Linux 6.14 Sees Last Minute Fix For A Two Year Old Regression Causing A 30% Performance Drop
Submitted today ahead of the Linux 6.14 stable release expected Sunday is a lone scheduler fix for the kernel. This patch is for reverting a change made to the Linux kernel two years ago that ended up regressing some workloads with a significant performance hit.
Merged nearly two years ago to the day was this patch to the core scheduler code for trying to reduce the cost of the sched_move_task handling when using the CONFIG_SCHED_AUTOGROUP configuration. With a simple loop of a bash script launching many "sleep" commands as separate processes, this patch ended up reducing that execution time by around 57%.
But it turns out that the attempted optimization two years ago has been hurting other Linux workloads. With today's patch to revert the code, this two-year-old code was found to regression UnixBench's spawn test by around 30%. This performance regression was reported by an Amazon engineer when running tests on their AWS cloud with auto-group enabled. Other workloads besides UnixBench are ultimately affected as well.
Ingo Molnar sent out the scheduler pull request today with this lone revert and commented:
Barring any reservations by Linus Torvalds, this performance regression fix/revert should be merged to Linux 6.14 Git later today.
Merged nearly two years ago to the day was this patch to the core scheduler code for trying to reduce the cost of the sched_move_task handling when using the CONFIG_SCHED_AUTOGROUP configuration. With a simple loop of a bash script launching many "sleep" commands as separate processes, this patch ended up reducing that execution time by around 57%.
But it turns out that the attempted optimization two years ago has been hurting other Linux workloads. With today's patch to revert the code, this two-year-old code was found to regression UnixBench's spawn test by around 30%. This performance regression was reported by an Amazon engineer when running tests on their AWS cloud with auto-group enabled. Other workloads besides UnixBench are ultimately affected as well.
Ingo Molnar sent out the scheduler pull request today with this lone revert and commented:
"This is admittedly a bit late in the cycle, and the regression is old, but the performance impact is substantial for the affected workloads so I didn't want to delay this fix."
Barring any reservations by Linus Torvalds, this performance regression fix/revert should be merged to Linux 6.14 Git later today.