PSD Projection Benchmarks

Benchmarks for the psd_projection library.

Datasets

Datasets are generated using the Matrix Depot Julia package. The datasets are stored in the data directory, and can be generated using the data/generate.jl script. The datasets were generated in parallel, so you may not get the same results if you generate them yourself.

Generating the datasets

Note that this approach requires around 130 GB of disk space, can take a long time, and might not give the same results as the datasets we used in our experiments. We recommend using the pre-generated datasets instead.

To generate the datasets, you need to have Julia installed. Then, you can run the following command in the terminal:

datasets=("cauchy" "chebspec") # see run.sh for the full list of datasets
instance_sizes=(10 20) # we used 5000, 10000, 20000 in our experiments
julia data/generate.jl --datasets "${datasets[@]}" --instance_sizes "${instance_sizes[@]}"

Downloading the datasets

This part is coming soon.

Running the benchmarks

We provide several scripts to run the benchmarks:

• run.sh: runs the benchmarks using the psd_projection library; requires the library to cloned, as well as CMake;
• run_full.sh: runs the benchmarks by calling benchmark_full.cu, a self-contained CUDA program that does not require the psd_projection library;
• run_final_1.sh: same as run_full.sh, but runs the first half of the datasets;
• run_final_2.sh: same as run_full.sh, but runs the second half of the datasets.

We used run_final_1.sh and run_final_2.sh for our experiments. All scripts assume that the datasets are already generated (see above).

Results are stored in the results directory, in a CSV file named psd_results.csv. The scripts also generate an output.txt file with the output of the benchmark and useful debugging information.