lkml.org 
[lkml]   [2021]   [Feb]   [8]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Date
SubjectRe: [PATCH 3/8] sched/fair: Tweak misfit-related capacity checks
On Fri, 5 Feb 2021 at 21:07, Valentin Schneider
<valentin.schneider@arm.com> wrote:
>
> On 05/02/21 18:17, Vincent Guittot wrote:
> > On Fri, 5 Feb 2021 at 18:00, Valentin Schneider
> >> >> @@ -8253,7 +8260,7 @@ check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
> >> >> static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd)
> >> >> {
> >> >> return rq->misfit_task_load &&
> >> >> - (rq->cpu_capacity_orig < rq->rd->max_cpu_capacity ||
> >> >> + (capacity_greater(rq->rd->max_cpu_capacity, rq->cpu_capacity_orig) ||
> >> >
> >> > Why do you add a margin here whereas there was no margin before ?
> >> >
> >>
> >> Comparing capacities without any sort of filter can lead to ping-ponging
> >> tasks around (capacity values very easily fluctuate by +/- 1, if not more).
> >
> > max_cpu_capacity reflects the max of the cpu_capacity_orig values
> > don't aim to change and can be considered as static values.
> > It would be better to fix this rounding problem (if any) in
> > topology_get_cpu_scale instead of computing a margin every time it's
> > used
> >
>
> That's embarrassing, I was convinced we had something updating
> rd->max_cpu_capacity with actual rq->capacity values... But as you point
> out that's absolutely not the case, it's all based on rq->capacity_orig,
> which completely invalidates patch 5/8.
>
> Welp.
>
> Perhaps I can still keep 5/8 with something like
>
> if (!rq->misfit_task_load)
> return false;
>
> do {
> if (capacity_greater(group->sgc->max_capacity, rq->cpu_capacity))
> return true;
>
> group = group->next;
> } while (group != sd->groups);

I don't catch what you want to achieve with this while loop compared
to the original condition which is :
trigger a load_balance :
- if there is CPU with higher original capacity
- or if the capacity of this cpu has significantly reduced because of
pressure and there is maybe others with more capacity even if it's one
with highest original capacity

>
> return false;
>
> This works somewhat well for big.LITTLE, but for DynamIQ systems under a
> single L3 this ends up iterating over all the CPUs :/

\
 
 \ /
  Last update: 2021-02-08 18:38    [W:0.194 / U:0.028 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site