References:

• https://sea.ucar.edu/sites/default/files/4_mpi4py.pdf
• https://mpi4py.readthedocs.io/en/stable/tutorial.html#running-python-scripts-with-mpi

Reduce (Comm.reduce | Comm.Reduce):

It works exactly like standard MPI, every process has to call the routine.

Functions prototypes:

def reduce(
    sendobj: Any, #Data to be reduced
    op: Op | ((Any, Any) -> Any) = SUM, #Reduce operation
    root: int = 0 #Root rank
) -> (Any | None)
#Returns reduced data for the root process and none for the other processes.

def Reduce(
    sendbuf: BufSpec | InPlace, #Data to be reduced
    recvbuf: BufSpec | None, #Reduced data for the root, none for others
    op: Op = SUM, #Reduce operation
    root: int = 0 #Root rank
) -> None

The default possible operations are the same as in standard MPI, and you can see all of them here: https://learn.microsoft.com/en-us/message-passing-interface/mpi-op-enumeration. However, for Python objects, you can pass custom reduce functions instead of using a default operation with the 'reduce' function.

Usage examples:

Python object example:

from mpi4py import MPI

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

parsedSquaredRank = (rank+1)**2
print('Process ' + str(rank) + ' sending: ' + str(parsedSquaredRank))

reducedData = comm.reduce(parsedSquaredRank, MPI.SUM)

if(rank == 0):
  print('Reduced data: ', end="")
  print(reducedData)

Output:

Python object example (with custom reduce function):

from mpi4py import MPI

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

parsedSquaredRank = (rank+1)**2
print('Process ' + str(rank) + ' sending: ' + str(parsedSquaredRank))

reducedData = comm.reduce(parsedSquaredRank, lambda a,b: a+b)

if(rank == 0):
  print('Reduced data: ', end="")
  print(reducedData)

Output:

Buffer-like object example:

from mpi4py import MPI
import numpy

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

parsedSquaredRank = numpy.array([(rank+1)**2], dtype='i')
print('Process ' + str(rank) + ' sending: ' + str(parsedSquaredRank))

reducedData = numpy.empty(1, dtype='i')
comm.Reduce(parsedSquaredRank, reducedData, MPI.SUM)

if(rank == 0):
  print('Reduced data: ', end="")
  print(reducedData)

Output: