ndzip: A High-Throughput Parallel Lossless Compressor for Scientific Data

ndzip compresses and decompresses multidimensional univariate grids of single- and double-precision IEEE 754 floating-point data. We implement

• a single-threaded CPU compressor
• an OpenMP-backed multi-threaded compressor
• a SYCL-based GPU compressor (currently hipSYCL + NVIDIA only)
• a CUDA-based GPU compressor (experimental)

All variants generate and decode bit-identical compressed stream.

ndzip is currently a research project with the primary use case of speeding up distributed HPC applications by increasing effective interconnect bandwidth.

Prerequisites

• CMake >= 3.15
• Clang >= 10.0.0
• Linux (tested on x86_64 and POWER9)
• Boost >= 1.66
• Catch2 >= 2.13.3 (optional, for unit tests and microbenchmarks)

Additionaly, for GPU support

• CUDA >= 11.0 (not officially compatible with Clang 10/11, but a lower version will optimize insufficiently!)
• An Nvidia GPU with Compute Capability >= 6.1
• For the SYCL version: A recent version of hipSYCL profiling functionality

Building

Make sure to set the right build type and enable the full instruction set of the target CPU architecture:

-DCMAKE_BUILD_TYPE=Release -DCMAKE_CXX_FLAGS="-march=native"

If unit tests and microbenchmarks should also be built, add

-DNDZIP_BUILD_TEST=YES

For GPU support with SYCL

1. Build and install hipSYCL

git clone https://github.com/illuhad/hipSYCL
cd hipSYCL
cmake -B build -DCMAKE_INSTALL_PREFIX=../hipSYCL-install -DWITH_CUDA_BACKEND=YES -DCMAKE_BUILD_TYPE=Release
cmake --build build --target install -j

2. Build ndzip with SYCL

cmake -B build -DCMAKE_PREFIX_PATH='../hipSYCL-install/lib/cmake' -DHIPSYCL_PLATFORM=cuda -DCMAKE_CUDA_ARCHITECTURES=75 -DHIPSYCL_GPU_ARCH=sm_75 -DCMAKE_BUILD_TYPE=Release -DCMAKE_CXX_FLAGS="-U__FLOAT128__ -U__SIZEOF_FLOAT128__ -march=native"
cmake --build build -j

Replace sm_75 and 75 with the string matching your GPU's Compute Capability. The -U__FLOAT128__ define is required due to an open bug in Clang.

For GPU support with CUDA (experimental)

a) Either build ndzip with CUDA + NVCC ...

cmake -B build -DCMAKE_CUDA_ARCHITECTURES=75 -DCMAKE_BUILD_TYPE=Release -DCMAKE_CXX_FLAGS="-march=native"
cmake --build build -j

Replace sm_75 and 75 with the string matching your GPU's Compute Capability.

b) ... or with CUDA + Clang

cmake -B build -DCMAKE_CUDA_COMPILER="$(which clang++)" -DCMAKE_CUDA_ARCHITECTURES=75 -DCMAKE_BUILD_TYPE=Release -DCMAKE_CXX_FLAGS="-U__FLOAT128__ -U__SIZEOF_FLOAT128__ -march=native"
cmake --build build -j

The -U__FLOAT128__ define is required due to an open bug in Clang.

Compressing and decompressing files

build/compress -n <size> -i <uncompressed-file> -o <compressed-file> [-t float|double]
build/compress -d -n <size> -i <compressed-file> -o <decompressed-file> [-t float|double]

<size> are one to three arguments depending on the dimensionality of the input grid. In the multi-dimensional case, the first number specifies the width of the slowest-iterating dimension.

By default, compress uses the single-threaded CPU compressor. Passing -e cpu-mt or -e sycl / -e cuda selects the multi-threaded CPU compressor or the GPU compressor if available, respectively.

Running unit tests

Only available if tests have been enabled during build.

build/encoder_test
build/sycl_bits_test  # only if built with SYCL support
build/sycl_ubench     # GPU microbenchmarks, only if built with SYCL support
build/cuda_bits_test  # only if built with CUDA support

See also

• Benchmarking ndzip

References

If you are using ndzip as part of your research, we kindly ask you to cite

• Fabian Knorr, Peter Thoman, and Thomas Fahringer. "ndzip: A High-Throughput Parallel Lossless Compressor for Scientific Data". In 2021 Data Compression Conference (DCC), IEEE, 2021. [DOI] Preprint PDF

• Knorr, Fabian, Peter Thoman, and Thomas Fahringer. "ndzip-gpu: efficient lossless compression of scientific floating-point data on GPUs". In SC'21: Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis, ACM, 2021. [DOI] [Preprint PDF]