As discussed offline, we think there are a few improvements we could make in how we do performance benchmarks for H3:
use some standard benchmarking framework instead of rolling our own (which is what we currently do)
a framework could provide nice features like warm-ups, advanced stats (min vs avg), etc.
what options currently exist for C?
standardize the benchmark format to make it more machine-readable (like in Go), which would make it easier to automatically compare timings between diffs
avoid doing benchmarks on laptops and/or finding a dedicated benchmarking machine
random (but deterministic) inputs for improved input coverage
benchmark performance tracking/plotting over the timeline of diffs, like what is done in Python using asv for libraries like numpy
For context, this issue came up after seeing a lot of variance in the benchmarks I was running on my laptop for https://github.com/uber/h3/pull/496
ajfriend
commented
3 years ago
As discussed offline, we think there are a few improvements we could make in how we do performance benchmarks for H3:
asvfor libraries likenumpyFor context, this issue came up after seeing a lot of variance in the benchmarks I was running on my laptop for https://github.com/uber/h3/pull/496