Skip to content

[Investigate] Custom MPI_Reduce operations #99

@PhilipVinc

Description

@PhilipVinc

I was pleasantly surprised by seeing that mpi4jax automatically supports custom MPI_Reduce operations. See for example the following:

import numpy as np
from mpi4py import MPI
import mpi4jax
import jax
import jax.numpy as jnp
from functools import partial

rank = MPI.COMM_WORLD.rank

# create numpy arrays to reduce
src = (np.arange(8) + rank*8).reshape(4,2)
src[0] = rank
src = jnp.array(src)
dst = np.zeros_like(src)

MPI.COMM_WORLD.barrier()
print("starting")
MPI.COMM_WORLD.barrier()

def myadd(xmem, ymem, dt):
    x = np.frombuffer(xmem, dtype=src.dtype)
    y = np.frombuffer(ymem, dtype=src.dtype)

    z = x + y

    print("Rank %d reducing %s (%s) and %s (%s), yielding %s" % (rank, x, type(x), y, type(y), z))

    y[:] = z

op = MPI.Op.Create(myadd, commute=True)

#MPI.COMM_WORLD.Reduce(src, dst, op)

jax.jit(partial(mpi4jax.reduce, op=op, root=0))(src)

if MPI.COMM_WORLD.rank == 0:
    print("ANSWER: %s" % dst)

MPI.COMM_WORLD.barrier()
print("-------------------------------")
MPI.COMM_WORLD.barrier()


jax.jit(partial(mpi4jax.allreduce, op=op))(src)

if MPI.COMM_WORLD.rank == 0:
    print("ANSWER: %s" % dst)

However this works because mpi4py calls back into the python runtime, which will slow down the execution.
Ideally I'd like to use numba CFFI to define custom operations without calling back into python...

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions