Linux 5.4 Should Improve NUMA Hugepage Allocation Performance

It turns out Linux 5.3 shipped with potentially subpar performance for the allocation of hugepages but that should be rectified in the now open Linux 5.4 cycle for trying to provide a sane default allocation strategy on NUMA boxes.

With Linux 5.4, the kernel will now avoid the reclaim operation when compaction isn't likely to succeed. It will also allow hugepages to fallback to remote nodes when seeking hugepages via madvise but the node-local hugepage allocation failed. Additionally, Linux 5.4 reverts two patches from Linux 5.3 that ended up regressing other workloads unintentionally.

Ultimately, if all goes well and long story short, Linux 5.4 should be offering a more sane default allocation strategy for hugepages on NUMA systems. That in turn should help performance. More details via this Git merge that happened on Saturday.