Benchmarking FastSASA

FastSASA provides a reproducible benchmark command for structures and trajectories. It records timing, SASA results, input size, backend, precision, CPU, GPU, operating system, and the exact command in CSV form.

The repository contains a small structure manifest, not the structure files themselves. Public structures are downloaded from RCSB when requested. Downloaded inputs and benchmark results are ignored by git.

Build

cmake -S . -B build -DFASTSASA_BUILD_NATIVE_TESTS=ON
cmake --build build -j4

Standard Benchmark

Fetch the standard structures and run the same matrix used by other testers:

python3 tools/fetch_benchmark_corpus.py \
  --output-dir benchmark_corpus/structures

python3 tools/fastsasa_benchmark.py standard \
  --fastsasa build/fastsasa \
  --profile standard \
  --output-dir profiles/standard_benchmark

The main result is:

profiles/standard_benchmark/fastsasa_benchmark_results.csv

The output directory also contains the input manifest and benchmark_run.json with system and command metadata.

Include Public Trajectories

The standard structure run does not download trajectories. Add the trajectory option when you want end-to-end frame throughput:

python3 tools/fastsasa_benchmark.py standard \
  --fastsasa build/fastsasa \
  --profile standard \
  --fetch-standard-trajectories \
  --trajectory-frames 100 \
  --trajectory-batches '1 8 32' \
  --output-dir profiles/standard_with_trajectory

This fetches one representative trajectory from each configured public record, not every file in each archive. The standard subset covers both PDB/XTC and PSF/DCD workflows.

Benchmark Your Own Trajectory

Use the suite mode with name|topology|trajectory:

python3 tools/fastsasa_benchmark.py suite \
  --fastsasa build/fastsasa \
  --backend auto \
  --precision fp64 \
  --trajectory-selection protein \
  --trajectory-batches '1 8 32' \
  --trajectory-frames 100 \
  --trajectories 'run1|topology.psf|trajectory.dcd' \
  --output profiles/run1.csv

Trajectory benchmarks require an atom policy. Use protein for a normal protein-only benchmark, all for every topology atom, or a FastSASA selection expression for another system definition.

Structure Matrix

Use corpus mode to choose algorithms, resolutions, backend, and precision:

python3 tools/fastsasa_benchmark.py corpus \
  --fastsasa build/fastsasa \
  --structure-dir benchmark_corpus/structures \
  --backend vulkan \
  --precision fp32 \
  --points '100 500' \
  --slices '10 20' \
  --output profiles/structures_vulkan_fp32.csv

Use --include-nondefault for the larger optional structures. Run python3 tools/fastsasa_benchmark.py MODE --help for the complete options for standard, suite, or corpus.

Precision Reports

After collecting several structure or trajectory CSVs:

python3 tools/fastsasa_precision_report.py \
  --input-dir profiles --summary profiles/structure_precision.csv

python3 tools/fastsasa_trajectory_precision_report.py \
  --input-dir profiles \
  --detail profiles/trajectory_precision_detail.csv \
  --summary profiles/trajectory_precision.csv

FastSASA benchmark tools measure FastSASA itself. Comparisons with other tools belong in a separate validation environment so those programs and their data do not become FastSASA package dependencies.