Re: [RFC PATCH 0/5] introduce sched-idle balancing

[Vincent Guittot <vincent.guittot@linaro.org>]

On Fri, 25 Feb 2022 at 11:46, Abel Wu <wuyun.abel@bytedance.com> wrote:
>
> On 2/25/22 4:29 PM, Vincent Guittot Wrote:
> > On Fri, 25 Feb 2022 at 07:46, Abel Wu <wuyun.abel@bytedance.com> wrote:
> >>
> >> Hi Peter,
> >>
> >> On 2/24/22 11:20 PM, Peter Zijlstra Wrote:
> >>> On Thu, Feb 17, 2022 at 11:43:56PM +0800, Abel Wu wrote:
> >>>> Current load balancing is mainly based on cpu capacity
> >>>> and task util, which makes sense in the POV of overall
> >>>> throughput. While there still might be some improvement
> >>>> can be done by reducing number of overloaded cfs rqs if
> >>>> sched-idle or idle rq exists.
> >>>
> >>> I'm much confused, there is an explicit new-idle balancer and a periodic
> >>> idle balancer already there.
> >>
> >> The two balancers are triggered on the rqs that have no tasks on them,
> >> and load_balance() seems don't show a preference for non-idle tasks so
> >
> > The load balance will happen at the idle pace if a sched_idle task is
> > running on the cpu so you will have an ILB on each cpu that run a
> > sched-idle task
>
> I'm afraid I don't quite follow you, since sched-idle balancer doesn't
> touch the ILB part, can you elaborate on this? Thanks.

I was referring to your sentence " The two balancers are triggered on
the rqs that have no tasks on them". When there is only sched-idle
tasks on a rq, the load_balance behave like the Idle Load Balance when
there is no task i.e. as often

>
> >
> >> there might be possibility that only idle tasks are pulled during load
> >> balance while overloaded rqs (rq->cfs.h_nr_running > 1) exist. As a
> >
> > There is a LB_MIN feature (disable by default) that filters task with
> > very low load ( < 16) which includes sched-idle task which has a max
> > load of 3

but we could easily change this like if !sched_idle_cpus then LB can
migrate only cfs tasks otherwise can migrate sched_idle task as well.
Instead of creating another side channel

>
> This feature might not that friendly to the situation that only
> sched-idle tasks are running in the system. And this situation
> can last more than half a day in our co-location systems in which
> the training/batch tasks are placed under idle groups or directly
> assigned to SCHED_IDLE.
>
> >
> >> result the normal tasks, mostly latency-critical ones in our case, on
> >> that overloaded rq still suffer waiting for each other. I observed this
> >> through perf sched.
> >>
> >> IOW the main difference from the POV of load_balance() between the
> >> latency-critical tasks and the idle ones is load.
> >>
> >> The sched-idle balancer is triggered on the sched-idle rqs periodically
> >> and the newly-idle ones. It does a 'fast' pull of non-idle tasks from
> >> the overloaded rqs to the sched-idle/idle ones to let the non-idle tasks
> >> make full use of cpu resources.
> >>
> >> The sched-idle balancer only focuses on non-idle tasks' performance, so
> >> it can introduce overall load imbalance, and that's why I put it before
> >> load_balance().
> >
> > According to the very low weight of a sched-idle task, I don't expect
> > much imbalance because of sched-idle tasks. But this also depends of
> > the number of sched-idle task.
> >
> >
> >>
> >> Best Regards,
> >>          Abel