Linux 5.10.17 Backports CPUFreq Patches From 5.11 - Benchmarks
Released yesterday was the Linux 5.10.17 LTS kernel and what makes this point release a bit more notable than usual is that it backports the CPUFreq patches from 5.11 that were used for addressing the earlier AMD performance regression on Linux 5.11 and often leading to net improvements as well over prior kernel series. The CPUFreq patches were back-ported while the AMD frequency invariance support was not, so what does the performance look like for the Linux 5.10 LTS kernel? Here are some benchmarks.
The newly-released Linux 5.10.17 kernel back-ported "cpufreq: ACPI: Extend frequency tables to cover boost frequencies" and "cpufreq: ACPI: Update arch scale-invariance max perf ratio if CPPC is not there". See this prior article for additional context but long story short these were the two patches to address the performance issue on Linux 5.11 when AMD frequency invariance was introduced for Zen 2 / Zen 3 and used when using the likes of the Schedutil governor.
Curious if these CPUFreq patches meant any AMD performance change now on Linux 5.10 without frequency invariance support, I ran some benchmarks of Linux 5.10.16 vs. 5.10.17 vs. 5.11.0 while using the default Schedutil governor throughout. Tests were done on the Ryzen 9 5900X system that previously exhibited the performance drop during 5.11 development.
But for most workloads, the Linux 5.10.17 performance was the same as the 5.10.16 kernel. The Linux 5.11 kernel remains performing much faster than Linux 5.10 LTS series on AMD Zen 2/3.
Out of 46 tests carried out since yesterday on this Ryzen 9 5900X, the Linux 5.11 kernel was leading in the majority of them, as expected...
Taking the geometric mean of those 46 results, with those tests Linux 5.11 is about 4~5% faster than Linux 5.10 LTS even with this new point release.
All the benchmarks for those interested over on this OpenBenchmarking.org page. Meanwhile I'm already working on benchmarks looking at the flow of new patches heading into Linux 5.12... Stay tuned!
The newly-released Linux 5.10.17 kernel back-ported "cpufreq: ACPI: Extend frequency tables to cover boost frequencies" and "cpufreq: ACPI: Update arch scale-invariance max perf ratio if CPPC is not there". See this prior article for additional context but long story short these were the two patches to address the performance issue on Linux 5.11 when AMD frequency invariance was introduced for Zen 2 / Zen 3 and used when using the likes of the Schedutil governor.
Curious if these CPUFreq patches meant any AMD performance change now on Linux 5.10 without frequency invariance support, I ran some benchmarks of Linux 5.10.16 vs. 5.10.17 vs. 5.11.0 while using the default Schedutil governor throughout. Tests were done on the Ryzen 9 5900X system that previously exhibited the performance drop during 5.11 development.
But for most workloads, the Linux 5.10.17 performance was the same as the 5.10.16 kernel. The Linux 5.11 kernel remains performing much faster than Linux 5.10 LTS series on AMD Zen 2/3.
Out of 46 tests carried out since yesterday on this Ryzen 9 5900X, the Linux 5.11 kernel was leading in the majority of them, as expected...
Taking the geometric mean of those 46 results, with those tests Linux 5.11 is about 4~5% faster than Linux 5.10 LTS even with this new point release.
All the benchmarks for those interested over on this OpenBenchmarking.org page. Meanwhile I'm already working on benchmarks looking at the flow of new patches heading into Linux 5.12... Stay tuned!
7 Comments