Go Wiki: Benchmarks

These are benchmarks collected from the community used to measure the effects of changes to the Go core (compiler, runtime, garbage collector, and libraries). They should have the following properties:

• they matter; someone cares, perhaps in a dollars-and-cents way, that they run well
• they are go-gettable and don’t require customized steps for building the benchmark
• they run under go test -bench ...
• they run relatively quickly, ideally a single “run” takes less than a second (there should perhaps be a separate set of longer-running benchmarks)
• their timings are not gratuitously noisy
• they run cleanly in a restricted environment, such as a Docker or rkt container
• they’re not gratuitously redundant with other benchmarks already in the list; we don’t need ten microbenchmarks of Go transcendental functions

These benchmarks change over time, and that is okay. Their intended use is for performance testing of proposed changes; is the geometric mean better, were any benchmarks made substantially worse?

Information for each benchmark includes (or should include):

• a short name for the benchmark
• the path to go get the benchmark
• a regexp for the benchmark suite excluding individual benchmarks that might be noisy, long-running, or redundant
• (ideally) a contact person for questions about the benchmarks

There is a benchmark runner that automates downloading, building, and running these benchmarks under various (user-defined) configurations. Benchmarking noise on Linux can be somewhat reduced with perflock.

A few have been proposed but have so far failed to make the cut (for fetch, build, or noise problems):

This content is part of the Go Wiki.