Re: [PATCH] sched/pelt: Fix task util_est update filtering

[Vincent Donnefort <vincent.donnefort@arm.com>]

On Thu, Feb 25, 2021 at 04:26:50PM +0100, Vincent Guittot wrote:
> On Mon, 22 Feb 2021 at 10:24, Vincent Donnefort
> <vincent.donnefort@arm.com> wrote:
> >
> > On Fri, Feb 19, 2021 at 11:48:28AM +0100, Vincent Guittot wrote:
> > > On Tue, 16 Feb 2021 at 17:39, <vincent.donnefort@arm.com> wrote:
> > > >
> > > > From: Vincent Donnefort <vincent.donnefort@arm.com>
> > > >
> > > > Being called for each dequeue, util_est reduces the number of its updates
> > > > by filtering out when the EWMA signal is different from the task util_avg
> > > > by less than 1%. It is a problem for a sudden util_avg ramp-up. Due to the
> > > > decay from a previous high util_avg, EWMA might now be close enough to
> > > > the new util_avg. No update would then happen while it would leave
> > > > ue.enqueued with an out-of-date value.
> > > >
> > > > Taking into consideration the two util_est members, EWMA and enqueued for
> > > > the filtering, ensures, for both, an up-to-date value.
> > > >
> > > > This is for now an issue only for the trace probe that might return the
> > > > stale value. Functional-wise, it isn't (yet) a problem, as the value is
> > >
> > > What do you mean by "it isn't (yet) a problem" ? How could this become
> > > a problem ?
> >
> > I wrote "yet" as nothing prevents anyone from using the ue.enqueued signal.
> 
> Hmm.. you are not supposed to use it outside the helper functions so
> this is irrelevant IMO which means that only the trace probe is
> impacted

I'll remove it.

> 
> >
> > >
> > > > always accessed through max(enqueued, ewma).
> > > >
> > >
> > > This adds more tests and or update of  struct avg.util_est. It would
> > > be good to have an idea of the perf impact. Especially because this
> > > only fixes a tracing problem
> >
> > I ran hackbench on the big cores of a SD845C board. After 100 iterations of
> > 100 loops runs, the geometric mean of the hackbench test is 0.1% lower
> > with this patch applied (2.0833s vs 2.0858s). The p-value, computed with
> > the ks_2samp [1] is 0.37. We can't conclude that the two distributions are
> > different. This patch, in this scenario seems completely harmless.
> 
> For such kind of change,  perf bench sched pipe is better to highlight
> any perf regression. I have done a quick test and i haven't seen
> noticeable difference

Thanks. I'll add your results to the commit message.

> 
> >
> > Shall I include those results in the commit message?
> >
> > [1] https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.ks_2samp.html
> >
> > >
> > >
> > > > This problem has been observed using LISA's UtilConvergence:test_means on
> > > > the sd845c board.
> > > >
> > > > Signed-off-by: Vincent Donnefort <vincent.donnefort@arm.com>
> > > >
> > > > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > > > index 794c2cb945f8..9008e0c42def 100644
> > > > --- a/kernel/sched/fair.c
> > > > +++ b/kernel/sched/fair.c
> > > > @@ -3941,24 +3941,27 @@ static inline void util_est_dequeue(struct cfs_rq *cfs_rq,
> > > >         trace_sched_util_est_cfs_tp(cfs_rq);
> > > >  }
> > > >
> > > > +#define UTIL_EST_MARGIN (SCHED_CAPACITY_SCALE / 100)
> > > > +
> > > >  /*
> > > > - * Check if a (signed) value is within a specified (unsigned) margin,
> > > > + * Check if a (signed) value is within the (unsigned) util_est margin,
> > > >   * based on the observation that:
> > > >   *
> > > >   *     abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
> > > >   *
> > > > - * NOTE: this only works when value + maring < INT_MAX.
> > > > + * NOTE: this only works when value + UTIL_EST_MARGIN < INT_MAX.
> > > >   */
> > > > -static inline bool within_margin(int value, int margin)
> > > > +static inline bool util_est_within_margin(int value)
> > > >  {
> > > > -       return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
> > > > +       return ((unsigned int)(value + UTIL_EST_MARGIN - 1) <
> > > > +               (2 * UTIL_EST_MARGIN - 1));
> > > >  }
> > > >
> > > >  static inline void util_est_update(struct cfs_rq *cfs_rq,
> > > >                                    struct task_struct *p,
> > > >                                    bool task_sleep)
> > > >  {
> > > > -       long last_ewma_diff;
> > > > +       long last_ewma_diff, last_enqueued_diff;
> > > >         struct util_est ue;
> > > >
> > > >         if (!sched_feat(UTIL_EST))
> > > > @@ -3979,6 +3982,8 @@ static inline void util_est_update(struct cfs_rq *cfs_rq,
> > > >         if (ue.enqueued & UTIL_AVG_UNCHANGED)
> > > >                 return;
> > > >
> > > > +       last_enqueued_diff = ue.enqueued;
> > > > +
> > > >         /*
> > > >          * Reset EWMA on utilization increases, the moving average is used only
> > > >          * to smooth utilization decreases.
> > > > @@ -3992,12 +3997,19 @@ static inline void util_est_update(struct cfs_rq *cfs_rq,
> > > >         }
> > > >
> > > >         /*
> > > > -        * Skip update of task's estimated utilization when its EWMA is
> > > > +        * Skip update of task's estimated utilization when its members are
> > > >          * already ~1% close to its last activation value.
> > > >          */
> > > >         last_ewma_diff = ue.enqueued - ue.ewma;
> > > > -       if (within_margin(last_ewma_diff, (SCHED_CAPACITY_SCALE / 100)))
> > > > +       last_enqueued_diff -= ue.enqueued;
> > > > +       if (util_est_within_margin(last_ewma_diff)) {
> > > > +               if (!util_est_within_margin(last_enqueued_diff)) {
> > > > +                       ue.ewma = ue.enqueued;
> 
> why do you set ewma directly with latest enqueued value ?

The idea was to align both ewma and enqueued, as the diff is < 1% anyway.

I'll remove that in v2.


> 
> > > > +                       goto done;
> > > > +               }
> > > > +
> > > >                 return;
> > > > +       }
> > > >
> > > >         /*
> > > >          * To avoid overestimation of actual task utilization, skip updates if
> > > > --
> > > > 2.25.1
> > > >