Re: [PATCH 1/4] sched/fair: Be less aggressive in calling cpufreq_update_util()

[Qais Yousef]

On 12/11/23 07:56, Lukasz Luba wrote:
> 
> 
> On 12/10/23 20:51, Qais Yousef wrote:
> > On 12/08/23 10:05, Lukasz Luba wrote:
> > > Hi Qais,
> > > 
> > > On 12/8/23 01:52, Qais Yousef wrote:
> > > 
> > > [snip]
> > > 
> > > > @@ -6704,14 +6677,6 @@ enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
> > > >    	 */
> > > >    	util_est_enqueue(&rq->cfs, p);
> > > > -	/*
> > > > -	 * If in_iowait is set, the code below may not trigger any cpufreq
> > > > -	 * utilization updates, so do it here explicitly with the IOWAIT flag
> > > > -	 * passed.
> > > > -	 */
> > > > -	if (p->in_iowait)
> > > > -		cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
> > > > -
> > > 
> > > Why this io wait boost is considered as the $subject says 'aggressive'
> > > calling?
> > 
> > This will trigger a frequency update along with the iowait boost. Did I miss
> > something?
> 
> Yes, it will change CPU freq and it was the main goal for this code
> path. We have tests which check how that works on different memory
> types.
> 
> Why do you want to remove it?

It seems you missed this hunk? I of course didn't remove it altogether if
that's what you mean :)

	@@ -6772,6 +6737,8 @@ enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
	 enqueue_throttle:
		assert_list_leaf_cfs_rq(rq);

	+       cpufreq_update_util(rq, p->in_iowait ? SCHED_CPUFREQ_IOWAIT : 0);
	+
		hrtick_update(rq);
	 }

> 
> Did you run some tests (e.g. fio, gallery, etc) to check if you still
> have a decent performance out some new ufs/nvme memories?

PCMark storage gives

before*: 29681
after: 30014

* no patches applied including remove-margins one


Cheers

--
Qais Yousef

> 
> Beata & Dietmar have presented at LPC2021 a proposal to have a per-task
> io boost, with a bit more controllable way of the trade off power vs.
> performance [1]. IMO the io wait boost could evolve, not simply die.
> 
> Regards,
> Lukasz
> 
> [1] https://lpc.events/event/11/contributions/1042/