lkml.org 
[lkml]   [2017]   [Mar]   [31]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [RESEND PATCH 2/2] sched/fair: Optimize __update_sched_avg()
On Fri, Mar 31, 2017 at 03:55:40AM -0700, Paul Turner wrote:

> > I'm thinking that we can compute the middle segment, by taking the max
> > value and chopping off the ends, like:
> >
>
> >
> > p
> > c2 = 1024 \Sum y^n
> > n=1
> >
> > inf inf
> > = 1024 ( \Sum y^n - \Sum y^n - y^0 )
> > n=0 n=p
> >
>
> Very nice!
> Minor nit: Second sum needs to be from n=p+1

Correct.

> > +static u32 __accumulate_pelt_segments(u64 periods, u32 d1, u32 d3)
> > {
> > + u32 c1, c2, c3 = d3; /* y^0 == 1 */
> >
> > /*
> > * c1 = d1 y^(p+1)
> > */
> > + c1 = decay_load((u64)d1, periods);
> >
> > /*
> > + * p
> > + * c2 = 1024 \Sum y^n
> > + * n=1
> > *
> > + * inf inf
> > + * = 1024 ( \Sum y^n - \Sum y^n - y^0 )
> > + * n=0 n=p+1
> > */
> > + c2 = LOAD_AVG_MAX - decay_load(LOAD_AVG_MAX, periods) - 1024;
>
> decay_load(LOAD_AVG_MAX, periods + 1)

So here, @periods == p+1, see also c1. Yes, this is confusing [*].

In particular, I think the decay terms for c1 and this should be the
same. We cut off this tail end of the series to replace it with c1 after
all.

[*] hysterically p used to be off by 1, which is where the p+1 came
from, but now periods includes it. I was thinking of doing a patch
correcting all the comments to fully eradicate the whole +1 business.


> I computed all the values vs true value that the old/new computations
> result in, and it's very close. Absolutely it's approximately 2x off
> the previous computation, e.g. if the old value was -15 (relative to
> true value) than the new computation is -30.
>
> This is definitely more than good enough. If we want more precision,
> then the correction factor of:
> +clamp(periods, 0, 45)

Can you do a patch with coherent comment explaining where that
correction term comes from?

\
 
 \ /
  Last update: 2017-03-31 13:40    [W:0.137 / U:0.468 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site