Re: [PATCH v2] sched/fair: update scale invariance of PELT

[Peter Zijlstra]

On Thu, Apr 13, 2017 at 05:16:20PM +0200, Vincent Guittot wrote:
> On 13 April 2017 at 15:39, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:

> > OK, so the reason util_avg varies is because we compute it wrong. And I
> > think we can easily fix that once we pull out all the factors (which
> > would mean your patch and the pulling out of weight patch which still
> > needs to be finished).
> 
> That would be great to remove this unwanted variation.

So the problem with the _avg stuff is that we include the d3 segment,
that is the unfinished current window. Since we only re-compute the _avg
whenever we roll over, the intent already seems to be to only compute it
on completed windows.

But because 'complicated/expensive', its hard to not include d3 and thus
we get the wobble.

Once we compute pure running/runnable sums, without extra contrib
factors, we can simply subtract our d3 term from sum when doing the
division and change the divider to LOAD_AVG_MAX*y, getting the stable
_avg over all completed windows.

(you could do the same with factors, but then we get to do a bunch of
extra multiplications which aren't free).

> >
> > But you're comparing against util_sum here, that behaves slightly
> > different. I think you want 'util_sum >= 1024 * (LOAD_AVG_MAX - 1024)'
> > instead.
> 
> yes, the variation happens on the util_sum

Well, for util_sum its simple to ignore the current window, which is
what the suggested equation does (note that LOAD_AVG_MAX*y ==
LOAD_AVG_MAX-1024).