WI4MPI is an MPI wrapper and interoperability library that enables seamless runtime conversions between various MPI implementations. Easily switch between MPI libraries without recompiling your applications!
Key Features:
-
🚀 Compile Once, Run Anywhere: Compile your application once using WI4MPI's MPI interface wrappers, and transparently execute it with different MPI implementations at runtime, including OpenMPI, IntelMPI, MPICH, MVAPICH, and MPC.
-
🔄 Dynamic MPI Conversion: Seamlessly switch between MPI implementations (e.g., OpenMPI ↔ IntelMPI) without recompiling.
-
🐳 Container-friendly: Easily replace containerized MPI stacks with optimized MPI libraries available on your HPC cluster.
-
🌐 Flexible HPC Integration: Ideal for configuring environment modules or Slurm scripts to maximize portability and performance across various HPC environments.
With WI4MPI, achieve seamless interoperability, optimized performance, and simplified management of MPI applications across diverse computing environments.
Check the full documentation here: https://wi4mpi.readthedocs.io/en/latest/
Two modes are provided:
Preload, for dynamic conversion:
Interface, for building an universal MPI code:
- GNU/Linux OS
- Mandatory: C and Fortran compilers
- Optional: Python 3 and Jinja 2
Python 3 is mandatory if you need to generate Wi4MPI code. As specified in the generator requirements, it depends on logging, colorlog & docopts Python packages.
WI4MPI is available on Spack:
$ spack install wi4mpi
Check the WI4MPI Spack tutorial here: Installalling Wi4MPI using Spack
Here is a quick introduction to compile WI4MPI with pre-generated sources (using wi4mpi-x.y.z.tar.gz archive
provided with releases, starting with Wi4MPI 4.0):
$ cd build
$ cmake -DCMAKE_INSTALL_PREFIX=/path-install/wi4mpi-%version ..
$ make -j 16
$ make install
For more details (and in particular if you need to regenerate the sources) please check the full documentation.
Wi4MPI documentation is built with Sphinx, choose your preferred format.
Example with html:
$ cd doc/
$ make html
$ firefox build/html/index.html
The full documentation in hosted by readthedocs: https://wi4mpi.readthedocs.io/en/latest/
Source the setup env bash file:
$ source share/wi4mpi/setup-env.sh
Run mpirun wrapper command:
$ mpirun -np 4 -F openmpi -T intelmpi ./mpi_hello.exe
You are using Wi4MPI-%version in preload mode from OMPI to INTEL
Hello_world 0
Hello_world 2
Hello_world 3
Hello_world 1
WI4MPI supports transparent MPI translation within containerized environments. This feature allows applications using MPI inside containers (e.g., Docker, Podman, Singularity) to leverage optimized MPI libraries provided by HPC centers, without modifying the container itself.
Currently, a detailed tutorial is available for using WI4MPI with OpenMPI and IntelMPI inside Podman containers:
If you need assistance or examples for other container technologies, feel free to reach out or contribute.
- vincent.ducrot.tgcc@cea.fr
- tony.delforge.tgcc@cea.fr
- adrien.cotte.tgcc@cea.fr
- clement.fontenaille.tgcc@cea.fr
- kevin.juilly.tgcc@cea.fr
- marc.joos@cea.fr
If you are referencing Wi4MPI in a publication, you can cite the following paper:
- E. A. León, M. Joos et al., "On-the-Fly, Robust Translation of MPI Libraries," 2021 IEEE International Conference on Cluster Computing (CLUSTER), 2021, pp. 504-515, doi: 10.1109/Cluster48925.2021.00026.
Copyright 2015-2022 CEA/DAM/DIF
WI4MPI is distributed under the BSD-3-Clause OR CeCILL-B.