You read that right; a single line of code from Intel has shown the Linux kernel could receive a performance jump of up to 4,000 percent by way of will-it-scale.per_process_ops.

A Intel kernel testing bot discovered this, and the change was added to commit efa7df3e3bb5 (“mm: align larger anonymous mappings on THP boundaries”). That commit was related to efficient memory management and mapping (mm and mmap) techniques that use Transparent Hugepages (THPs) and Page Middle Directory (PMD).

Before anyone gets too excited, the discovered performance increase was isolated to a synthetic test case, so real-world workloads will probably never see such incredible gains. However, it does make a solid case for how well Linux is capable of performing.

One issue with the change is that it was shown to significantly regress certain workloads by up to 600 percent (when using the cactusBSSN benchmark on certain platforms). Those regressions are related to the benchmark's access pattern that suffers from translation lookaside buffer (TLB) or cache aliasing from aligned boundaries of the individual areas.

According to author Vlastimil Babka from SUSE, “To fix the regression but still try to benefit from THP-friendly anonymous mapping alignment, add a condition that the size of the mapping must be a multiple of PMD size instead of at least PMD size. In case of many odd-sized mapping like the cactusBSSN creates, those will stop being aligned and with gaps between, and instead naturally merge again.”

You can read more about this discovery from this entry in the Kernel mailing list and this post from Vlastimil Babka.