More Benchmarks From Linux 5.5 Looking Like A Scheduler Snafu Even On Smaller CPUs
For the Linux 5.5 kernel that's about half-way through its development phase we have been pointing out some rather significant performance regressions affecting both AMD EPYC and Intel Xeon servers but there are also regressions to be found with desktop class systems too.
In still running through more Linux 5.5 performance tests this weekend with still seeing lower performance in a number of workloads, here are some of the latest numbers on additional and distinct platforms.
Like the dual socket servers, the AMD Threadripper 3970X was seeing some regressions off Linux 5.5 compared to 5.4 and 5.3 stable series.
Though on the Threadripper box, some of the workloads negatively impacted on the larger core count, dual socket servers weren't showing lower performance off Linux 5.5.
But when running the Hackbench Linux kernel scheduler benchmark there was indeed lower performance spotted with Linux 5.5.
And some of the database workloads were regressed on Linux 5.5 like seen with the previous server tests.
The schbench scheduler test also pointed to a regression on Linux 5.5.
Meanwhile I also fired up some Linux 5.4 vs. 5.5 Git tests on a completely different system, an Intel Core i9 9900KS desktop.
Hackbench continued pointing to lower scheduler performance on Linux 5.5 with all of the systems we have tried so far.
Some of the other test profiles seeing slower performance on Linux 5.5 with the higher core count systems were also pointing to slightly slower performance on the Core i9-9900KS but not by as wide of margins.
With Linux 5.5 carrying some admittedly large changes to its load balancing code and an acknowledged risk of performance regressions, presumably that is at least part of the Linux 5.5 slowdown we're seeing in a number of workloads, especially given the Hackbench numbers. But I'm spending more time this holiday weekend further investing these performance changes on more workloads and more systems.
In still running through more Linux 5.5 performance tests this weekend with still seeing lower performance in a number of workloads, here are some of the latest numbers on additional and distinct platforms.
Like the dual socket servers, the AMD Threadripper 3970X was seeing some regressions off Linux 5.5 compared to 5.4 and 5.3 stable series.
Though on the Threadripper box, some of the workloads negatively impacted on the larger core count, dual socket servers weren't showing lower performance off Linux 5.5.
But when running the Hackbench Linux kernel scheduler benchmark there was indeed lower performance spotted with Linux 5.5.
And some of the database workloads were regressed on Linux 5.5 like seen with the previous server tests.
The schbench scheduler test also pointed to a regression on Linux 5.5.
Meanwhile I also fired up some Linux 5.4 vs. 5.5 Git tests on a completely different system, an Intel Core i9 9900KS desktop.
Hackbench continued pointing to lower scheduler performance on Linux 5.5 with all of the systems we have tried so far.
Some of the other test profiles seeing slower performance on Linux 5.5 with the higher core count systems were also pointing to slightly slower performance on the Core i9-9900KS but not by as wide of margins.
With Linux 5.5 carrying some admittedly large changes to its load balancing code and an acknowledged risk of performance regressions, presumably that is at least part of the Linux 5.5 slowdown we're seeing in a number of workloads, especially given the Hackbench numbers. But I'm spending more time this holiday weekend further investing these performance changes on more workloads and more systems.
9 Comments