Linux 5.11 Sees Quick Fix For A Context Switching Performance Regression
Linux 5.11-rc1 is due for release later today and at least one regression has seen a quick resolution.
Intel's kernel test robot saw a ~1.6% performance regression in one of its scheduler benchmarks. The change causing the regression was trying to ensure only per-CPU KThreads run during hotplug.
Thanks to that timely testing by Intel, There is a fix of optimizing the finish_lock_switch() function within the scheduler code. Kernel developer Peter Zijlstra noted, "Even though we were careful to replace a single load with another single load from the same cacheline. Restore finish_lock_switch() to the exact state before the offending patch and solve the problem differently."
That fix was sent in today ahead of the Linux 5.11-rc1 release. Great seeing one more performance regression resolved timely albeit a small one in this case.
As to the separate regression we've been looking at with AMD performance regressing on Linux 5.11 when using the common default "Schedutil" governor, I have some new data likely going up in a few hours once those tests complete looking at some different configurations.