Apache-2.0 licensed and maintained by Erik Schnetter

Module documentation for

This version can be pinned in stack with:mpi-hs-,15234


MPI bindings for Haskell


MPI (the Message Passing Interface) is a widely used standard for distributed-memory programming on HPC (High Performance Computing) systems. MPI allows exchanging data (messages) between programs running in parallel. There are several high-quality open source MPI implementations (e.g. MPICH, MVAPICH, OpenMPI) as well as a variety of closed-source implementations. These libraries can typically make use of high-bandwidth low-latency communication hardware such as InfiniBand.

This library mpi-hs provides Haskell bindings for MPI. It is based on ideas taken from haskell-mpi, Boost.MPI for C++, and MPI for Python.

mpi-hs provides two API levels: A low-level API gives rather direct access to the actual MPI API, apart from certain “reasonable” mappings from C to Haskell (e.g. output arguments that are in C stored via a pointer are in Haskell regular return values). A high-level API simplifies exchanging arbitrary values that can be serialized.


This is a typical MPI C code:

#include <stdio.h>
#include <mpi.h>

int main(int argc, char** argv) {
  MPI_Init(&argc, &argv);
  int rank, size;
  MPI_Comm_rank(MPI_COMM_WORLD, &rank);
  MPI_Comm_size(MPI_COMM_WORLD, &size);
  printf("This is process %d of %d\n", rank, size);
  int msg = rank;
  MPI_Bcast(&msg, 1, MPI_INT, 0, MPI_COMM_WORLD):
  printf("Process %d says hi\n", msg);
  return 0;

The Haskell equivalent looks like this:

{-# LANGUAGE TypeApplications #-}

import qualified Control.Distributed.MPI as MPI
import Foreign
import Foreign.C.Types

main :: IO ()
main = do
  rank <- MPI.commRank MPI.commWorld
  size <- MPI.commSize MPI.commWorld
  putStrLn $ "This is process " ++ show rank ++ " of " ++ show size
  let msg = MPI.fromRank rank
  buf <- mallocForeignPtr @CInt
  withForeignPtr buf $ \ptr -> poke ptr msg
  MPI.bcast (buf, 1::Int) MPI.rootRank MPI.commWorld
  msg' <- withForeignPtr buf peek
  putStrLn $ "Process " ++ show msg' ++ " says hi"

The high-level API simplifies exchanging data; no need to allocate a buffer nor to use poke or peek:

{-# LANGUAGE TypeApplications #-}

import qualified Control.Distributed.MPI as MPI
import qualified Control.Distributed.MPI.Storable as MPI

main :: IO ()
main = MPI.mainMPI $ do
  rank <- MPI.commRank MPI.commWorld
  size <- MPI.commSize MPI.commWorld
  putStrLn $ "This is process " ++ show rank ++ " of " ++ show size
  let msg = MPI.fromRank rank :: Int
  msg' <- MPI.bcast (Just msg) MPI.rootRank MPI.commWorld
  putStrLn $ "Process " ++ show msg' ++ " says hi"


mpi-hs requires an external MPI library to be available on the system. How to install such a library is beyond the scope of these instructions.

In some cases, the MPI library will be installed in /usr/include, /usr/lib, and /usr/bin, respectively. In this case, no further configuration is necessary, and mpi-hs will build out of the box with stack build.

For convenience, this package offers Cabal flags to handle several common cases where the MPI library is not installed in a standard location:

  • OpenMPI on Debian Linux (package libopenmpi-dev): use --flag mpi-hs:openmpi-debian
  • OpenMPI on Ubuntu Linux (package libopenmpi-dev): use --flag mpi-hs:openmpi-ubuntu
  • OpenMPI on macOS with MacPorts (package openmpi): use --flag mpi-hs:openmpi-macports
  • MPICH on Debian Linux (package libmpich-dev): use --flag mpi-hs:mpich-debian
  • MPICH on Ubuntu Linux (package libmpich-dev): use --flag mpi-hs:mpich-ubuntu
  • MPICH on macOS with MacPorts (package mpich): use --flag mpi-hs:mpich-macports

For example, using the MPICH MPI implementation installed via MacPorts on macOS, you build mpi-hs like this:

stack build --flag mpi-hs:mpich-macports

In the general case, if your MPI library/operating system combination is not supported by a Cabal flag, you need to describe the location of your MPI library in stack.yaml. This might look like:

  - /usr/lib/openmpi/include
  - /usr/lib/openmpi/lib

To remove default systems paths from the search paths, disable the system-mpi flag, which is enable by default: --flag mpi-hs:-system-mpi.

Testing the MPI installation with a C program

To test your MPI installation independently of using Haskell, copy the example MPI C code into a file mpi-example.c, and run these commands:

mpicc -c mpi-example.c
mpicc -o mpi-example mpi-example.o
mpiexec -n 3 ./mpi-example

All three commands must complete without error, and the last command must output something like

This is process 0 of 3
This is process 1 of 3
This is process 2 of 3

where the lines will be output in a random order. (The output might even be jumbled, i.e. the characters in these three lines might be mixed up.)

If these commands do not work, then this needs to be corrected before mpi-hs can work. If additional compiler options or libraries are needed, then these need to be added to the stack.yaml configuration file (for include and library paths; see extra-include-dirs and extra-lib-dirs there) or the package.yaml configuration file (for additional libraries; see extra-libraries there).

Examples and Tests

Running the example

To run the example provided in the src directory:

stack build
mpiexec stack exec version
mpiexec -n 3 stack exec example1
mpiexec -n 3 stack exec example2

Running the tests

There are two test cases provided in tests. The first (mpi-test) tests the low-level API, the second (mpi-test-storable) tests the high-level API:

stack build --test --no-run-tests
mpiexec -n 3 stack exec -- $(stack path --dist-dir)/build/mpi-test/mpi-test
mpiexec -n 3 stack exec -- $(stack path --dist-dir)/build/mpi-test-storable/mpi-test-storable

Related packages

There are three companion packages that provide the same high-level API via different serialization packages. These are separate packages to reduce the number of dependencies of mpi-hs:

Used by 3 packages in nightly-2020-09-16(full list with versions):