Re: [PATCH 2/2] sched/fair: Scale wakeup granularity relative to nr_running

[Peter Zijlstra]

On Wed, Sep 22, 2021 at 08:22:43PM +0200, Vincent Guittot wrote:

> @@ -7161,6 +7162,13 @@ static void check_preempt_wakeup(struct rq *rq,
> struct task_struct *p, int wake_
>                 return;
> 
>         update_curr(cfs_rq_of(se));
> +       delta_exec = se->sum_exec_runtime - se->prev_sum_exec_runtime;
> +       /*
> +        * Ensure that current got a chance to move forward
> +        */
> +       if (delta_exec < sysctl_sched_min_granularity)
> +               return;
> +

I think we tried that at some point; IIRC the problem with this is that
if the interactive task fails to preempt, that preemption is lost. IOW
interactivity suffers.

Basically if you don't want wake-up preemptions, use SCHED_BATCH, then
those tasks will not preempt one another, but the SCHED_NORMAL tasks
will preempt them.