Re: [PATCH] sched/pelt: Add UTIL_AVG_UNCHANGED flag for last_enqueued_diff

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

On Thu, May 06, 2021 at 08:46:08PM +0800, Xuewen Yan wrote:
> Hi
> On Thu, May 6, 2021 at 8:28 PM Vincent Donnefort
> <vincent.donnefort@arm.com> wrote:
> >
> > On Thu, May 06, 2021 at 07:09:36PM +0800, Xuewen Yan wrote:
> > > From: Xuewen Yan <xuewen.yan@unisoc.com>
> > >
> > > The UTIL_AVG_UNCHANGED flag had been cleared when the task util changed.
> > > And the enqueued is equal to task_util with the flag, so it is better
> > > to add the UTIL_AVG_UNCHANGED flag for last_enqueued_diff.

Could we change the description here a bit? I don't think this is accurately
explaning the issue. Would probably be interesting to mention that by not
setting the flag, which is the LSB, we add +1 to the diff. This is therefore
reducing slightly UTIL_EST_MARGIN.

> > >
> > > Fixes: b89997aa88f0b sched/pelt: Fix task util_est update filtering
> > >
> > > Signed-off-by: Xuewen Yan <xuewen.yan@unisoc.com>
> > > ---
> > >  kernel/sched/fair.c | 2 +-
> > >  1 file changed, 1 insertion(+), 1 deletion(-)
> > >
> > > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > > index e5e457fa9dc8..94d77b4fa601 100644
> > > --- a/kernel/sched/fair.c
> > > +++ b/kernel/sched/fair.c
> > > @@ -3996,7 +3996,7 @@ static inline void util_est_update(struct cfs_rq *cfs_rq,
> > >       if (ue.enqueued & UTIL_AVG_UNCHANGED)
> > >               return;
> > >
> > > -     last_enqueued_diff = ue.enqueued;
> > > +     last_enqueued_diff = (ue.enqueued | UTIL_AVG_UNCHANGED);
> > >
> > >       /*
> > >        * Reset EWMA on utilization increases, the moving average is used only
> > > --
> > > 2.29.0
> > >
> >
> > Hi,
> >
> > We do indeed for the diff use the flag for the value updated and no flag for the
> > value before the update. However, last_enqueued_diff is only used for the margin
> > check which is an heuristic and is not an accurate value (~1%) and as we know
> The last_enqueued_diff is compared with the UTIL_EST_MARGIN which is
> "1024/100 = 10",
> and The LSB may cause ~10% error.

I meant ~1% being the original margin. With the bit set, we would use 0.87% instead
of 0.97%.

> > we already loose the LSB in util_est, I'm not sure this is really necessary.
> I'm also not very sure, maybe the calculation will be more rigorous
> with the flag?
> >
> > --
> > Vincent
> >