Re: [RFC PATCHSET v2] sched/fair: fix load balancer behavior when cgroup is in use

From: Vincent Guittot
Date: Fri May 05 2017 - 04:47:18 EST

Hi Tejun,

On 4 May 2017 at 22:28, Tejun Heo <tj@xxxxxxxxxx> wrote:
> Hello,
> v1 posting can be found at
> The patchset is still RFC and based on v4.11. I used Peter's updated
> calc_cfs_shares() instead of scaling manually and updated so that
> runnable_load_avg is propagated independently from load_avg. Due to
> the way sched_entity and cfs_rq loads are calculated, this requires an
> extra runnable_load_avg calculation for group sched_entities, but the
> end result is cleaner and actually makes sense.
> Vincent, can you please verify whether the regression that you see is
> gone with this version?

schbench results looks better with this version
Latency percentiles (usec)
50.0000th: 212
75.0000th: 292
90.0000th: 385
95.0000th: 439
*99.0000th: 671
99.5000th: 7992
99.9000th: 12176
min=0, max=14855

p99 is back to a normal value but p99.5 stays higher than mainline

I have also checked load_avg and runnable_load_avg value and there is
something incorrect. I will provide details on the related patch

> Thanks.
> --
> tejun