benchmarks

Backend benchmarks

The backend benchmark measures complete fixed-step RK4 updates for the infinite-plane model. Each step performs four direct all-pairs velocity evaluations. The timed interval starts only after backend construction, memory allocation, CUDA context creation/state upload, and warm-up steps. The final CUDA download and checksum are also outside the timed interval.

Configure and build all available benchmark executables:

cmake -S . -B build/benchmarks -DCMAKE_BUILD_TYPE=Release
cmake --build build/benchmarks --parallel --target \
  point_vortex_backend_benchmark_serial \
  point_vortex_backend_benchmark_cpu \
  point_vortex_backend_benchmark_mpi \
  point_vortex_backend_benchmark_cuda \
  point_vortex_backend_benchmark_cuda_mixed

Run five calibrated trials at each problem size and create the README-ready plot:

python3 benchmarks/run_benchmarks.py --overwrite
python3 benchmarks/plot_benchmarks.py

By default CPU/OpenMP uses one thread per detected physical core, MPI uses the same number of single-threaded ranks, CPU serial uses a library compiled without OpenMP, and both CUDA modes use one GPU. Override the parallel CPU choices with --cpu-threads and --mpi-ranks. Raw trials go to results.csv; machine/build metadata and the exact timing exclusions go to system.json.

The plot's timing panel reports absolute RK4 step time. Its speedup panel is normalized by the single-core CPU serial backend, which is the conventional baseline for total parallel speedup.

The checked-in plot extends the default sweep to larger populations with three five-second scout trials:

python3 benchmarks/run_benchmarks.py --resolutions 16384 32768 65536 --trials 3 \
  --target-seconds 5 --warmup-seconds 1 \
  --output benchmarks/results_large.csv \
  --metadata benchmarks/system_large.json --overwrite
python3 benchmarks/plot_benchmarks.py \
  --input benchmarks/results.csv benchmarks/results_large.csv \
          benchmarks/results_serial_mixed.csv benchmarks/results_serial_mixed_large.csv

For the sustained high-resolution check, use one-minute timed trials and a time-calibrated warm-up. This makes clock settling and thermal effects visible at the most expensive resolution:

python3 benchmarks/run_benchmarks.py --resolutions 65536 --trials 3 \
  --target-seconds 60 --warmup-seconds 5 \
  --output benchmarks/results_sustained_65536.csv \
  --metadata benchmarks/system_sustained_65536.json --overwrite

The repository also retains the corresponding one-minute check at N=8192 so the short and sustained measurements can be compared at both ends of the extended range. The serial CPU and mixed-CUDA trials at that size can be reproduced with:

python3 benchmarks/run_benchmarks.py --backends serial cuda_mixed \
  --resolutions 8192 --trials 3 --target-seconds 60 --warmup-seconds 5 \
  --output benchmarks/results_serial_mixed_sustained_8192.csv \
  --metadata benchmarks/system_serial_mixed_sustained_8192.json --overwrite

Serial CPU and mixed-CUDA sustained results are stored in the matching results_serial_mixed_sustained_8192.csv and results_serial_mixed_sustained_65536.csv files.