CODEWITHSUNDEEP
pythonmpicluster-computinghpcmpi4py
Commoner
Commoner

Reputation: 1768

mpi4py | comm.bcast does not work

I am working on a simple python script to test mpi4py. Specifically, I want to broadcast a scalar and an array from a given processor (say rank 0), so that all other processors have access to the values of the broadcasted scalar and the array in subsequent steps.

This is what I have done till now:

from __future__ import division
from mpi4py import MPI
import numpy as np

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

if rank==0:
    scal = 55.0
    mat = np.array([[1,2,3],[4,5,6],[7,8,9]])

    arr = np.ones(5)
    result = 2*arr

    comm.bcast([ result , MPI.DOUBLE], root=0)
    comm.bcast( scal, root=0)
    comm.bcast([ mat , MPI.DOUBLE], root=0) 

for proc in range(1, 3):
    if (rank == proc):
        print "Rank: ", rank, ". Array is: ", result
        print "Rank: ", rank, ". Scalar is: ", scal
        print "Rank: ", rank, ". Matrix is: ", mat

But, I get the following errors:

NameError: name 'mat' is not defined
    print "Rank: ", rank, ". Matrix is: ", mat

Also, in my output (print "Rank: ", rank, ". Scalar is: ", scal and print "Rank: ", rank, ". Array is: ", arr), I do not see the values of scal and array. What am I missing here? I will really appreciate any help.

Upvotes: 2

Views: 5093

Answers (1)

jcgiret
jcgiret

Reputation: 758

I see here two errors:

  • Your variable scal and your numpy arrays mat, arr and results are only defined on rank 0. They should be defined on all the MPI ranks. Indeed, as the data are broadcasted on all the ranks, the variables and the Numpy arrays must be allocated to store the received results.
  • bcast is intended for Python objects, and are pickled (e.g. serialized) in order to be sent. Bcast is intended for Numpy array. So use the different calls accordingly with respect to what you are sending/receiving. Moreover, they have to be called on all the ranks.

As I am using Python 3, I have also corrected the print calls. Yet, you should not notice any issue with Python 2 due to the print_function import from future

Finally, I advise you to have a look at the MPI4Py tutorials here: http://mpi4py.scipy.org/docs/usrman/tutorial.html. I think they cover a large spectrum of what you may do with MPI4Py.

Here is something working:

from __future__ import division, print_function
from mpi4py import MPI
import numpy as np

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


scal = None
mat = np.empty([3,3], dtype='d')

arr = np.empty(5, dtype='d')
result = np.empty(5, dtype='d')


if rank==0:
    scal = 55.0
    mat[:] = np.array([[1,2,3],[4,5,6],[7,8,9]])

    arr = np.ones(5)
    result = 2*arr

comm.Bcast([ result , MPI.DOUBLE], root=0)
scal = comm.bcast(scal, root=0)
comm.Bcast([ mat , MPI.DOUBLE], root=0)

print("Rank: ", rank, ". Array is:\n", result)
print("Rank: ", rank, ". Scalar is:\n", scal)
print("Rank: ", rank, ". Matrix is:\n", mat)

Upvotes: 8

Related Questions