Re: [PATCH v3] sched/fair: unlink misfit task from cpu overutilized

[Qais Yousef]

On 01/16/23 14:56, Dietmar Eggemann wrote:
> On 16/01/2023 12:23, Vincent Guittot wrote:
> > On Mon, 16 Jan 2023 at 10:07, Dietmar Eggemann <dietmar.eggemann@xxxxxxx> wrote:
> >>
> >> On 13/01/2023 14:40, Vincent Guittot wrote:
> 
> [...]
> 
> >>> @@ -6132,6 +6135,7 @@ static inline bool cpu_overutilized(int cpu)
> >>>       unsigned long rq_util_min = uclamp_rq_get(cpu_rq(cpu), UCLAMP_MIN);
> >>>       unsigned long rq_util_max = uclamp_rq_get(cpu_rq(cpu), UCLAMP_MAX);
> >>>
> >>> +     /* Return true only if the utlization doesn't fit its capacity */
> >>
> >> s/utlization/utilization
> >> s/its/CPU ?
> >>
> >>>       return !util_fits_cpu(cpu_util_cfs(cpu), rq_util_min, rq_util_max, cpu);
> >>>  }
> >>
> >> cpu_overutilized() is the only place where we now only test for
> >> !util_fits_cpu(). The new comment says we only care about utilization
> >> not fitting CPU capacity.
> >>
> >> Does this mean the rq uclamp values are not important here and we could
> >> go back to use fits_capacity()?
> >>
> >> Not sure since util_fits_cpu() is still coded differently:
> > 
> > uclamp_min is not important but uclamp_max still cap the utilization
> 
> OK, makes sense.
> 
> I.e. we could pass in `rq_util_min = 0` to avoid fetching it
> unnecessary? In case `fits == 1` before the uclamp_min condition in
> util_fits_cpu() it doesn't matter if we switch to return `-1` when
> called from cpu_overutilized(). Detail though ...
> 
> [...]
> 
> >>> @@ -6940,12 +6945,28 @@ select_idle_capacity(struct task_struct *p, struct sched_domain *sd, int target)
> >>>
> >>>               if (!available_idle_cpu(cpu) && !sched_idle_cpu(cpu))
> >>>                       continue;
> >>> -             if (util_fits_cpu(task_util, util_min, util_max, cpu))
> >>> +
> >>> +             fits = util_fits_cpu(task_util, util_min, util_max, cpu);
> >>> +
> >>> +             /* This CPU fits with all capacity and performance requirements */
> >>
> >> In task_fits_cpu() `utilization and performance (better uclamp)
> >> requirements` term was used. I assume it's the same thing here?
> >>
> >>> +             if (fits > 0)
> >>>                       return cpu;
> >>> +             /*
> >>> +              * Only the min performance (i.e. uclamp_min) doesn't fit. Look
> >>> +              * for the CPU with highest performance capacity.
> >>                                             ^^^^^^^^^^^^^^^^^^^^
> >>
> >> Do we use a new CPU capacity value `performance capacity (1)` here?
> >>
> >> Which I guess is `capacity_orig_of(cpu) - thermal_load_avg(cpu_rq(cpu)`.
> >>
> >> I'm asking since util_fits_cpu() still uses: `capacity_orig_thermal (2)
> >> = capacity_orig - arch_scale_thermal_pressure()` when checking whether
> >> to return -1. Shouldn't (1) and (2) be the same?
> > 
> > I'm all in favor of both being capacity_orig_of(cpu) -
> > thermal_load_avg(cpu_rq(cpu) like the capacity inversion detection
> 
> I think we need a handy name for this new capacity value, which seems to
> be `capacity_orig - capacity reduced by thermal`. And we should either
> use `thermal_load_avg` or `thermal pressure` for the latter part. And
> then we should use this consistently in all these places:
> util_fits_cpu(), feec(), sic().

So we had reports from Xuewen that not using instantaneous pressure causes
problems.

Lukasz came up with this patch to help address the problem, but it's still
waiting discussions. I think we need to discuss this problem more there.

	https://lore.kernel.org/lkml/20220429091245.12423-1-lukasz.luba@xxxxxxx/

At the moment there's no good solution and either comes with its own set of
caveat(s). Being consistent is not an option now AFAICT? We need to improve
thermal_load_avg() response time first somehow.

For now, to make best decision - we look at instantaneous. But when falling
back - I think using the long term pressure signal makes more sense because we
wouldn't be doing this fallback path if everything works as expected thermal
wise.

At least, that's what I think is a good trade-of for now :)


Thanks!

--
Qais Yousef