Re: [RFC PATCH v4 1/2] sched/fair: Introduce short duration task check

[Chen Yu]

Hi Peter,
On 2023-01-16 at 11:33:26 +0100, Peter Zijlstra wrote:
> On Thu, Jan 05, 2023 at 12:33:16PM +0100, Dietmar Eggemann wrote:
> > On 16/12/2022 07:11, Chen Yu wrote:
> > 
> > [...]
> > 
> > > @@ -5995,6 +6005,18 @@ enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
> > >  
> > >  static void set_next_buddy(struct sched_entity *se);
> > >  
> > > +static inline void dur_avg_update(struct task_struct *p, bool task_sleep)
> > > +{
> > > +	u64 dur;
> > > +
> > > +	if (!task_sleep)
> > > +		return;
> > > +
> > > +	dur = p->se.sum_exec_runtime - p->se.prev_sum_exec_runtime_vol;
> > > +	p->se.prev_sum_exec_runtime_vol = p->se.sum_exec_runtime;
> > 
> > Shouldn't se->prev_sum_exec_runtime_vol be set in enqueue_task_fair()
> > and not in dequeue_task_fair()->dur_avg_update()? Otherwise `dur` will
> > contain sleep time.
> > 
> > Like we do for se->prev_sum_exec_runtime in set_next_entity() but for
> > one `set_next_entity()-put_prev_entity()` run section.
> > 
> > AFAICS, you want to measure the exec_runtime sum over all run sections
> > between enqueue and dequeue.
> 
> You were thinking of the dynamic PELT window size thread? (which is what
> I had to think of when I looked at this).
> 
> I think we can still do that with this prev_sum_exec_runtime_vol (can't
> say I love the name though).
I agree that this name is not accurate, maybe prev_sleep_sum_exec_runtime?
I'm open to any other name for this : )

Currently I'm checking Prateek's data on Zen3 and Yicong's data on Arm64,
and their data suggested that: inhibiting the spreading of short wakee is not
always a good idea on a system with small LLC. Meanwhile, according to my
test on a system with large number of CPUs in 1 LLC, short duration wakee become
a trouble maker if spreading them on different CPUs, which could trigger unexpected
race condition. I'm thinking of taking nr_llc_cpu into consideration when defining
a short duration task, and do some experiment on this.

thanks,
Chenyu