Re: [PATCH 2/2] sched: adjust SCHED_IDLE interactions

[Josh Don <joshdon@google.com>]

On Mon, Aug 16, 2021 at 5:52 AM Peter Zijlstra <peterz@infradead.org> wrote:
>
> On Wed, Aug 11, 2021 at 03:31:49PM +0200, Vincent Guittot wrote:
> > On Fri, 30 Jul 2021 at 04:00, Josh Don <joshdon@google.com> wrote:
>
>
> > > @@ -4216,7 +4228,15 @@ place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
> > >                 if (sched_feat(GENTLE_FAIR_SLEEPERS))
> > >                         thresh >>= 1;
> > >
> > > -               vruntime -= thresh;
> > > +               /*
> > > +                * Don't give sleep credit to a SCHED_IDLE entity if we're
> > > +                * placing it onto a cfs_rq with non SCHED_IDLE entities.
> > > +                */
> > > +               if (!se_is_idle(se) ||
> > > +                   cfs_rq->h_nr_running == cfs_rq->idle_h_nr_running)
>
> I really dislike that second clause, either never do this for idle or
> always, but not sometimes when the planets are aligned just right.

Yep, switched this to always for idle entities.

>
> > Can't this condition above create unfairness between idle entities ?
> > idle thread 1 wake up while normal thread is running
> > normal thread thread sleeps immediately after
> > idle thread 2 wakes up just after and gets some credits compared to the 1st one.
>
> No. Strictly speaking cfs is unfair here. But it's a really tricky case.
>
> Consider a task that is running 50% competing against a task that's
> running 100%. What's fair in that situation, a 50/50 split, or a 25/75
> split? What if that 50% is 50% of a minute?
>
> What we do here is fudge the vruntime such that we end up with a 50/50
> split provided the period over which it blocks is less than a slice.
> After that it gradually converges to the 'expected' 25/75 split that
> results from strict runnable competition.
>
> By not letting idle tasks participate in this, we avoid idle tasks
> 'stealing' the !runnable time and they revert back to strict runnable
> competition only.

I like Vincent's suggestion to use a reduced threshold for idle
entities, that's been working pretty well. And it retains some
idle<->idle fairness when we have idle waking onto other idle threads.