mpi4py hangs when trying to send large data

问题

i've recently encountered a problem trying to share large data among several processors using the command 'send' from the mpi4py-library. Even a 1000x3 numpy float array is too large to be sent. Any ideas how to overcome this problem?

thx in advance.

回答1:

I've found a simple solution. Divide data into small enough chunks...

回答2:

I encountered this same problem with Isend (not with Send). It appears that the problem was due to the sending process terminating before the receiver had received the data. I fixed this by including a comm.barrier() call at the end of each of the processes.

回答3:

Point-to-point send/recv of large data works:

#!/usr/bin/env python
from mpi4py import MPI
import numpy

comm = MPI.COMM_WORLD
rank = comm.Get_rank()

if rank == 0:
    data = numpy.arange(300*100000, dtype='f')
    comm.Send([data, MPI.FLOAT], dest=1, tag=77)
elif rank == 1:
    data = numpy.empty(300*100000, dtype='f')
    comm.Recv([data, MPI.FLOAT], source=0, tag=77)
    print data

Running this with two processors:

% ~/work/soft/mpich/bin/mpiexec -np 2 ./send-numpy.py
[  0.00000000e+00   1.00000000e+00   2.00000000e+00 ...,   2.99999960e+07
   2.99999980e+07   3.00000000e+07]

来源：https://stackoverflow.com/questions/21528275/mpi4py-hangs-when-trying-to-send-large-data