Profiling givp

When investigating performance regressions or proposing optimizations to the hot path (_neighborhood_*, local_search_vnd, _run_grasp_loop), please attach a profile so reviewers can verify the impact.

Quick CPU profile with py-spy

py-spy is a sampling profiler that attaches to a running Python process without instrumentation overhead.

pip install py-spy
py-spy record -o profile.svg -- \
  python -m pytest -m performance tests/benchmark/test_performance.py --benchmark-only -k "sphere and 30"

Open profile.svg in a browser to inspect the flamegraph.

Line-level profile with scalene

scalene reports CPU and memory usage per line, which is useful for spotting accidental allocations in the hot path.

pip install scalene
scalene --html --outfile scalene.html benchmarks/runner.py

Reproducible benchmark runs

The repository ships an opt-in pytest-benchmark module at tests/benchmark/test_performance.py. Use --benchmark-autosave to compare runs over time:

pytest -m performance tests/benchmark/test_performance.py --benchmark-only --benchmark-autosave
pytest-benchmark compare 0001 0002 --columns=mean,stddev,ops

Pin the master RNG via the public seed= parameter to make timing comparisons deterministic across runs.

Julia benchmarks

The Julia port includes a BenchmarkTools.jl-based benchmark suite under julia/benchmarks/. It covers four classic test functions (sphere, rosenbrock, rastrigin, ackley) at dimensions 5 and 10:

cd julia
julia --project=. benchmarks/benchmarks.jl

Results are saved to julia/benchmarks/results.json. On subsequent runs, the script automatically compares against the previous results and flags regressions (>10% time increase).