Skip to content

REPRODUCIBILITY.md

Latest commit

 

History

History
52 lines (37 loc) · 1.62 KB

REPRODUCIBILITY.md

File metadata and controls

52 lines (37 loc) · 1.62 KB
Raw

Reproducibility Protocol

This document defines a minimal reproducibility baseline for performance and validation claims.

Environment Capture

Record the following in every benchmark or validation report:

  • Git commit hash
  • GPU model and compute capability
  • Driver version (nvidia-smi)
  • CUDA toolkit version (nvcc --version)
  • Host CPU, RAM, and OS
  • CMake preset and build type

Determinism and Seeds

  • N-body initialization uses fixed seed 42 by default.
  • Any new randomized initialization must expose a seed parameter.
  • Reports must state the seed used.

Baseline Commands

conan install . --build=missing -s build_type=Release -s compiler.cppstd=17 --output-folder=build/conan
cmake -B build/release -G Ninja -DCMAKE_BUILD_TYPE=Release -DCMAKE_TOOLCHAIN_FILE=build/conan/conan_toolchain.cmake
cmake --build build/release --parallel
ctest --test-dir build/release --output-on-failure
./build/release/bin/nbody_bench --benchmark_format=json --benchmark_out=nbody_results.json
./build/release/bin/sph_bench --benchmark_format=json --benchmark_out=sph_results.json

Validation Thresholds

Use the acceptance criteria in README.md:

  • N-body energy drift below 0.1% over 1000 steps
  • Plummer sphere profile statistical pass criteria
  • Fused vs two-pass agreement below 1e-3 per particle

Reporting Template

Include this table in every benchmark report:

Metric Value Baseline Delta
N-body steps/s (10K)
N-body steps/s (32K)
SPH steps/s (30K)
Energy drift (%)

If a metric regresses by more than 5%, attach Nsight evidence and root-cause analysis.