Re: [PATCH v2 1/3] sched/pelt: Don't sync hardly util_sum with uti_avg

[Vincent Guittot <vincent.guittot@linaro.org>]

On Fri, 7 Jan 2022 at 12:43, Dietmar Eggemann <dietmar.eggemann@arm.com> wrote:
>
> On 05/01/2022 14:57, Vincent Guittot wrote:
> > On Wed, 5 Jan 2022 at 14:15, Dietmar Eggemann <dietmar.eggemann@arm.com> wrote:
> >>
> >> On 04/01/2022 14:42, Vincent Guittot wrote:
> >>> On Tue, 4 Jan 2022 at 12:47, Dietmar Eggemann <dietmar.eggemann@arm.com> wrote:
> >>>>
> >>>> On 22/12/2021 10:38, Vincent Guittot wrote:
> >>
> >> [...]
> >>
> >>>> I still wonder whether the regression only comes from the changes in
> >>>> update_cfs_rq_load_avg(), introduced by 1c35b07e6d39.
> >>>> And could be fixed only by this part of the patch-set (A):
> >>>
> >>> I have been able to trigger the warning even with (A) though It took
> >>> much more time.
> >>> And I have been able to catch wrong situations  (with additional
> >>> traces) in the 3 places A, B and C
> >>
> >> OK. By wrong situation you mean '_sum < _avg * MIN_DIVIDER' ?
> >
> > not only.
> > also util_sum == 0 but util_avg !=0 in different places although these
>
> Ah OK, I saw this one as part of '_sum < _avg * MIN_DIVIDER'.
>
> > situation didn't trigger sched_warn because some other sync happened
> > before the periodic call of __update_blocked_fair
> > or if nr_running == 1 and  and task's util_avg/sum > cfs' util_avg/sum
> > just before removing the task
>
> I see.
>
> >>>> @@ -3677,15 +3706,22 @@ update_cfs_rq_load_avg(u64 now, struct cfs_rq
> >>>> *cfs_rq)
> >>>>
> >>>>     r = removed_load;
> >>>>     sub_positive(&sa->load_avg, r);
> >>>> -   sa->load_sum = sa->load_avg * divider;
> >>>> +   sub_positive(&sa->load_sum, r * divider);
> >>>> +   sa->load_sum = max_t(u32, sa->load_sum, sa->load_avg * MIN_DIVIDER);
> >>>>
> >>>>     r = removed_util;
> >>>>     sub_positive(&sa->util_avg, r);
> >>>> -   sa->util_sum = sa->util_avg * divider;
> >>>> +   sub_positive(&sa->util_sum, r * divider);
> >>>> +   sa->util_sum = max_t(u32, sa->util_sum, sa->util_avg * MIN_DIVIDER);
> >>>>
> >>>>     r = removed_runnable;
> >>>>     sub_positive(&sa->runnable_avg, r);
> >>>> -   sa->runnable_sum = sa->runnable_avg * divider;
> >>>> +   sub_positive(&sa->runnable_sum, r * divider);
> >>>> +   sa->runnable_sum = max_t(u32, sa->runnable_sum,
> >>>> +                                 sa->runnable_avg * MIN_DIVIDER);
> >>>>
> >>>> i.e. w/o changing update_tg_cfs_X() (and
> >>>> detach_entity_load_avg()/dequeue_load_avg()).
> >>>>
> >>>> update_load_avg()
> >>>>   update_cfs_rq_load_avg()    <---
> >>>>   propagate_entity_load_avg()
> >>>>     update_tg_cfs_X()         <---
> >>>>
> >>>>
> >>>> I didn't see the SCHED_WARN_ON() [cfs_rq_is_decayed()] when looping on
> >>>> hackbench in several different sched group levels on
> >>>> [Hikey620 (Arm64, 8 CPUs, SMP, 4 taskgroups: A/B C/D E/F G/H), >12h uptime].
> >>>
> >>> IIRC, it was with hikey960 with cgroup v1
> >>> As a side note, I never trigger the problem with dragonboard845 and cgroup v2
> >>
> >> OK, just started a test on hikey960 cgroupv1. Let's see if I can catch it.
>
> Still no sign of the issue (hikey960, cgroupv1,  4 taskgroups: A/B C/D
> E/F G/H > 45h uptime
>
> >>>>> @@ -3780,7 +3799,11 @@ static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *s
> >>>>>
> >>>>>       dequeue_load_avg(cfs_rq, se);
> >>>>>       sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
> >>>>> -     cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * divider;
> >>>>> +     sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
> >>>>> +     /* See update_tg_cfs_util() */
> >>>>> +     cfs_rq->avg.util_sum = max_t(u32, cfs_rq->avg.util_sum,
> >>>>> +                                       cfs_rq->avg.util_avg * MIN_DIVIDER);
> >>>>> +
> >>>>
> >>>> Maybe add a:
> >>>>
> >>>> Fixes: fcf6631f3736 ("sched/pelt: Ensure that *_sum is always synced
> >>>> with *_avg")
> >>>
> >>> I spent time thinking about adding fixes tag. There is no crash/warn
> >>> so far so should we propagate it back in LTS for better performance ?
> >>
> >> Not sure I understand. What do you mean by 'should we propagate it back
> >> in LTS'?
> >
> > Sorry I had any stables in mind and not only LTS.
> >
> > Some of the changes done in PELT signal propagation that replace
> > subtracting util_sum  by using util_avg * divider instead, are related
> > to other problems with sched group hierarchy and
> > throttling/unthrottling. I'm not 100% confident that using fixes tag
> > to backport this on stables doesn't need to backport more patches on
> > other areas in order to not resurrect old problems. So I wonder if
> > it's worth  backporting this on stables
>
> OK, I see. So only 1c35b07e6d39 (i.e. the util _sum/_avg change in
> update_cfs_rq_load_avg() (1)) caused the CPU frequency regression. That
> was the reason why I initially suggested to split the patch-set
> differently. But you said that you saw the issue also when (1) is fixed.

Ok, I think that we were not speaking about the same setup. I wrongly
read that you were saying that
sa->util_sum = max_t(u32, sa->util_sum, sa->util_avg * MIN_DIVIDER);
was only needed in update_cfs_rq_load_avg() but not in the other places.

But what you said is that we only need the below to fix the perf
regression raised by rick ?
     r = removed_util;
     sub_positive(&sa->util_avg, r);
 -   sa->util_sum = sa->util_avg * divider;
 +   sub_positive(&sa->util_sum, r * divider);
 +   sa->util_sum = max_t(u32, sa->util_sum, sa->util_avg * MIN_DIVIDER);

The WARN that I mentioned in my previous email was about not adding
the max_t in all 3 places. I rerun some test today and I triggered the
WARN after a detach without the max_t line.

I can probably isolate the code above in a dedicated patch for the
regression raised by Rick and we could consider adding a fixes tag; I
will run more tests with only this part during the weekend.
That being said, we need to stay consistent in all 3 places where we
move or propagate some *_avg. In particular, using  "sa->util_sum =
sa->util_avg * divider" has the drawback of filtering the contribution
not already accounted for in util_avg and the impact is much larger
than expected. It means that  although fixing only
update_cfs_rq_load_avg() seems enough for rick's case, some other
cases could be impacted by the 2 other places and we need to fixed
them now that we have a better view of the root cause