Re: [sched/fair] ddb3b1126f: hackbench.throughput -25.9% regression

[Chen Yu]

On Thu, Mar 31, 2022 at 11:42 AM Tim Chen <tim.c.chen@xxxxxxxxxxxxxxx> wrote:
>
> On Wed, 2022-03-30 at 17:46 +0800, kernel test robot wrote:
> >
> > Greeting,
> >
> > FYI, we noticed a -25.9% regression of hackbench.throughput due to commit:
> >
>
> Will try to check the regression seen.
>
Double check  that the regression could be reproduced on top of the
latest sched/core branch:
parent ("sched/fair: Don't rely on ->exec_start for migration")
fbc  ("sched/fair: Simple runqueue order on migrate")

parent                                fbc
91107           -40.8%      53897        hackbench.throughput

and it is consistent with lkp's original report that the  context
switch count is much higher with patch applied:
9591919          +510.3%   58534937
hackbench.time.involuntary_context_switches
36451523          +281.5%  1.391e+08
hackbench.time.voluntary_context_switches

Considering that this patch  'raises' the priority of the migrated
task, by giving it the cfs_rq->min_vruntime,
it is possible that the migrated task would preempt the current
running task more easily.

0.00           +12.2       12.21
perf-profile.calltrace.cycles-pp.enqueue_entity.enqueue_task_fair.ttwu_do_activate.try_to_wake_up.autoremove_wake_function
and the patched version has  spent more time on enqueue_entity(),
which might be caused by setting sched entity hierarchy from leaf to
root,
which was mentioned in another thread.

-- 
Thanks,
Chenyu