cheshmi / sc23-ad-sparse-fusion

Sparse Fusion Benchmark

This directory contains scripts for reproducing results in the sparse fusion paper at SC23.

How to run the artifact

• The sympiler-bench repository should be cloned:

	git clone https://github.com/cheshmi/sc23-ad-sparse-fusion.git
	cd sc23-ad-sparse-fusion

• The singularity image should be copied to the same directory that the code is cloned. The *.sif is provided as part of the artifact ( see here).

You can test the image by running the following command from the current directory:

    singularity exec --env LD_LIBRARY_PATH=/opt/intel/oneapi/mkl/latest/lib/intel64 artifact.sif /source/fusion/build/demo/sptrsv_sptrsv_demo

The output is a set of comma separated values such as specifications for a random matrix and execution time of different tools for the matrix.

• You can run the following script and come back in half an hour:

bash run_sparse_fusion_ad.sh

Otherwise, you can follow below steps:

• The datasets should be downloaded by calling:

    bash dl_matrix_dataset.sh 

Matrices are downloaded into the mm directory in the current directory (This requires internet connection). Only 10 matrices are selected to speedup the evaluation time.

• The sparse fusion (Experiment 1) experiment can be executed by emitting:

	bash run_sfusion.sh

For running on compute node:

	sbatch run_sfusion.sh

You might need to update scripts with new absolute paths to the dataset and the image file. Also singularity module should be loaded for running on a server.

• The MKL evaluation (Experiment 2) can be done by running:

	bash run_mkl.sh

• The DAGP evaluation (Experiment 3) can be reproduced by calling:

bash run_dagp.sh

• The Gauss-Seidel case study (Experiment 4) can be reproduced by calling:

	bash run_gs.sh

• Upon successful completion of experiments, all results should be stored as comma separated values (CSV) files under the ./logs/ directory and are ready to be plotted. You can call:

bash plot.sh

to create plots. Plots are stored in the current directory as PDF files.