Profiling

Rust benchmarking tools work at two levels. hyperfine benchmarks whole commands from the shell, running them repeatedly with warmup and reporting mean, spread and outliers, which is the quickest way to compare two binaries or flags. For functions inside your code, Criterion.rs runs statistically rigorous microbenchmarks on stable Rust and reports whether a change is a real regression or noise.

Divan is a newer benchmarking library with a simpler attribute-based API and support for measuring allocations. Bencher tracks benchmark results across commits in CI, so regressions are caught in pull requests rather than after release. Divan also lets one benchmark run over several types or input sizes, and can run it on multiple threads to measure contention, which is awkward to set up in Criterion.

For profiling, Bytehound records every allocation and shows where memory is used and leaked. hotpath instruments functions to show where time is spent, and flame produces flame graphs from manual instrumentation. For sampling CPU profiles, cargo-flamegraph and the system profilers perf, Instruments and samply work well with Rust as long as you keep debug symbols in release builds.