Raw benchmark results as used in the paper

Throughput / ratio averages

Note that if you try reproducing the "contrib" csvs from scratch, the missing decompressor for cudpp-compress will break the plot script, so you have to fake decompressor timings with something like "sed s/;;/;99999;/" (these are of course not reported in the paper)

For MPC, GFC, nvCOMP LZ4, nvCOMP Cascaded and cudppCompress:

• bench-2070-contrib.csv
• bench-3090-contrib.csv
• bench-a100-contrib.csv
• bench-v100-contrib.csv

For ndzip-gpu (second number in the file name is the thread block size; we select 256 for single-precision and 512 for double-precision on all systems)

• bench-2070-ndzip-gpu-128.csv
• bench-2070-ndzip-gpu-256.csv
• bench-2070-ndzip-gpu-384.csv
• bench-2070-ndzip-gpu-512.csv
• bench-2070-ndzip-gpu-640.csv
• bench-2070-ndzip-gpu-768.csv
• bench-3090-ndzip-gpu-128.csv
• bench-3090-ndzip-gpu-256.csv
• bench-3090-ndzip-gpu-384.csv
• bench-3090-ndzip-gpu-512.csv
• bench-3090-ndzip-gpu-640.csv
• bench-3090-ndzip-gpu-768.csv
• bench-a100-ndzip-gpu-128.csv
• bench-a100-ndzip-gpu-256.csv
• bench-a100-ndzip-gpu-384.csv
• bench-a100-ndzip-gpu-512.csv
• bench-a100-ndzip-gpu-640.csv
• bench-a100-ndzip-gpu-768.csv
• bench-v100-ndzip-gpu-128.csv
• bench-v100-ndzip-gpu-256.csv
• bench-v100-ndzip-gpu-384.csv
• bench-v100-ndzip-gpu-512.csv
• bench-v100-ndzip-gpu-640.csv
• bench-v100-ndzip-gpu-768.csv

For CPU benchmarks on the Ryzen 3900X (RTX 2070) system

• bench-2070-cpu.csv

Invoke the plot script separately for each system and CSV file like so:

    python3 src/benchmark/plot_benchmark.py bench-2070-ndzip-gpu-256.csv  # For single-precision on RTX 2070
    python3 src/benchmark/plot_benchmark.py bench-2070-ndzip-gpu-512.csv  # For double-precision on RTX 2070
    python3 src/benchmark/plot_benchmark.py bench-2070-contrib.csv        # For all GPU 3rd-party compressors
    python3 src/benchmark/plot_benchmark.py bench-2070-cpu.csv            # For CPU reference compressors

Compression ratio / throughput per dataset

• bench-v100-optimal.csv

• python3 ratiotable.py bench-v100-optimal.csv

    \documentclass[a4paper,8pt]{scrartcl}
    \usepackage{booktabs}
    \usepackage{graphicx}
    \usepackage{tabularx}
    \newcommand\algohead[1]{\rotatebox[origin=l]{90}{\hspace{-0.8ex}\ttfamily\small#1\hspace{0.5ex}}}
    \begin{document}
    \hspace{-20em}
    \renewcommand\arraystretch{1.1}
    \setlength\tabcolsep{0.45ex}
    ...
    \end{document}

Thread block size effect

• python3 group_size_stats.py bench-*-ndzip-gpu-*.csv

Average kernel runtimes

• profile-v100.log

• python3 kernel_times.py