Re: [PATCH v1] sched/fair: update_pick_idlest() Select group with lowest group_util when idle_cpus are equal

[Mel Gorman]

On Mon, Nov 02, 2020 at 09:44:18AM -0500, Phil Auld wrote:
> > I'm going to have a look at the regressions and see if  patches that
> > have been queued for v5.10 or even more recent patch can help or if
> > the patch should be adjusted
> >
> 
> Fwiw, we have pulled this in, along with some of the 5.10-rc1 fixes in this
> area and in the load balancing code.
> 

I assume you mean a distro kernel but in this case, all the bisections
were vanilla mainline.

> We found some load balancing improvements and some minor overall perf
> gains in a few places, but generally did not see any difference from before
> the commit mentioned here.
> 
> I'm wondering, Mel, if you have compared 5.10-rc1? 
> 

No, but it's queued now -- 5.9 vs 5.9-revert vs 5.10-rc2 vs
5.10-rc2-revert. It's only one machine queued but hopefully it'll
reproduce. Both 5.9 and 5.10 are being tested in case one of the changes
merged in 5.10 mask the problem. Ordinarily I would have checked first
but I'm backlogged so I took a report-first-test-later approach this
time around.

> We don't have everything though so it's possible something we have
> not pulled back is interacting with this patch, or we are missing something
> in our testing, or it's better with the later fixes in 5.10 or ...
> something else :)
> 

Add userspace differences, core counts, CPU generation, volume of scheduler
changes with interactions, different implementations of tests, masking from
cpufreq changes, phase of the moon and just general plain old bad luck.

> I'll try to see if we can run some direct 5.8 - 5.9 tests like these. 
> 

That would be nice. While I often see false positive bisections for
performance bugs, the number of identical reports and different machines
made this more suspicious.

-- 
Mel Gorman
SUSE Labs