Home

Awesome

h5fortran-mpi

DOI

ci ci

Easy to use object-oriented Fortran parallel HDF5-MPI interface. The h5fortran-mpi API can be used with or with MPI. A very similar NetCDF4 interface is nc4fortran.

Many computer systems default to the serial HDF5 API, which lacks the HDF5 parallel MPI layer. The scripts/CMakeLists.txt can build the HDF5-MPI stack if needed. To use HDF5-MPI features, the computer must have a working MPI library installed already (e.g. OpenMPI, MPICH, Intel MPI, MS-MPI).

Some OS have an installable parallel HDF5 package:

While HDF5 1.10.2 is the oldest working HDF5 version, in general for bugfixes and performance HDF5 ≥ 1.10.5 is recommended. For highest performance with parallel compressed writes consider HDF5 ≥ 1.12.2.

Compressed parallel HDF5

Compression is useful in general to save significant disk space and speed up write/read. HDF5-MPI file compression requires HDF5 >= 1.10.2 and MPI-3. As noted above, HDF5 >= 1.10.5 is recommended for stability and performance.

Windows limitations

Microsoft Windows does not currently support native HDF5 parallel file compression. Windows Subsystem for Linux can be used for HDF5 parallel file compression. Native Windows users can read HDF5 compressed files but without using MPI.

Native Windows MPI options are currently limited to MS-MPI and Intel MPI. Currently Windows MS-MPI is MPI-2. A quirk with Intel oneAPI on Windows despite having MPI-3 is that with HDF5 1.10.x and at least through HDF5 1.12.1 the collective filtered parallel compression file I/O does not work. We test for this in CMake and set the compile flags appropriately.

Windows users that need file compression may use Windows Subsystem for Linux (e.g. Ubuntu) and install libhdf5-mpi-dev.

Build this project

Build this project like:

cmake -B build
cmake --build build

If you have previously built / installed a parallel HDF5 library, refer to it (saving build time) like:

cmake -B build -DHDF5_ROOT=~/lib_par
cmake --build build

To build without MPI (serial HDF5 file operations only):

cmake -B build -Dhdf5_parallel=off

Cray computers may use a CMake toolchain file to work with Intel or GCC backend.

Fortran Package Manager (FPM) users build like:

fpm build --flag -Dh5fortran_HAVE_PARALLEL
# omitting this flag builds the serial API only

fpm test

Notes

To build and install the HDF5 parallel library use the script:

cmake -B build_hdf5 -S scripts --install-prefix=$HOME/lib_par

cmake --build build_hdf5

that will build and install HDF5 under ~/lib_par (or other directory of your choice).