Re: [PATCH 1/2] x86,sched: Add support for frequency invariance

[Vincent Guittot]

On Tue, 17 Sep 2019 at 16:21, Giovanni Gherdovich <ggherdovich@xxxxxxx> wrote:
>
> Hello Srinivas,
>
> On Fri, 2019-09-13 at 15:52 -0700, Srinivas Pandruvada wrote:
> > On Mon, 2019-09-09 at 04:42 +0200, Giovanni Gherdovich wrote:
> >
> > ...
> >
> > > +
> > > +/*
> > > + * APERF/MPERF frequency ratio computation.
> > > + *
> > > + * The scheduler wants to do frequency invariant accounting and
> > > needs a <1
> > > + * ratio to account for the 'current' frequency, corresponding to
> > > + * freq_curr / freq_max.
> >
> > I thought this is no longer the restriction and Vincent did some work
> > to remove this restriction.
>
> If you're referring to the patch
>
>   23127296889f "sched/fair: Update scale invariance of PELT"
>
> merged in v5.2, I'm familiar with that and from my understanding you still
> want a <1 scaling factor. This is my recalling of the patch:
>
> Vincent was studying some synthetic traces and realized that util_avg reported
> by PELT didn't quite match the result you'd get computing the formula with pen
> and paper (theoretical value). To address this he changed where the scaling
> factor is applied in the PELT formula.
>
> At some point when accumulating the PELT sums, you'll have to measure the time
> 'delta' since you last updated PELT. What we have after Vincent's change is
> that this time length 'delta' gets itself scaled by the freq_curr/freq_max
> ratio:
>
>     delta = time since last PELT update
>     delta *= freq_percent
>
> In this way time goes at "wall clock speed" only when you're running at max
> capacitiy, and goes "slower" (from the PELT point of view) if we're running at
> a lower frequency. I don't think Vincent had in mind a faster-than-wall-clock
> PELT time (which you'd get w/ freq_percent>1).

Yes, I haven't really planned to have time going faster that wall
clock but I don't see any algorithm problem at least if that would be
the case.
There will be a reduced maximum delta update of clock pelt but that
will still be large enough

>
> Speaking of which, Srinivas, do you have any opinion and/or requirement about
> this? I confusely remember Peter Zijlstra saying (more than a year ago, now)
> that you would like an unclipped freq_curr/freq_max ratio, and may not be
> happy with this patch clipping it to 1 when freq_curr > 4_cores_turbo. If
> that's the case, could you elaborate on this?
> Ignore that if it doesn't make sense, I may be mis-remembering.
>
>
> Thanks,
> Giovanni