To build from source, these packages are necessary (Debian and derivatives):

sudo apt install git gcc g++ python3 python3-dev python3-pip python3-build python3-venv nasm

Then, the latest unreleased development version can be built with:

python3 -m pip install --force-reinstall 'git+https://github.com/mxmlnkn/rapidgzip.git@main#egginfo=rapidgzip&subdirectory=python/rapidgzip'

And to build locally, you can use build and install the wheel:

git clone --recursive https://github.com/mxmlnkn/rapidgzip.git
cd rapidgzip/python/rapidgzip/
rm -rf dist
python3 -m build
python3 -m pip install --force-reinstall --user --break-system-packages dist/rapidgzip-*.whl

Decompression of Gzip-Compressed Base64 Data

This benchmark uses random data, that has been base64 encoded and then gzip-compressed. This is the next best case for rapidgzip after the trivial case of purely random data, which cannot be compressed and therefore can be decompressed with a simple memory copy. This next best case results in mostly Huffman-coding compressed data with only very few LZ77 back-references. Without LZ77 back-references, parallel decompression can be done more independently and therefore faster than in the case of many LZ77 back-references.

Decompression of Gzip-Compressed FASTQ Data

This benchmarks uses gzip-compressed FASTQ data. That's why the TAR file is repeated as often as there are number of cores in the benchmark to hold the decompression times roughly constant in order to make the benchmark over this large a range feasible. This is almost the worst case for rapidgzip because it contains many LZ77 back-references over very long ranges. This means that a fallback to ISA-L is not possible and it means that the costly two-staged decoding has to be done for almost all the data. This is also the reason why it fails to scale above 64 cores, i.e, to the second CPU socket. The first and second decompression stages are completely independently submitted to a thread pool, which on this NUMA architecture means, that data needs to be costly transferred from one processor socket to the other if the second step for a chunk is not done on the same processor as the first. This should be fixable by making the ThreadPool NUMA-aware.

These three scaling plots were created with rapidgzip 0.9.0, while the ones in the paper were created with 0.5.0.

Scaling Benchmarks on Ryzen 3900X

These benchmarks on my local workstation with a Ryzen 3900X that only has 12 cores (24 virtual cores), but the base frequency is much higher than the 2xAMD EPYC 7702 CPU.

Decompression With Existing Index

Decompression From Scratch

Benchmarks for Different Compressors

This benchmark compresses the enlarged Silesia TAR with different gzip implementations, each with different compression levels. Rapidgzip is then used to decompress the resulting files with 128 cores.

Rapidgzip can parallelize decompression for almost all tested cases. The only exception is files compressed with igzip -0, because these files contain only a single several gigabytes large deflate block. This is the only known tool to produce such a pathological deflate block.

The decompression bandwidth for the other compressors, varies quite a lot. The fastest decompression is reached with 22 GB/s for bgzip-compressed files because the bgzip format is directly supported, which enables rapidgzip to avoid the two-stage decompression method and also enables rapidgzip to offload all of the work to ISA-L. Files compressed with bgzip -l 0 decompress slightly slower with "only" 18 GB/s, because it creates a fully non-compressed gzip stream and therefore is more I/O bound than the other bgzip-generated files.

Decompression of pigz-generated files is the slowest with 6 GB/s as opposed to 10-14 GB/s for gzip and igzip. It is not clear why that is. It might be because pigz generates small deflate blocks and adds flush markers.

The values in this chart are higher than in table 3 in the paper because the measurements were done with rapidgzip 0.10.1 instead of version 0.5.0.

Benchmarks for Different Decompressors

These benchmarks use different compressors and different decompressors to show multiple things:

• Single-core decompression of rapidgzip is close to igzip and roughly twice as fast as bgzip, which uses zlib.
• Decompression bandwidth with ISA-L can somewhat compete with zstd and is only 25% slower.
• Both bgzip and pzstd can only parallelize decompression of files compressed with bgzip / pzstd. This especially means that files compressed with the standard zstd tool cannot be decompressed in parallel and tops out at ~800 MB/s.
• Even for bgzip-compressed files, rapidgzip is always faster than bgzip for decompression, thanks to ISA-L and better multi-threading.
• Rapidgzip scales higher than pzstd for decompression with many cores, and can be more than twice as fast when an index exists: 24.3 GB/s vs. 9.5 GB/s.

The values in this chart are higher than in table 4 in the paper because the measurements were done with rapidgzip 0.10.1 instead of version 0.5.0.

A gzip decompressor tool based on the rapidgzip backend from ratarmount
Usage:
  rapidgzip [OPTION...] positional parameters

 Actions options:
  -d, --decompress        Force decompression. Only for compatibility. No
                          compression supported anyways.
      --import-index arg  Uses an existing gzip index.
      --export-index arg  Write out a gzip index file.
      --count             Prints the decompressed size.
  -l, --count-lines       Prints the number of newline characters in the
                          decompressed data.
      --analyze           Print output about the internal file format
                          structure like the block types.

 Advanced options:
      --chunk-size arg      The chunk size decoded by the parallel workers
                            in KiB. (default: 4096)
      --verify              Verify CRC32 checksum. Will slow down
                            decompression and there are already some
                            implicit and explicit checks like whether the
                            end of the file could be reached and whether
                            the stream size is correct.
      --no-verify           Do not verify CRC32 checksum. Might speed up
                            decompression and there are already some
                            implicit and explicit checks like whether the
                            end of the file could be reached and whether
                            the stream size is correct.
      --io-read-method arg  Option to force a certain I/O method for
                            reading. By default, pread will be used when
                            possible. Possible values: pread, sequential,
                            locked-read (default: pread)
      --index-format arg    Option to select an output index format.
                            Possible values: gztool, gztool-with-lines,
                            indexed_gzip. (default: indexed_gzip)

 Decompression options:
  -c, --stdout                  Output to standard output. This is the
                                default, when reading from standard input.
  -f, --force                   Force overwriting existing output files.
                                Also forces decompression even when piped
                                to /dev/null.
  -o, --output arg              Output file. If none is given, use the
                                input file name with '.gz' stripped or
                                '<input file>.out'. If no input is read
                                from standard input and not output file is
                                given, then will write to standard output.
  -k, --keep                    Keep (do not delete) input file. Only for
                                compatibility. This tool will not delete
                                anything automatically!
  -P, --decoder-parallelism arg
                                Use the parallel decoder. If an optional
                                integer >= 1 is given, then that is the
                                number of decoder threads to use. Note that
                                there might be further threads being
                                started with non-decoding work. If 0 is
                                given, then the parallelism will be
                                determined automatically. (default: 0)
      --ranges arg              Decompress only the specified byte ranges.
                                Example: 10@0,1KiB@15KiB,5L@20L to
                                decompress the first 10 bytes, 1024 bytes
                                at offset 15 KiB, as well as the 5 lines
                                after skipping the first 20 lines.

 Output options:
  -h, --help                   Print this help message.
  -q, --quiet                  Suppress noncritical error messages.
  -v, --verbose                Print debug output and profiling statistics.
  -V, --version                Display software version.
      --oss-attributions       Display open-source software licenses.
      --oss-attributions-yaml  Display open-source software licenses in
                               YAML format for use with Conda.

If no file names are given, rapidgzip decompresses from standard input to standard output.
If the output is discarded by piping to /dev/null, then the actual decoding step might
be omitted if neither -l nor -L nor --force are given.

Examples:

Decompress a file:
  rapidgzip -d file.gz

Decompress a file in parallel:
  rapidgzip -d -P 0 file.gz

List information about all gzip streams and deflate blocks:
  rapidgzip --analyze file.gz

sudo apt-get install libopenmpi-dev openmpi-bin gcc-11-plugin-dev llvm-dev libclang-dev libunwind-dev \
                     libopen-trace-format-dev otf-trace libpapi-dev

# Install Score-P (to /opt/scorep)
SCOREP_VERSION=8.0
wget "https://perftools.pages.jsc.fz-juelich.de/cicd/scorep/tags/scorep-${SCOREP_VERSION}/scorep-${SCOREP_VERSION}.tar.gz"
tar -xf "scorep-${SCOREP_VERSION}.tar.gz"
cd "scorep-${SCOREP_VERSION}"
./configure --with-mpi=openmpi --enable-shared --without-llvm --without-shmem --without-cubelib --prefix="/opt/scorep-${SCOREP_VERSION}"
make -j $( nproc )
make install

# Add /opt/scorep to your path variables on shell start
cat <<EOF >> ~/.bashrc
if test -d /opt/scorep; then
    export SCOREP_ROOT=/opt/scorep
    export PATH=$SCOREP_ROOT/bin:$PATH
    export LD_LIBRARY_PATH=$SCOREP_ROOT/lib:$LD_LIBRARY_PATH
fi
EOF

echo -1 | sudo tee /proc/sys/kernel/perf_event_paranoid

# Check whether it works
scorep --version
scorep-info config-summary