Linux 5.18 Scheduler Updates Improve NUMA Balancing For AMD EPYC Servers
When it comes to NUMA balancing within the scheduler of Linux 5.18, there are improvements for dealing with CPU-less nodes and for tuning systems that have multiple LLC cache domains per node. The latter change around multiple caches per node happens to be of much use for modern AMD server processor designs.
This is the previously talked about scheduler change shown to have a very positive impact for some workloads on AMD EPYC servers.
Last month I ran some benchmarks and found this change did indeed help a number of workloads over the current state in Linux 5.17. I'll have additional AMD (and Intel) CPU benchmarks on Linux 5.18 once the merge window settles down for seeing how the performance is looking overall.
Aside from the NUMA balancing improvements to the scheduler code for Linux 5.18, there are also cleanups to the deadline scheduler code, CPU accounting fixes, preempt-dynamic support on AArch64, and other fixes.
The scheduler code has also pulled in its portion of the "Fast Kernel Headers" patches touching it while the bulk of the "FKH" patches are still undergoing review/testing for elsewhere in the kernel as part of that massive patch series.
The full list of scheduler updates for inclusion in Linux 5.18 can be found via this morning's pull request.