Linux Cluster-Aware Scheduling Being Extended To AMD Processors

Back in 2021 saw work on CPU cluster-aware scheduling by HiSilicon engineers for Arm processors as well as Intel engineers with a focus on their Jacobsville platform being comprised of clusters of Atom cores. That x86 cluster-aware scheduling was enabled for capable Intel processors while now two years later is being extended for AMD processors.

The cluster scheduler support was merged back in the Linux 5.16 kernel while patches posted today by AMD engineer K Prateek Nayak extend it for working on AMD processors.

The set of kernel patches posted today wire up the "l2c_id" for being set on AMD processors with the Topology Extension "TOPOEXT" feature so that the set of threads sharing the same L2 cache can be properly mapped to the same cluster ID.

Up to now even for large AMD EPYC server processors the same cluster ID would be reported across all CPU cores. With the two patches sent out today, the cluster ID is now properly associated to each distinct set of L2 cache.

These AMD patches are fairly straight-forward for getting the AMD CPU cluster scheduler support going, so hopefully they'll be managed to get mainlined soon -- possibly even for the upcoming v6.4 cycle.

Making use of the Linux cluster scheduler requires setting the "CONFIG_SCHED_CLUSTER" option at build time for enhancing the CPU scheduler's decision making based on the layout of the cores. No performance numbers were provided as part of this patch series while it will be interesting to benchmark the impact of cluster scheduling particularly for large AMD EPYC servers.