Re: Abysmal scheduler performance in Linus' tree?

[Chris Wilson]

Quoting Andy Lutomirski (2017-09-06 06:13:39)
> I'm running e7d0c41ecc2e372a81741a30894f556afec24315 from Linus' tree
> today, and I'm seeing abysmal scheduler performance.  Running make -j4
> ends up with all the tasks on CPU 3 most of the time (on my
> 4-logical-thread laptop).  taskset -c 0 whatever puts whatever on CPU
> 0, but plain while true; do true; done puts the infinite loop on CPU 3
> right along with the make -j4 tasks.
> 
> This is on Fedora 26, and I don't think I'm doing anything weird.
> systemd has enabled the cpu controller, but it doesn't seem to have
> configured anything or created any non-root cgroups.
> 
> Just a heads up.  I haven't tried to diagnose it at all.

There is an issue on !llc machines arising from numa_wake_wide() where
processes are not spread widely across the low-power cores:
https://patchwork.kernel.org/patch/9875581/

The patch we are using to fix the regression is
https://cgit.freedesktop.org/drm-intel/commit/?h=topic/core-for-CI&id=6c362d9657293d700a8299acbeb2f1e24378f488
-Chris