Datasets

loops ships with a single bundled matrix, datasets/chesapeake/chesapeake.mtx, so all examples can be run as smoke tests right out of the box. For benchmarking and the paper's experiments we use the SuiteSparse Matrix Collection.[^1]

Downloading SuiteSparse

The full collection is large enough that a download takes a long time and a lot of disk; we recommend running the command below inside a tmux session.

wget --recursive --no-parent --force-directories -l inf -X RB,mat \
     --accept "*.tar.gz" \
     "https://suitesparse-collection-website.herokuapp.com/"

Flag	Purpose
`--recursive`	Recursively follow links
`--no-parent`	Don't fetch links above the starting URL
`-l inf`	No depth limit
`-X RB,mat`	Skip the `RB` and `mat` subdirs (we only want MatrixMarket)
`--accept "*.tar.gz"`	Only download tarballs
`--force-directories`	Preserve the original site hierarchy

Total downloaded size: ~887 GB (uncompressed + compressed).

Uncompressing

After the wget completes, decompress every tarball in place:

find . -name '*.tar.gz' -execdir tar -xzvf '{}' \;

Picking a subset

The list of matrices used by the paper's runs lives in datasets/suitesparse.txt. When invoking scripts/run.sh you can either:

point DATASET_FILES_NAME at this list to reproduce the paper's experiment set, or
supply your own newline-separated list to run on a custom subset.

[^1]: Timothy A. Davis and Yifan Hu. 2011. The University of Florida Sparse Matrix Collection. ACM Transactions on Mathematical Software 38, 1, Article 1 (December 2011), 25 pages. DOI: https://doi.org/10.1145/2049662.2049663.