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

[Vincent Guittot]

On 13 April 2017 at 15:39, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> On Tue, Apr 11, 2017 at 09:52:21AM +0200, Vincent Guittot wrote:
>
>> > Secondly, what's up with the util_sum < LOAD_AVG_MAX * 1000 thing?
>>
>> The lost idle time makes sense only if the task can also be "idle"
>> when running at max capacity. When util_sum reaches the
>> LOAD_AVG_MAX*SCHED_CAPACITY_SCALE value, all tasks are considered to
>> be the same as we can't make any difference between a task running
>> 400ms or a task running 400sec. It means that these tasks are "always
>> running" tasks even at max capacity. In this case, there is no lost
>> idle time as they always run and tracking and adding back the lost
>> idle time because we run at lower capacity doesn't make sense anymore
>> so we discard it.
>
> Right, this is the point we reached yesterday with the too low F. At
> that point you cannot know and we assuming u=1, F<1 -> u=1, F=1, which
> is a sensible assumption.
>
>> Then an always running task can have a util_sum that is less than the
>> max value because of the rounding (util_avg varies between
>> [1006..1023]), so I use LOAD_AVG_MAX*1000 instead of LOAD_AVG_MAX*1024
>
> 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.

>
> 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