AMD Performance On Linux 5.11 Remains Mixed Due To Schedutil With Frequency Invariance

On the opposite side, when testing with an AMD EPYC 7F52 1P server I found Linux 5.11 to offer some nice improvements. PostgreSQL notably benefits as outlined previously but even for cases like Kvazaar that are regressing on smaller Ryzen CPUs, with the EPYC 7F52 saw some nice improvements with Linux 5.11 and only in one test case (GraphicsMagick with image rotation) happened to regress hard on Linux 5.11.

Or when running an Intel Core i9 10900K for reference purposes, there ended up being few tests with measurably different results (all of the tests without any statistically significant change are not shown in that side-by-side graph).

Likewise on an Intel Tiger Lake notebook with Core i7 1165G7 the performance from Linux 5.10 to 5.11 (and even 5.9) has largely been flat. That's what I have also been seeing with other Intel systems too.

After hammering many different systems with benchmarks on Linux 5.10 and 5.11, the same basic theme continues representing itself:

- Intel systems and older AMD systems are not seeing much change in performance on Linux 5.11 as expected, at least for the workloads being focused on as part of this frequency invariance investigation.

- AMD Zen 2 / Zen 3 systems seem quite prone to regressing on Linux 5.11 but depending upon the CPU there are also cases where Linux 5.11 is helping with the performance as the frequency invariance patches try to do.

- The AMD changes appear to indeed be caused by the frequency invariance support when running under the Schedutil governor. If running the same tests on Linux 5.10 and 5.11 but instead using the "performance" governor, the performance is generally quite flat except for PostgreSQL where we have noted separate changes with this new kernel.

Presumably this wasn't caught by the AMD and SUSE engineers working on the patches as it appears that Ryzen and/or smaller CPUs tend to see more regressions on Linux 5.11 with this support than the beefier EPYC processors, which at least based on the kernel patches is where the testing was taking place. Additionally, some of the workloads like some games and video encoders may have not been used in the evaluation of the frequency invariance behavior.

The AMD frequency invariance support at this stage is using a midpoint of max_boost and max_p since the maximum frequency/performance is uknown given dynamic power allocation between cores. The linked patch notes that using the midpoint over the maximum non-boost frequency or maximum boost frequency was slightly favored from the experimental data. Indeed when tweaking the performance ratio did help recover lost performance in some of the workloads but adversely impacting others.

So at this point there doesn't seem to be any straight-forward solution for improving the Linux 5.11 behavior short of switching over to the performance governor if you find yourself using Schedutil by default. But hopefully with time this AMD frequency invariance support will be adequately improved or ideally finding more desktop/server Linux distributions default to using the more reliable performance governor at least until Schedutil is less of a wreck.

Aside from that headache, Linux 5.11 is looking good for AMD customers with some prominent new features being supported.