Skip to content

Latest commit

 

History

History
83 lines (54 loc) · 2.37 KB

collectives-n-to-1.md

File metadata and controls

83 lines (54 loc) · 2.37 KB
Raw

Collective communication: many to one

Next step is to look into how to use collective communication to collect data from all tasks to a single task, i.e. how to move data from many to one.

Gather

Gather collects an equal amount of data from all the processes in a communicator to one process. One can think of it as an inverse scatter operation that allows one e.g. to collect partial results from the tasks.

Segments A, B, etc. may contain multiple elements just like in scatter.

An example of gathering a list of single values (rank) from each process as well as a numpy array of multiple elements (data) that are then stored in a larger receive array (buffer):

from mpi4py import MPI
from numpy import arange, zeros

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

data = arange(10, dtype=float) * (rank + 1)
buffer = zeros(size * 10, float)

n = comm.gather(rank, root=0)     # returns the value
comm.Gather(data, buffer, root=0) # in-place modification

Reduce

Reduce gathers data from all the processes in a communicator and applies an operation on the data before storing the result in a single process. Essentially, it is just like gather, but with an additional operation applied to the gathered data.

The operation that is applied on the data can be anything from a sum to a logical XOR and anything in between. MPI supports a wide variety of operations that can be used, including e.g. maximum value (MPI.MAX), minimum value (MPI.MIN), summation (MPI.SUM), and product (MPI.PROD).

An example of reduction using MPI.SUM to calculate the total sum of all received values:

from mpi4py import MPI
from numpy import arange, zeros

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

data = arange(10 * size, dtype=float) * (rank + 1)
buffer = zeros(size * 10, float)

n = comm.reduce(rank, op=MPI.SUM, root=0)     # returns the value
comm.Reduce(data, buffer, op=MPI.SUM, root=0) # in-place modification

Other common collective operations

  • Scatterv (each process receives different amount of data)
  • Gatherv (each process sends different amount of data)