Re: [PATCH v2 3/3] cgroup/rstat: Add run_delay accounting for cgroups

[Abel Wu]

Hi Michal,

On 1/27/25 10:10 PM, Michal Koutný Wrote:
> Hello.
>
> On Sat, Jan 25, 2025 at 01:25:12PM +0800, Abel Wu <wuyun.abel@xxxxxxxxxxxxx> wrote:
> > The per-task and per-cpu accounting have already been tracked by
> > t->sched_info.run_delay and rq->rq_sched_info.run_delay respectively.
> > Extends this to also include cgroups.
> >
> > The PSI indicator, "some" of cpu.pressure, loses the insight into how
> > severely that cgroup is stalled. Say 100 tasks or just 1 task that gets
> > stalled at a certain point will show no difference in "some" pressure.
> > IOW "some" is a flat value that not weighted by the severity (e.g. # of
> > tasks).
>
> IIUC below are three examples of when "some" tasks are waiting for CPU:
>
> a)
>    t1 |----|
>    t2 |xx--|
>
> b)
>    t1 |----|
>    t2 |x---|
>    t3 |-x--|
>
> c)
>    t1 |----|
>    t2 |xx--|
>    t3 |xx--|
>
> (- means runnable on CPU, x means runnable waiting on RQ)
>
> Which pair from a), b), c) is indistinguishable via PSI? (Or can you
> please add your illustrative example?)

PSI tracks stall times for each cpu, and

	tSOME[cpu] = time(nr_delayed_tasks[cpu] != 0)

which turns nr_delayed_tasks[cpu] into boolean value, hence loses
insight into how severely this task group is stalled on this cpu.

Thanks,
	Abel