Re: [PATCH 1/2] sched/fair: Record the short sleeping time of a task

From: Chen Yu
Date: Wed Sep 27 2023 - 02:18:03 EST


Hi Mathieu,

On 2023-09-26 at 06:29:40 -0400, Mathieu Desnoyers wrote:
> On 9/26/23 06:11, Chen Yu wrote:
> > During task wakeup, the wakee firstly checks if its previous
> > running CPU is idle. If yes, choose that CPU as its first
> > choice. However, in most cases, the wakee's previous CPU
> > could be chosen by someone else, which breaks the cache
> > locality.
> >
> > Proposes a mechanism to reserve the task's previous
> > CPU for a short while. In this reservation period, other
> > tasks are not allowed to pick that CPU until a timeout.
> > The reservation period is defined as the average short
> > sleep time of the task. To be more specific, it is the
> > time delta between this task being dequeued and enqueued.
> > Only the sleep time shorter than sysctl_sched_migration_cost
> > will be recorded. If the sleep time is longer than
> > sysctl_sched_migration_cost, give the reservation period
> > a penalty by shrinking it to half. In this way, the 'burst'
> > sleeping time of the task is honored, meanwhile, if that
> > task becomes a long-sleeper, the reservation time of that
> > task is shrunk to reduce the impact on task wakeup.
> >
> > Suggested-by: Mathieu Desnoyers <mathieu.desnoyers@xxxxxxxxxxxx>
> > Signed-off-by: Chen Yu <yu.c.chen@xxxxxxxxx>
> > ---
> > include/linux/sched.h | 3 +++
> > kernel/sched/fair.c | 21 +++++++++++++++++++++
> > 2 files changed, 24 insertions(+)
> >
> > diff --git a/include/linux/sched.h b/include/linux/sched.h
> > index dc37ae787e33..4a0ac0276384 100644
> > --- a/include/linux/sched.h
> > +++ b/include/linux/sched.h
> > @@ -561,6 +561,9 @@ struct sched_entity {
> > u64 vruntime;
> > s64 vlag;
> > u64 slice;
> > + u64 prev_dequeue_time;
> > + /* the reservation period of this task during wakeup */
> > + u64 sis_rsv_avg;
> > u64 nr_migrations;
> > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > index d0877878bcdb..297b9470829c 100644
> > --- a/kernel/sched/fair.c
> > +++ b/kernel/sched/fair.c
> > @@ -6456,6 +6456,24 @@ enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
> > struct sched_entity *se = &p->se;
> > int idle_h_nr_running = task_has_idle_policy(p);
> > int task_new = !(flags & ENQUEUE_WAKEUP);
> > + u64 last_dequeue = p->se.prev_dequeue_time;
> > + u64 now = sched_clock_cpu(task_cpu(p));
> > +
> > + /*
> > + * If the task is a short-sleepting task, there is no need
> > + * to migrate it to other CPUs. Estimate the average short sleeping
> > + * time of the wakee. This sleep time is used as a hint to reserve
> > + * the dequeued task's previous CPU for a short while. During this
> > + * reservation period, select_idle_cpu() prevents other wakees from
> > + * choosing this CPU. This could bring a better cache locality.
>
> "This could bring a better cache locality." could be rephrased as
> "This improves cache locality for short-sleeping tasks."
>

OK, will do in the next version.

> Please add my:
>
> Reviewed-by: Mathieu Desnoyers <mathieu.desnoyers@xxxxxxxxxxxx>
>

Thanks very much!

best,
Chenyu