Posts tagged "when the benchmark lies"

I measured eight of the clocks that clock_gettime can read, and on my laptop two of them disagreed by 43,257 ns after 20 seconds. I measured their resolution, their cost, which ones enter the kernel, and which clocks Rust and Python use.

My toy server stops for half a second every five seconds. A client that waits for each answer saw 4 slow requests out of 1.36 million, and a p99 of 30 microseconds. A client with a fixed schedule saw 1,608 slow requests and a p99 of 450 milliseconds.

I timed four million single Vec::push calls: mean 22.8 ns, slowest 315,684 ns. The slow ones were not the reallocations but page faults, and huge pages made them bigger. Then a simulated request handler where the mean is a time that no request had.

On a laptop with two kinds of cores, I measured the same benchmark on a P-core and on an E-core, without taskset, and with two threads on one core. The E-core needed more cycles, not only a lower clock, and a busy sibling thread made a P-core as slow as an E-core.

I measured the work of one loop and the clock of its core at the same time, for one minute. The clock moved between 3.3 and 4.2 GHz without any change in my program, and the work per cycle stayed the same in every phase. Here is the trace, and what I can do about it on a laptop.