Monday, November 11, 2019

Linux syscall performance regressions explained

There is a paper in SOSP 2019 that explains Linux system call performance for recent kernels. Explaining performance is what I do so I appreciate the effort that went into this paper:
  • Linux syscall perf mostly improved from 3.0 to 4.0 then went bad starting with 4.7 thanks to 5 security changes - Spectre, Meltdown, harden usercopy, randomize slab freelist, user pagefault handling.
  • Linux has a lot of variance in syscall performance across releases, more than I am used to with MySQL. But MySQL still has regressions across release just without so much variance. Interesting posts for Postgres perf across versions are here, here and here. AFAIK Linux can't move as fast as we need it to and avoid the variance, so the it is caught and fixed after the fact.
  • Results are for Intel CPUs. It would be interesting to see results for AMD and for Intel with and without HT enabled.

No comments:

Post a Comment