Linux 5.11 Is Now Looking Great For AMD Zen 2 / Zen 3 Performance
Not only is the AMD "CPU frequency invariance regression" from that new support with the in-development Linux 5.11 kernel on course to address the performance shortcomings I outlined last month, but with the patched kernel for a number of workloads the performance is now ahead of where it was at with Linux 5.10.
At the end of last year I pointed out that the early Linux 5.11 kernel code was regressing hard for AMD performance when using the "Schedutil" scheduler utilization governor as is now often the default behavior of the kernel. As bisected back then and outlined in full with that aforelinked article, the performance regressions seen on AMD Zen 2 and Zen 3 systems stem from the introduction of CPU frequency invariance support added this cycle.
Now the shortcoming with that initial CPU frequency invariance support is in the process of being worked out. As outlined last night, a fix has been proposed for addressing that major performance regression for AMD systems with Linux 5.11.
That fix hasn't been merged yet to mainline but I have been eagerly testing the patch now. While less than 24 hours into the testing, I can already say: Linux 5.11 is now in much better shape for AMD. Not only is the performance issue resolved, but for multiple workloads the Linux 5.11 performance is better off than 5.10!
In this article are some preliminary figures using an AMD EPYC 7702 server and AMD Ryzen 9 5950X desktop. As outlined in prior articles, this CPU frequency invariance support in combination with the CPUFreq governor only impact Zen 2 hardware and newer due to its reliance on ACPI CPPC. Let's take a look at some of these initial numbers now off a patched kernel while in the coming days will be more workloads tested and more AMD CPUs being benchmarked. Follow-up articles will also offer a look at the CPU power efficiency/consumption as well as seeing how this Schedutil performance compares now to the performance governor - today's article is just the culmination of a day's worth of tests.