MPI for Python

4.1.1 · active · verified Sat Apr 11

mpi4py provides Python bindings for the Message Passing Interface (MPI) standard, enabling parallel programming with Python. It allows Python applications to leverage high-performance distributed memory architectures. The current version is 4.1.1, and it typically sees a few minor and patch releases per year, with major feature updates (like MPI standard support) in new minor versions.

Warnings

breaking Python 2.x support was dropped after mpi4py version 3.1.6. Versions 4.x and newer require Python 3.8+.
Fix: Upgrade to Python 3.8 or newer to use mpi4py versions 4.0.0+.
breaking Version 4.0.0 introduced support for the MPI-4.0 standard, including new features like sessions, persistent collectives, and partitioned communication. While mostly backward compatible, existing code might need updates to leverage new functionalities or if it relied on specific MPI-3 behaviors that changed in MPI-4.
Fix: Consult the official mpi4py documentation and MPI-4.0 standard for details on new features and potential API changes if migrating from older MPI standards.
gotcha mpi4py requires an underlying MPI implementation (e.g., Open MPI, MPICH, Intel MPI) to be installed on the system. It is a Python binding, not a standalone MPI library.
Fix: Ensure a compatible MPI library (and its development headers if building from source) is installed on your system before installing and using mpi4py.
gotcha MPI programs, including those written with mpi4py, must be executed using an MPI launcher like `mpirun` or `mpiexec`, not simply `python your_script.py`.
Fix: Run your mpi4py script with `mpirun -np N python your_script.py`, replacing `N` with the desired number of processes.
gotcha When installing mpi4py from source (e.g., if pre-built wheels are not available for your system/MPI configuration), you typically need MPI development headers installed (e.g., `libopenmpi-dev` or `mpich-dev` on Debian/Ubuntu, `openmpi-devel` or `mpich-devel` on CentOS/Fedora).
Fix: Install the appropriate MPI development package for your system's MPI distribution.
gotcha Pre-built `mpi4py` wheels available on PyPI are built against specific MPI Application Binary Interfaces (ABIs), typically MPICH or Open MPI. If your system's MPI implementation is ABI-incompatible with the wheel, you may encounter runtime errors and will need to build `mpi4py` from source against your specific MPI library.
Fix: Check `mpi4py` documentation for wheel ABI compatibility. If an incompatibility exists, build `mpi4py` from source (`pip install mpi4py --no-binary :all:`) after ensuring MPI development headers are present.

Install

pip install mpi4py Install mpi4py

Imports

MPI
```
from mpi4py import MPI
```
The 'MPI' object provides access to all MPI functions, types, and constants (e.g., MPI.COMM_WORLD, MPI.INT). Do not confuse 'mpi4py' (the package) with 'MPI' (the imported object).

Quickstart

This 'Hello, World!' example demonstrates how to initialize MPI, get the rank of the current process, and the total number of processes in the communicator. Save it as `hello.py` and run with `mpirun -np 4 python hello.py` for 4 processes.

from mpi4py import MPI
import os

# Get the default communicator (all processes)
comm = MPI.COMM_WORLD

# Get the rank of the current process
rank = comm.Get_rank()

# Get the total number of processes
size = comm.Get_size()

# Each process prints its rank and size
print(f"Hello from process {rank} of {size} on host {os.uname().nodename}")

# Use a barrier to ensure all processes reach this point before exiting
comm.Barrier()

view raw JSON →