Reproducing Results

Pre-generated CSVs and the performance_evaluation.ipynb notebook used to recreate every labeled figure in the paper live in plots/.

End-to-end run

  1. Update the dataset path in scripts/run.sh:
  2. Set DATASET_DIR to the directory containing the MM subdirectory (which itself holds per-matrix subdirectories of .mtx files). See Datasets for how to obtain it.

  3. Optionally change DATASET_FILES_NAME; it defaults to datasets/suitesparse.txt, the paper's matrix list.

  4. Kick off the run from the scripts/ directory:

bash cd scripts && ./run.sh

A complete run can take up to ~3 days — it sweeps the entire SuiteSparse collection four times, once per algorithm. The dominant cost is .mtx parsing from disk, not GPU time.

  1. Expect occasional matrix-level failures. Some files in SuiteSparse are mislabeled or otherwise malformed, and a few matrices are larger than the GPU can hold. These produce runtime exceptions or offset_t overflows and can be safely ignored — they don't affect the rest of the sweep.

  2. Run on a smaller subset by editing the stop condition in scripts/run.sh#L22 (it caps at the first 10 matrices by default), or removing the conditional in L22–L26 to process every .mtx file under DATASET_DIR.

Output

Each algorithm writes a CSV alongside run.sh. Drop these into plots/data/ to replace the bundled results, then open plots/performance_evaluation.ipynb (Jupyter — install with pip install jupyter or following the official instructions) and re-run all cells to regenerate the figures.

A representative line of CSV output:

kernel,dataset,rows,cols,nnzs,elapsed
merge-path,144,144649,144649,2148786,0.0720215
merge-path,08blocks,300,300,592,0.0170898
merge-path,1138_bus,1138,1138,4054,0.0200195