With a patch on hold, the performance of Linux 5.11 AMD Zen 2 / Zen 3 looks great in terms of out-of-the-box performance when using Schedutil, as it becomes the default CPU frequency scaling governor for multiple default distributions / cores. With the aforementioned Linux 5.11 regression, approached since the first invariance support of the AMD processor was first introduced, the performance of Schedutil from small Ryzen systems to large EPYC hardware looks pretty good. But how much remains in relation to the optimal performance of scaling the processor frequency with the “performance” governor? Here’s a look at those reference standards on Ryzen and EPYC for Schedutil vs. Performance on a Linux 5.11 patch patch.
As a continuation of last week’s test, analyzing the patched Linux 5.11 kernel that leads to better Schedutil behavior with AMD Zen 2 / Zen 3 hardware, this article provides some benchmarks on how well the governor compares Schedutil with patches to the “performance” governor on the same core. Most Linux distributions are not yet default to the performance governor, but using “ondemand” or newer kernels often uses the “schedulutil” for AMD hardware with the CPUFreq driver.
When you test on an AMD Ryzen 9 5950X, there’s still room to further adjust the Schedutil governor for better performance:
Ignoring some key exceptions for some faster running tests, in which the benefits of performance governor are more pronounced compared to Schedutil, which requires a (short) time to increase clock frequencies, for most tasks, the performance governor fails just a few percent faster performance. This is not so bad, given that this Linux 5.11 patch kernel for testing is much better than the considerable regressions seen earlier in cycle 5.11 and the patch that often leads to performance beyond where it Linux 5.10 stable worked.
218 different test cases were run on this Ryzen 9 5950X system compared to the two CPUFreq governors. The graph above shows only the workloads with a measurable and statistically significant difference.
If we take the geometric average of all 218 results, the performance governor was just over 1% faster than the default Schedutil governor on Linux 5.11.
But let’s see now the difference Schedutil vs. Performance on the AMD EPYC, while monitoring the power consumption of the processor, as shown by the AMD_Energy driver