A script to automatically download and install OpenMM build dependencies, build OpenMM from source, then run the OpenMM benchmark suite and save the results.
Can run with:
curl -LOf https://raw.githubusercontent.com/epretti/openmm-benchmark/refs/heads/master/openmm-benchmark.sh && chmod +x openmm-benchmark.sh && ./openmm-benchmark.sh
Arguments are passed on from openmm-benchmark.sh to OpenMM's benchmark.py,
so, e.g., you can run a multi-GPU benchmark by replacing
./openmm-benchmark.sh with
./openmm-benchmark.sh "--test=amber20-cellulose,amber20-stmv --device=0,1,..."
(Use the actual device indices for the GPUs you want to benchmark.)
Arguments to pass should be quoted into a single argument; if multiple arguments
are given to the script, benchmark.py will be run multiple times. This can be
useful to perform multiple sets of benchmarks at once with different sets of
settings. If no arguments are given at all, benchmarks will be run once with
default options.
Note that currently openmm-benchmark.sh is CUDA-specific. This is not the
case for the underlying OpenMM benchmarks, but support for other platforms in
this repository is pending.
The script patches benchmark.py after cloning the OpenMM repository to add a
--deterministic-forces flag that enables the DeterministicForces platform
option.
Run analyze.py with JSON files output by the benchmark script as arguments to
summarize benchmark results.