Re: [PATCH 1/2] sched/fair: Record the average duration of a task

[Chen Yu <yu.c.chen@intel.com>]

On 2024-08-05 at 10:08:35 +0530, Madadi Vineeth Reddy wrote:
> Hi Chen Yu,
> 
> On 30/06/24 18:39, Chen Yu wrote:
> > Hi Mike,
> > 
> > Thanks for your time and giving the insights.
> > 
> > On 2024-06-26 at 06:21:43 +0200, Mike Galbraith wrote:
> >> On Tue, 2024-06-25 at 15:22 +0800, Chen Yu wrote:
> >>>
> >>> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> >>> index 0935f9d4bb7b..7399c4143528 100644
> >>> --- a/kernel/sched/core.c
> >>> +++ b/kernel/sched/core.c
> >>> @@ -4359,6 +4359,8 @@ static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
> >>>         p->migration_pending = NULL;
> >>>  #endif
> >>>         init_sched_mm_cid(p);
> >>> +       p->prev_sleep_sum_runtime = 0;
> >>> +       p->duration_avg = 0;
> >>>  }
> >>
> >> Beginning life biased toward stacking?
> >>
> > 
> > OK, so I should change the short_task() to skip the 0 duration_avg, to avoid
> > task stacking in the beginning.
> >    
> >>>  DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
> >>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> >>> index 41b58387023d..445877069fbf 100644
> >>> --- a/kernel/sched/fair.c
> >>> +++ b/kernel/sched/fair.c
> >>>
> >>> @@ -6905,6 +6914,9 @@ static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
> >>>  
> >>>  dequeue_throttle:
> >>>         util_est_update(&rq->cfs, p, task_sleep);
> >>> +       if (task_sleep)
> >>> +               dur_avg_update(p);
> >>> +
> >>>         hrtick_update(rq);
> >>>  }
> >>>
> >>
> >> That qualifier looks a bit dangerous.  Microbench components tend to
> >> have only one behavior, but the real world goes through all kinds of
> >> nutty gyrations, intentional and otherwise.
> >>
> > 
> > Understand. Unfortunately I don't have access to production environment
> > so I have to rely on microbenchmarks and a OLTP to check the result. I
> > get feedback from Abel that the former version of this patch brought
> > benefit to some short tasks like Redis in the production environment[1].
> > https://lore.kernel.org/lkml/36ba3b68-5b73-9db0-2247-061627b0d95a@bytedance.com/
> 
> Since the discussion was about real-life workload performance, I ran the DayTrader
> workload with three users and three instances. The results show no performance
> regression, and a 1% performance gain was observed, which is within the standard
> deviation.
>

Thank you Madadi for your test, let me respin this version and send it out with the
fixes after talked with Mike.

thanks,
Chenyu