Linux Kernel Patches Boost NUMA Multi-Threaded Workloads

Red Hat developers have been working to improve the performance of multi-threaded workloads on NUMA systems.

Rik van Riel of Red Hat sent out two patches on Monday to refactor the NUMA (non-uniform memory access) balancing code as it spends too much CPU time scanning and faulting during multi-threaded workloads.

He commented on the patch series, "This patch set slows down NUMA PTE scanning when there are lots of shared faults, and when dealing with large NUMA groups that have a large fraction of shared faults."

With these two patches that affect just around one hundred lines of code in sched/fair.c, their tests have found 10 to 30% better performance on a four-node system in NAS benchmarks, 5~10% performance improvements in SPECjbb2005, and similar to small performance improvements in SPECjvm2008.

Interesting results and hopefully we'll see this code land come Linux 4.14 if these performance improvements are indeed valid.