perf tutorial

This page shows contributors how to benchmark the @mutable kernels with moon bench and how to read the result. The method is in the perf design; the full pipeline is in bench/README.md.

Quick start

moon bench -p perf --target native --release

Each line of the output is one case, named <operation>/<case id>, with the time per call.

Everyday tasks

Compare a kernel change

1. Run the benchmark on the unchanged code and keep the output.
2. Apply the change, run again, and compare the same case names.
3. Check that the result bits did not change with the perf_support tutorial checksum comparison, or that they changed only by rounding.

Run the full report

just bench

This builds the benchmarks in release mode, runs them, and writes bench/results/summary.md and summary.json with median, p90 and MAD per case.

Going further

Add the Rust nalgebra baseline with BENCH_FLAGS="--include-rust" just bench, and open the local dashboard with just bench-web.

Common pitfalls

• Benchmarking debug builds. Always pass --release.
• Comparing runs from different machines. Compare on one machine, under similar load.

Next steps

• perf_runner tutorial for raw samples of one case.
• perf_support API.