Re: [patch v3.18+ regression fix] sched: Further improve spurious CPU_IDLE active migrations

[Mike Galbraith <umgwanakikbuti@gmail.com>]

On Wed, 2016-08-31 at 17:52 +0200, Vincent Guittot wrote:
> On 31 August 2016 at 12:36, Mike Galbraith <umgwanakikbuti@gmail.com> wrote:
> > On Wed, 2016-08-31 at 12:18 +0200, Mike Galbraith wrote:
> > > On Wed, 2016-08-31 at 12:01 +0200, Peter Zijlstra wrote:
> > 
> > > > So 43f4d66637bc ("sched: Improve sysbench performance by fixing spurious
> > > > active migration") 's +1 made sense in that its a tie breaker. If you
> > > > have 3 tasks on 2 groups, one group will have to have 2 tasks, and
> > > > bouncing the one task around just isn't going to help _anything_.
> > > 
> > > Yeah, but frequently tasks don't come in ones, so, you end up with an
> > > endless tug of war between LB ripping communicating buddies apart, and
> > > select_idle_sibling() pulling them back together.. bouncing cow
> > > syndrome.
> > 
> 
> replacing +1 by +2 fixes this use case that involves 2 threads but
> similar behavior can happen with 3 tasks on system with 4 cores per MC
> as an example
> 
> IIUC, you have on
> - one side, periodic load balance that spreads the 2 tasks in the system
> - on the other side, wake up path that moves the task back in the same MC.

Yup.

> Isn't your regression more linked to spurious migration than where the
> task is scheduled ? I don't see any direct relation between the client
> and the server in this netperf test, isn't it ?

         netperf  4360 [004]  1207.865265:       sched:sched_wakeup: netserver:4361 [120] success=1 CPU:002
         netperf  4360 [004]  1207.865274:       sched:sched_wakeup: netserver:4361 [120] success=1 CPU:002
         netperf  4360 [004]  1207.865280:       sched:sched_wakeup: netserver:4361 [120] success=1 CPU:002
       netserver  4361 [002]  1207.865313:       sched:sched_wakeup: netperf:4360 [120] success=1 CPU:004
         netperf  4360 [004]  1207.865340:       sched:sched_wakeup: kworker/u16:4:89 [120] success=1 CPU:000
         netperf  4360 [004]  1207.865345:       sched:sched_wakeup: kworker/u16:5:90 [120] success=1 CPU:006
         netperf  4360 [004]  1207.865355:       sched:sched_wakeup: kworker/u16:5:90 [120] success=1 CPU:006
         netperf  4360 [004]  1207.865357:       sched:sched_wakeup: kworker/u16:4:89 [120] success=1 CPU:000
         netperf  4360 [004]  1207.865369:       sched:sched_wakeup: netserver:4361 [120] success=1 CPU:002
       netserver  4361 [002]  1207.865377:       sched:sched_wakeup: netperf:4360 [120] success=1 CPU:004
         netperf  4360 [004]  1207.865476:       sched:sched_wakeup: perf:4359 [120] success=1 CPU:003

It's not limited to this load, anything at all that is communicating
will do the same on these or similar processors.

This trying to be perfect looks like a booboo to me, as we are now
specifically asking our left hand undo what our right hand did to crank
up throughput.  For the diagnosed processor at least, one of those
hands definitely wants to be slapped.

This doesn't seem to be an issue for L3 equipped CPUs, but perhaps is
for some even modern processors, dunno (the boxen where regression was
detected are far from new).

> we could either remove the condition which tries to keep an even
> number of tasks in each group until busiest group becomes overloaded
> but it means that unrelated tasks may have to share same resources
> or we could try to prevent the migration at wake up. I was looking at
> wake_affine which seems to choose local cpu  when both prev and local
> cpu are idle. I wonder if local cpu is  really a better choice when
> both are idle

I don't see a great alternative to turning it off off the top of my
head, at least for processors with multiple LLCs.  Yeah, unrelated
tasks could end up sharing a cache needlessly, but will that hurt as
badly as tasks not munching tasty hot data definitely does?

	-Mike