Keyboard shortcuts

Press or to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

Native benchmark results

Human-facing benchmark reports in this directory use Markdown. Raw TSV files preserve every experiment and are the authoritative inputs for statistics.

WorkloadReportRaw samples
Coordinated retry, BiteOpt retry, and DE→CMA retry on GTOPbenchmark_gtop.mdbenchmark_gtop_100_raw.tsv, benchmark_gtop_tandem_100_raw.tsv, benchmark_biteopt_gtop_rust_100_raw.tsv, benchmark_de_cma_gtop_rust_100_raw.tsv
fcmaes versus independent Rust optimizer cratesoptimizer-comparison/comparison.mdraw artifacts
Core boundary: DE versus Nelder–Mead refinement and EGO at equal wall deadlinesoptimizer-boundary/results/decision-v2/comparison.mdrefiner-raw.tsv, bo-trace.tsv

Recreate the recorded native fcmaes workloads from the repository root:

cargo run --release -p fcmaes-examples --bin benchmark-gtop -- \
  --runs 100 --workers 32 --seed 1 \
  --raw-output benchmarks/benchmark_gtop_100_raw.tsv

python3 benchmarks/run_coordinated_tandem.py

cargo run --release -p fcmaes-examples --bin benchmark-biteopt-gtop -- \
  --algo biteopt --runs 100 --workers 24 --retries 24 \
  --evaluations 10000 --seed 1

cargo run --release -p fcmaes-examples --bin benchmark-biteopt-gtop -- \
  --algo de_cma --runs 100 --workers 24 --retries 24 \
  --evaluations 10000 --seed 1

The binaries print Markdown tables and accept --table-output PATH when a separate generated .md file is useful. The recorded slow Tandem run also includes benchmark_gtop_tandem_100_metadata.json with its exact configuration and total invocation time.

Run the dependency-isolated optimizer comparison with:

benchmarks/optimizer-comparison/run_all_external.sh

Reproduce the dependency-isolated optimizer-boundary experiment using the commands in its protocol. It keeps the experiment-only egobox-ego and Nelder–Mead implementations outside the root workspace and fcmaes-core dependency graph.

Wall times depend strongly on CPU, operating system, compiler version, and background load. Treat recorded timings as reproducibility data for the stated machine, not as universal performance guarantees.