Intel Updates x86 Hybrid CPU Cluster Scheduling For The Linux Kernel
Earlier this year Intel posted a new round of Linux cluster scheduling patches after their original implementation I found to be causing regressions and hurting performance at the time with Alder Lake when their original cluster scheduling work was being tackled in 2021. With the 2023 incarnation things appear to be in much better shape.
In June were the v2 patches and on Friday succeeded by a third version. This newest version simplifies how the sibling imbalance is computed and removes the asym packing bias, rounding is added to the sibling imbalance, and some basic changes.
The latest benchmarks are showing some minor performance improvements to a variety of single and threaded workloads, mostly though in the ~1% range.
The v3 cluster scheduling patches are too late for the Linux v6.5 kernel cycle but perhaps we'll finally see this code ready to go for v6.6 at the end of the summer. Those wishing to test out the v3 patches can find them out for review on the kernel mailing list. Once they get picked up by mainline I'll be running some fresh Intel hybrid CPU benchmarks on Linux.
1 Comment
