benchmarking.md

Benchmarking

routing/benches/fib_lookup.rs (Criterion) and fib_lookup_callgrind.rs (iai-callgrind) use the same fixtures and lookup code.

just profile=release bench  # wall-clock time
just bench-callgrind        # instructions and modelled cache traffic

The dev shell provides valgrind and iai-callgrind-runner. The runner version must match the iai-callgrind crate; nix/overlays/dataplane-dev.nix reads it from the workspace Cargo.toml.

Interpreting results

Criterion measures elapsed time on the host; run it on a quiet machine. Callgrind counts instructions and modelled cache events, which help compare changes in work but do not predict elapsed time. Fewer instructions can still mean slower execution after a data-layout change. "Estimated Cycles" is derived from counters, not a measured cycle count.

Callgrind scopes measurements to the benchmark function. Cachegrind counts the whole process, so its totals are not directly comparable. The cache geometry in fib_lookup_callgrind.rs is a fixed reference configuration, not a description of the current host.

Valgrind can change DPDK's CPU dispatch. Confirm which implementation ran before comparing its results with native measurements.

Fixtures

The shared fixture asserts that the packet matches the installed prefix. This prevents accidentally benchmarking a miss or the default route.

Callgrind counts argument destruction inside the measured function. The fixture is leaked during setup so its teardown is excluded from the measurement.