mpi_jl_pi.py

# This is an example of how to integrate python's MPI wrapper, mpi4py, with
# julia's MPI wrapper, MPI.jl. This example approximates pi using Monte Carlo
# integration.
#
# We make use of pyjulia which gives access to Julia from Python.  For some
# basic usage of pyjulia see the jupyter notebook PyJulia_Demo.ipynb in
# Julia_HPC_Guides/demos/notebooks. For more details (including installation
# instructions) see the pyjulia documentation:
#     https://pyjulia.readthedocs.io/en/latest/
# For more details on mpi4py and MPI.jl, see their documentation:
#     https://mpi4py.readthedocs.io/en/stable/
#     https://juliaparallel.github.io/MPI.jl/stable/
#
# WARNING: Before running this script, you must setup mpi4py and MPI.jl to use
# the SAME MPI library. This is described in the README.md found in the same
# directory as this script.

# Something in the below julia interface setup forks a process--setup julia before
# importing mpi4py to keep MPI from complaining about the fork
from julia.api import Julia
jl = Julia(compiled_modules=False)

import mpi4py
from mpi4py import MPI as pympi

import numpy as np

comm = pympi.COMM_WORLD

# Activate the desired julia environment
jl.using('Pkg')
from julia import Pkg
Pkg.activate(".")

jl.using('MPI')
jl.using('Random') # for seed! function
jl.using('Statistics') # for mean function

from julia import Main
from julia import MPI as jlmpi

# Convert pympi comm to jlmpi comm
Main.handle = pympi._handleof(comm) # make handle accessible to julia
jl.eval('comm = MPI.Comm(MPI.MPI_Comm(handle))') # create julia comm

# WARNING: You might think that we could use a statement like
#     Main.comm = jlmpi.Comm(jlmpi.MPI_Comm(pympi._handleof(comm)))
# to turn the python MPI comm into a julia MPI comm instead of the above `eval`.
# However, this will fail when using MPICH (it works with OpenMPI). The reason
# is that MPICH uses integers to differentiate MPI comms (OpenMPI uses raw
# pointers) . So for MPICH, `jlmpi.MPI_Comm(pympi._handleof(comm))` returns a
# `Cint` (which is a specialized julia Int32 for interfacing with C/Fortran
# libraries). When it comes back to python, it is converted to a python `int`
# which is then converted to a Julia Int64 when given to `jlmpi.Comm` as an
# argument. The result is a type error. We can avoid this MPICH incompatibility
# by using the above `eval` statement.

# Initialize Julia MPI without initializing the libmpi--this is part of the
# initialization done by MPI.Init() in MPI.jl
jl.eval(
    'function init_mpi(); for f in MPI.mpi_init_hooks; f(); end; return; end;'
)

Main.init_mpi()

rank = jlmpi.Comm_rank(Main.comm)
size = jlmpi.Comm_size(Main.comm)

if rank == 0:
    jl.eval('N = Vector{Int}([1000000])')
else:
    jl.eval('N = Vector{Int}(undef, 1)')

# python doesn't like the '!' character so use an eval call here
jl.eval('MPI.Bcast!(N, 0, comm)')
Main.n = Main.N[0]

jl.eval('Random.seed!({})'.format(rank))

Main.x = Main.rand(Main.n)
Main.y = Main.rand(Main.n)
jl.eval('val = [4.0 * mean(x.^2 + y.^2 .< 1.0)]')
# We could also use the following:
# Main.val = 4.0 * Main.mean(x**2 + y**2 < 1.0)

Main.rbuf = jlmpi.Gather(Main.val, 0, Main.comm)

if rank == 0:
    print("pi estimate = {}".format(Main.mean(Main.rbuf)))

exit()