[Python] How to send/recv from a specific index in mpi4py

Here is a translation of an C+MPI example in Python+Numpy+mpi4py. The goal of this example was to show that the message received is put in memory and that memory is in one dimension

from mpi4py import MPI
import numpy as np
# Init MPI
comm = MPI.COMM_WORLD
rank = comm.Get_rank()
# Define parameters
nb_lines = 6
nb_columns = 5
tag = 100
# Init matrix on each process
a = np.full((nb_lines, nb_columns), rank,dtype=np.float32)
# Define a contiguous type_line
type_line = MPI.FLOAT.Create_contiguous(nb_columns)
type_line.Commit()

if rank == 0:
# Send first line
comm.Send([a, 1, type_line], 1, tag)
elif rank == 1:
# Receive
comm.Recv([a[nb_lines-2,nb_columns-1:], nb_columns, MPI.FLOAT], 0, tag)
print(f"{a}")


The ouput with 2 processes is the same as in C

[[1. 1. 1. 1. 1.]
[1. 1. 1. 1. 1.]
[1. 1. 1. 1. 1.]
[1. 1. 1. 1. 1.]
[1. 1. 1. 1. 0.]
[0. 0. 0. 0. 1.]]


Even if it seems to work, I find the syntax a[nb_lines-1,nb_columns-1:] weird but with the expression a[nb_lines-1,nb_columns-1] i have an error BufferError: scalar buffer is readonly.

I have not find any example that use index in communication buffer, so my question is what it is the good way to use a specific index in communication buffer with mpi4py+numpy ?

Continue reading...