Re: [PATCH 0/8] sched/fair: misfit task load-balance tweaks

[Valentin Schneider]

On 04/02/21 13:03, Dietmar Eggemann wrote:
> How did you verify the benefit of these changes?
>
> It's clear that you need a platform with capacity_orig diffs <20%
> between CPU types (like Pixel4 - SD855 (4x261, 3x871, 1x1024) or QC's
> RB5 platform - SD865 (4x284, 3x871, 1*1024)) but which
> benchmark/testcase did you use?

Benchmark is the usual culprit:

  https://lisa-linux-integrated-system-analysis.readthedocs.io/en/master/kernel_tests.html#lisa.tests.scheduler.misfit.StaggeredFinishes

This test spawns 1 CPU hog per CPU, and screams whenever a CPU of capacity
X is running a hog while another CPU of capacity Y > X has been idling for
"too long" (a few ms). IOW it makes sure upmigration happens in a timely
manner.

Some of the test platforms (Juno (4+2 big.LITTLE), HiKey960 (4+4
big.LITTLE)) show some improvements due to the last 2 patches.

As for systems with CPUs in the [819-1024] "deadzone", Ionela's been kindly
running said test on said RB5, and the upmigrations look just fine with the
patches applied.