MPI is based on a single program, multiple data (SPMD) model, where multiple processes are launched running independent programs, which then communicate as necessary via messages.

Basic example

A script should include using MPI and MPI.Init() statements before calling any MPI operaions, for example

# examples/01-hello.jl
using MPI

println("Hello world, I am $(MPI.Comm_rank(comm)) of $(MPI.Comm_size(comm))")

Calling MPI.Finalize() at the end of the program is optional, as it will be called automatically when Julia exits.

The program can then be launched via an MPI launch command (typically mpiexec, mpirun or srun), e.g.

$ mpiexec -n 3 julia --project examples/01-hello.jl
Hello world, I am rank 0 of 3
Hello world, I am rank 2 of 3
Hello world, I am rank 1 of 3

The mpiexec function is provided for launching MPI programs from Julia itself.

CUDA-aware MPI support

If your MPI implementation has been compiled with CUDA support, then CUDA.CuArrays (from the CUDA.jl package) can be passed directly as send and receive buffers for point-to-point and collective operations (they may also work with one-sided operations, but these are not often supported).

If using Open MPI, the status of CUDA support can be checked via the MPI.has_cuda() function.


In order to ensure MPI routines are called in the correct order at finalization time, MPI.jl maintains a reference count. If you define an object that needs to call an MPI routine during its finalization, you should call MPI.refcount_inc() when it is initialized, and MPI.refcount_dec() in its finalizer (after the relevant MPI call).

For example

mutable struct MyObject
    function MyObject(args...)
        obj = new(args...)
        # MPI call to create object
        finalizer(obj) do x
            # MPI call to free object
        return obj