Linux 5.5-rc3 Benchmarks Are Still Pointing To Slips In Performance

Early on in the Linux 5.5 cycle during the merge window we saw some wild swings in performance including some positive gains but also performance regressions. Given last weekend's Linux 5.5-rc3 release having merged some scheduler fixes and other fallout from early on in Linux 5.5, I was curious to see if those regressions have been addressed... Sadly, they are not.

Those earlier tests were on Intel Xeon Cascade Lake but even on AMD EPYC I can hit similar performance woes on Linux 5.5. Sadly, even with Linux 5.5-rc3 there are some apparent regressions.

Using an AMD EPYC 7742 2P server, I ran some fresh tests of Linux 5.5 Git compared to Linux 5.4 while also tossing in Linux 5.3 stable too, just for ruling out noisy tests, etc. There are many cases of threaded workloads seeing performance regressions on Linux 5.5 compared to both 5.4 and 5.3. No changes to the system hardware or software besides swapping out the kernel builds.






The Hackbench results seem to indicate a big regression in the scheduler.






Again these Linux 5.5 performance hits seem to happen both on Intel and AMD, at least for larger core count systems. Will work on doing some bisecting this weekend now that these regressions have been in mainline for a few weeks with no cure. If you wish to support all of our daily Linux benchmarking and to see more bisecting work and more, consider showing your support for Phoronix this holiday season. On a geometric mean basis across dozens of benchmarks, Linux 5.5-rc3 is about 6% lower than Linux 5.4 stable.