar 2020 at 21:28, Daniel Lezcano <daniel.lezcano@...aro.org> wrote:
> >>>>
> >>>> In the idle CPU selection process occuring in the slow path via the
> >>>> find_idlest_group_cpu() function, we pick up in priority an idle CPU
> >>>> with the shallowest idle state otherwise we fall back to the least
> >>>> loaded CPU.
> >>>
> >>> The idea makes sense but this path is only used by fork and exec so
> >>> I'm not sure about the real impact
> >>
> >> I agree the fork / exec path is called much less often than the wake
> >> path but it makes more sense for the decision.
> > 
> > Looking at the flow in find_idlest_cpu(), AFAICT,
> > find_idlest_group_cpu() is not actually making the final choice of CPU,
> > so going through a lot of trouble there looking at idle states is
> > pointless. Is there something I don't see?
> > 
> > We fellow sd->child until groups == CPUs which which means that
> > find_idlest_group() actually makes the final choice as the final group
> > passed to find_idlest_group_cpu() is single-CPU group. The flow has been
> > like that for years. Even before you added the initial idle-state
> > awareness.
> > 
> > I agree with Vincent, if this should really make a difference it should
> > include wake-ups existing tasks too. Although I'm aware it would be a
> > more invasive change. As said from the beginning, the idea is fine, but
> > the current implementation should not make any measurable difference?
> 
> I'm seeing the wake-ups path so sensitive, I'm not comfortable to do any
> changes in it. That is the reason why the patch only changes the slow path.

Right. I'm not against being cautious at all. It would be interesting to
evaluate how bad it really is. The extra time-stamping business cost is
the same, so it really down how much we dare to use the information in
the fast-path and change the CPU selection policy. And of course, how
much can be gained by the change.

> 
> >>>> 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.

> 
> > The patch doesn't affect the number of idle state
> > enter/exit cycles, so you spend the amount of energy on those
> > transitions. The main change is that idle time get spread out, so CPUs
> > are less likely to be in the process of entering an idle state when they
> > are asked to wake back up again.
> > 
> > Isn't it fair to say that we expect the total number of wake-ups remains
> > unchanged? Total busy and idle times across all CPUs should remain the
> > same too? Unless chosen idle-state is changed, which I don't think we
> > expect either, there should be no net effect on energy? The main benefit
> > is reduced wake-up latency I think.
> > 
> > Regarding chosen idle state, I'm wondering how this patch affects the
> > cpuidle governor's idle state selection. Could the spreading of wake-ups
> > trick governor to pick a shallower idle-state for some idle CPUs because
> > we actively spread wake-ups rather than consolidating them? Just a
> > thought.
> 
> May be I missed the point, why are we spreading the tasks?

Picking the CPU with the smallest break-even time-stamp means you pick
the CPU that has been idle longest in the shallowest idle-state. If you
periodically one-shot spawn tasks at a rate which is long enough that
the shallowest state is the same for several CPUs, you would end up
picking the least recently used CPU each time effectively spreading the
wake-ups across all the CPUs in the same state.

Thinking more about it, it might not be a real problem as if one of the
CPUs suddenly choose a shallower idle-state, it would become the target
all new tasks from that point onwards.

> We are taking the decision on the same sched domain, no?

I'm not sure I get the relation to the sched_domain?

> 
> >>>> At the selection process, we use the shallowest CPU but in addition we
> >>>> choose the one with the minimal break even deadline instead of relying
> >>>> on the idle_timestamp. When the CPU is idle, the timestamp has less
> >>>> meaning because the CPU could have wake up and sleep again several times
> >>>> without exiting the idle loop. In this case the break even deadline is
> >>>> more relevant as it increases the probability of choosing a CPU which
> >>>> reached its break even.
> > 
> > I guess you could improve the idle time stamping without adding the
> > break-even time, they don't have to go together?
> 
> Yes, we can add the idle start time when entering idle in the
> cpuidle_enter function which is different from the idle_timestamp which
> gives the idle task scheduling. I sent a RFC for that [1].
> 
> However, each time we would like to inspect the deadline, we will have
> to compute it, so IMO it makes more sense to pre-compute it when
> entering idle in addition to the idle start.
> 
> [1] https://lkml.org/lkml/2020/3/16/902

Yes, I saw that patch too. Seems to make sense :-)

> 
> >>>> Tested on:
> >>>>  - a synquacer 24 cores, 6 sched domains
> >>>>  - a hikey960 HMP 8 cores, 2 sched domains, with the EAS and energy probe
> >>>>
> >>>> sched/perf and messaging does not show a performance regression. Ran
> >>>> 50 times schbench, adrestia and forkbench.
> >>>>
> >>>> The tools described at https://lwn.net/Articles/724935/
> >>>>
> >>>>  --------------------------------------------------------------
> >>>> | Synquacer             | With break even | Without break even |
> >>>>  --------------------------------------------------------------
> >>>> | schbench *99.0th      |      14844.8    |         15017.6    |
> >>>> | adrestia / periodic   |        57.95    |              57    |
> >>>> | adrestia / single     |         49.3    |            55.4    |
> >>>>  --------------------------------------------------------------
> >>>
> >>> Have you got some figures or cpuidle statistics for the syncquacer ?
> >>
> >> No, and we just noticed the syncquacer has a bug in the firmware and
> >> does not actually go to the idle states.
> > 
> > I would also like some statistics to help understanding what actually
> > changes.
> > 
> > I did some measurements on TX2, which only has one idle-state. I don't
> > see the same trends as you do. adrestia single seems to be most affected
> > by the patch, but _increases_ with the break_even patch rather than
> > decrease. I don't trust adrestia too much though as the time resolution
> > is low on TX2.
> > 
> > TX2			tip		break_even
> > ----------------------------------------------------
> > adrestia / single	5.21		5.51
> > adrestia / periodic	5.75		5.67
> > schbench 99.0th		45465.6		45376.0
> > hackbench		27.9851		27.9775
> > 
> > Notes:
> > adrestia: Avg of 100 runs: adrestia -l 25000
> > schbench: Avg of 10 runs: schbench -m16 -t64
> > hackbench: Avg of 10 runs: hackbench -g 20 -T 256 -l 100000
> 
> Thanks for testing. Is that a Jetson TX2 from Nvidia? If that is the
> case, IIRC, it has some kind of switcher for the CPUs in the firmware, I
> don't know how that can interact with the testing.

Sorry, I should have been clearer. It is a ThunderX2. 2x 32-core (128
threads) = 256 HW threads.

Morten

Powered by blists - more mailing lists