Re: [RFC v2 2/2] sched/fair: introduce sched-idle balance

[Abel Wu]

On 4/14/22 8:08 AM, Josh Don Wrote:
> > > >    /*
> > > >     * Use locality-friendly rq->overloaded to cache the status of the rq
> > > >     * to minimize the heavy cost on LLC shared data.
> > > > @@ -7837,6 +7867,22 @@ int can_migrate_task(struct task_struct *p, struct lb_env *env)
> > > >           if (kthread_is_per_cpu(p))
> > > >                   return 0;
> > > >
> > > > +       if (unlikely(task_h_idle(p))) {
> > > > +               /*
> > > > +                * Disregard hierarchically idle tasks during sched-idle
> > > > +                * load balancing.
> > > > +                */
> > > > +               if (env->idle == CPU_SCHED_IDLE)
> > > > +                       return 0;
> > > > +       } else if (!static_branch_unlikely(&sched_asym_cpucapacity)) {
> > > > +               /*
> > > > +                * It's not gonna help if stacking non-idle tasks on one
> > > > +                * cpu while leaving some idle.
> > > > +                */
> > > > +               if (cfs_rq_busy(env->src_rq) && !need_pull_cfs_task(env->dst_rq))
> > > > +                       return 0;
> > >
> > > These checks don't involve the task at all, so this kind of check
> > > should be pushed into the more general load balance function. But, I'm
> > > not totally clear on the motivation here. If we have cpu A with 1
> > > non-idle task and 100 idle tasks, and cpu B with 1 non-idle task, we
> > > should definitely try to load balance some of the idle tasks from A to
> > > B. idle tasks _do_ get time to run (although little), and this can add
> > > up and cause antagonism to the non-idle task if there are a lot of
> > > idle threads.
> >
> > CPU_SCHED_IDLE means triggered by sched_idle_balance() in which pulls
> > a non-idle task for the unoccupied cpu from the overloaded ones, so
> > idle tasks are not the target and should be skipped.
> >
> > The second part is: if we have cpu A with 1 non-idle task and 100 idle
> > tasks, and B with >=1 non-idle task, we don't migrate the last non-idle
> > task on A to B.
>
> It could be possible that we do want to migrate the last non-idle task
> from A to B, if the weight sum of idle tasks on A is very high (easily
> possible with affinity restrictions). So I think we should leave
> regular load balance alone here if it really wants to move the
> non-idle task, and wrap this entire block in an if (env->idle ==
> CPU_SCHED_IDLE).

Makes sense. I will fix it in next version.

Thanks & BR,
Abel