1. 21
  1. 2

    A thing I remember seeing years ago was someone benchmarking mysqld compiled with different gcc options and finding that it used less time on CPU if compiled with -Os instead of -O2.

    At the time I assumed it was due to L1i misses but maybe it was TLB misses instead? Both are transparent to the OS so both show up as just “cpu time” usage - unless you look at hardware performance counters, which I don’t remember them doing.

    1. 1

      Very interesting work. Second to last graph has deceptive X axes though. The next one is fine, so it’s probably an accident.