Ubuntu Generic vs. Low-Latency Linux Kernel Benchmarks For HPC & Desktop
With it mostly being some Kconfig changes these days for their low-latency kernel build, they are looking at incorporating them into their generic kernel build. This would save on having to maintain the "lowlatency" kernel build with the additional build resources, QA/testing, and simplicity on the user side with more versatility out of the generic kernel build.
As part of weighing the change, Canonical kernel engineers have acknowledged this could cause some performance regressions particularly in the area of HPC. So with the current Linux 6.5-based kernel on Ubuntu 23.10 I ran benchmarks of their existing "generic" and "lowlatency" kernel builds.
This round of testing was on an AMD EPYC 8534P "Siena" 64-core server and simply running the same set of benchmarks in the same configuration on the default "generic" kernel against the same kernel version in its "lowlatency" form. HPC and other general workloads were the focus to see what the performance implications may be of using the low-latency kernel build.
Of 148 benchmarks run on both kernel flavors, the generic kernel was faster by just 1%...
Only in a handful of the 148 benchmarks was there a measurable difference in these HPC/workstation/desktop benchmarks. Most of the differences were within the Stress-NG kernel micro-benchmarks, which is to be expected. There were also some minor fluctuations within the video encoding and PostgreSQL database benchmarks, among others. Most of the time though that the generic kernel was faster tended to be by 4% or less.
See all of the benchmarks in full for those interested. It will be interesting to see what Canonical decides around the low-latency kernel version for the Ubuntu 24.04 LTS due out in April along with various other performance tuning/enhancements.
44 Comments
