Re: [PATCH] sched/pelt: sync util/runnable_sum with PELT window when propagating

[Vincent Guittot]

On Fri, 24 Apr 2020 at 14:07, Dietmar Eggemann <dietmar.eggemann@xxxxxxx> wrote:
>
> On 23/04/2020 18:17, Vincent Guittot wrote:
> > On Thu, 23 Apr 2020 at 16:30, Dietmar Eggemann <dietmar.eggemann@xxxxxxx> wrote:
> >>
> >> On 22/04/2020 17:14, Vincent Guittot wrote:
>
> [..]
>
> >>     gcfs --> tg --> gse
> >>      ________________|
> >>      |
> >>      V
> >>
> >>     cfs ---> tg (root)
> >>
> >>      |
> >>      V
> >>
> >>     rq
> >>
> >
> > child cfs_rq aka gcfs_rq
> >   |
> > gse: group entity that represents child cfs_rq in parent cfs_rq
> >   |
> >   v
> > parent cfs_rq aka cfs_rq
>
> OK, I see. Maybe it's clearer to refer to child cfs_rq as gcfs_rq in
> this context.
>
> [...]
>
> >>>       /* Set new sched_entity's utilization */
> >>>       se->avg.util_avg = gcfs_rq->avg.util_avg;
> >>> -     se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
> >>> +     se->avg.util_sum = se->avg.util_avg * divider;
> >>
> >> divider uses cfs_rq but we sync se->avg.util_avg with gcfs_rq here.
> >
> > we sync the util_avg of gse with the new util_avg of gcfs_rq but gse
> > is attached to cfs_rq and as a result we have to use cfs_rq's
> > period_contrib
>
> I agree.
>
> But the decay windows (avg.last_update_time, avg.period_contrib) of
> cfs_rq and gcfs_rq are not always aligned, I guess?
>
> I see they are not after the online_fair_sched_group() ->
> attach_entity_cfs_rq() but later the are in sync as well.

cfs_rq and gcfs_rq use 2 different clocks:
- cfs_rq_clock_pelt(cfs_rq)
- cfs_rq_clock_pelt(gcfs_rq)

And they can be different in some cases like cfs bandwidth

>
> I ran a couple of different rt-app taskgroup tests and try to
>
> BUG_ON(se->avg.period_contrib != gcfs_rq->avg.period_contrib);
> BUG_ON(se->avg.last_update_time != gcfs_rq->avg.last_update_time)
>
> in update_tg_cfs_util() but they didn't trigger so far.
>
> Both, cfs_rq and gcfs_rq are in sync in update_tg_cfs_util() before and
> during a task runs in gcfs_rq.
>
> Are there cases where this wouldn't necessary happen in
> update_tg_cfs_util(), maybe a more complicated testcase?
>
> >> But since avg.period_contrib of cfs_rq and gcfs_rq are the same this
> >> should work.
> >>
> >>>       /* Update parent cfs_rq utilization */
> >>>       add_positive(&cfs_rq->avg.util_avg, delta);
> >>> -     cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
> >>> +     cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * divider;
> >>
> >> Looks like that avg.last_update_time of se (group entity), it's gcfs_rq
> >> and cfs_rq is always the same in update_tg_cfs_[util\|runnable].
> >>
> >> So that means the PELT windows are aligned for cfs_rqs and group se's?
> >
> > We want to align util_avg with util_sum and period_contrib otherwise
> > we might have some unalignment. It's quite similarly to what is done
> > in attach_entity_load_avg()
>
> I agree.
>
> >> And if we want to enforce this for cfs_rq and task, we have
> >> sync_entity_load_avg().
> >
> > It's not a matter of syncing the last_update_time
>
> I agree, this is not what you want to achieve here.
> But syncing 'last_update_time' and 'period_contrib' is what I understand
> as aligning the decay window (like in attach_entity_load_avg()).