Re: [PATCH] sched: consider WF_SYNC to find idle siblings

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

Hi Mel,
On 2022-11-01 at 09:41:57 +0000, Mel Gorman wrote:
> On Thu, Oct 27, 2022 at 01:26:03PM -0700, Andrei Vagin wrote:
> > From: Andrei Vagin <avagin@gmail.com>
> > 
> > WF_SYNC means that the waker goes to sleep after wakeup, so the current
> > cpu can be considered idle if the waker is the only process that is
> > running on it.
> > 
> > The perf pipe benchmark shows that this change reduces the average time
> > per operation from 8.8 usecs/op to 3.7 usecs/op.
> > 
> > Before:
> >  $ ./tools/perf/perf bench sched pipe
> >  # Running 'sched/pipe' benchmark:
> >  # Executed 1000000 pipe operations between two processes
> > 
> >      Total time: 8.813 [sec]
> > 
> >        8.813985 usecs/op
> >          113456 ops/sec
> > 
> > After:
> >  $ ./tools/perf/perf bench sched pipe
> >  # Running 'sched/pipe' benchmark:
> >  # Executed 1000000 pipe operations between two processes
> > 
> >      Total time: 3.743 [sec]
> > 
> >        3.743971 usecs/op
> >          267096 ops/sec
> > 
> 
> The WF_SYNC hint in unreliable as the waking process does not always
> go to sleep immediately. While it's great for a benchmark like a pipe
> benchmark as the relationship is strictly synchronous, it does not work
> out as well for networking which can use WF_SYNC for wakeups but either
> multiple tasks are being woken up or the waker does not go to sleep as
> there is sufficient inbound traffic to keep it awake. There used to be
> an attempt to track how accurate WF_SYNC was, using avg_overlap I think,
> but it was ultimately removed.
avg_overlap was removed 10 years ago because of accuracy problem that
"we are missing the necessary call to update_curr()" according to
commit e12f31d3e5d3 ("sched: Remove avg_overlap"). But in current code
I think this issue described in above commit does not exist anymore because
in current code the put_prev_task() would invoke update_curr() for each
entity, then calculating the avg_overlap is always using the update-to-date
runtime? If it is true, is it applicable to bring avg_overlap back?

Some benchmarks suffer from cross-CPU wakeup which introduces rq lock
contention. Similar to this patch, I tracked the average duration of the
task and place the wakee to a CPU where only 1 short-running task is running,
which is another direction to mitigate cross-CPU wakeup[1]. Not sure if we
could deal with more accurately?

[1] https://lore.kernel.org/lkml/6b81eea9a8cafb7634f36586f1744b8d4ac49da5.1666531576.git.yu.c.chen@intel.com/

thanks,
Chenyu