mpi-hs

MPI bindings for Haskell

• GitHub: Source code repository
• Hackage: Haskell package and documentation
• CircleCI: Continuous integration

Overview

MPI (the Message Passing Interface) is 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, and MPI for Python.

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

Example

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);
  MPI_Finalize();
  return 0;
}

The Haskell equivalent looks like this:

import Control.Distributed.MPI as MPI

main :: IO ()
main =
  do MPI.init
     rank <- MPI.commRank MPI.commWorld
     size <- MPI.commSize MPI.commWorld
     putStrLn $ "This is process " ++ show rank ++ " of " ++ show size
     MPI.finalize

Installing

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 many 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.

On Ubuntu, one MPI package is openmpi-dev. It installs into /usr/lib/openmpi/include, /usr/lib/openmpi/lib, and /usr/bin/. You need to ensure that these settings are present in stack.yaml:

extra-include-dirs:
  - /usr/lib/openmpi/include
extra-lib-dirs:
  - /usr/lib/openmpi/lib

On MacOS, one MPI package is the MacPorts package openmpi. It installs into /opt/local/include/openmpi-mp, /opt/local/lib/openmpi-mp, and /opt/local/bin. You need to ensure that these settings are present in stack.yaml:

extra-include-dirs:
  - /opt/local/include/openmpi-mp
extra-lib-dirs:
  - /opt/local/lib/openmpi-mp

Both these settings are there by default.

Testing the MPI installation

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:

cc -I/usr/lib/openmpi/include -c mpi-example.c
cc -o mpi-example mpi-example.o -L/usr/lib/openmpi/lib -lmpi
mpirun -np 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 order in which the lines are printed can be random. (The output might even be jumbled, i.e. the characters of the 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 src/Main.hs:

stack build
mpirun -np 3 stack exec example && echo SUCCESS || echo FAILURE

With OpenMPI, and when running on a single node (e.g. on a laptop or a workstation), these additional mpirun options might be useful:

mpirun -np 3 --mca btl self,vader --oversubscribe stack exec example && echo SUCCESS || echo FAILURE

The options --mca btl self,vader enable the shared memory byte transfer layer (called “vader”), and also disable any network communication.

The option --oversubscribe lets you run as many MPI processes on the local node as you want, without being limited by the physical number of cores. This is convenient for testing.

Other MPI implementations should have equivalent (but differently named) options.

Running the tests

There are four test cases provided in tests:

stack build --test --no-run-tests
mpirun-openmpi-mp -np 3 --mca btl self,vader --oversubscribe stack exec $(stack path --dist-dir)/build/mpi-test/mpi-test && echo SUCCESS || echo FAILURE
mpirun-openmpi-mp -np 3 --mca btl self,vader --oversubscribe stack exec $(stack path --dist-dir)/build/mpi-test-binary/mpi-test-binary && echo SUCCESS || echo FAILURE
mpirun-openmpi-mp -np 3 --mca btl self,vader --oversubscribe stack exec $(stack path --dist-dir)/build/mpi-test-serialize/mpi-test-serialize && echo SUCCESS || echo FAILURE
mpirun-openmpi-mp -np 3 --mca btl self,vader --oversubscribe stack exec $(stack path --dist-dir)/build/mpi-test-store/mpi-test-store && echo SUCCESS || echo FAILURE