CPU Cluster Scheduler Continues To Be Worked On For Linux With Promising Results
HiSilicon has been pursuing this "cluster scheduler" for the Linux kernel in order to enhance the performance of the Kunpeng 920 ARM server chip that has six or eight clusters per NUMA node and each cluster being comprised of four CPU cores with shared L3 cache access among the clusters. But there is also the possibility of this scheduler helping some x86 hardware too, like Intel's Jacobsville is noted for its clusters of Atom cores.
The HiSilicon Linux kernel work now up to its sixth round of patches is for exposing this topology and having a CPU scheduler to properly/efficiently deal with the layout. The goal with the scheduler is for spreading unrelated tasks among the multiple clusters to reduce contention and then to also gather related tasks within a cluster for improving cache affinity.
With benchmarks shown by HiSilicon engineers, the scheduler patches tested on the Kunpeng 920 server SoC were able to see the Stream memory benchmark performance improve by more than 20% compared to the stock kernel. The patches under the "request for comments" flag at this stage would expose the cluster scheduler functionality with the SCHED_CLUSTER Kconfig build switch.
While HiSilicon's primary focus is obviously on the ARM support, there is a patch tacked onto the series from an Intel Linux engineer for handling this SCHED_CLUSTER functionality for x86 too. These cluster optimizations will likely have more importance moving forward with new hardware platforms.
Those wanting to dig into this current work from HiSilicon on the Linux cluster scheduler can find the patches on the kernel mailing list.
4 Comments
