PnumpyDomainDecomp

Letting pnumpy do the domain decomposition

pnumpy does not impose any domain decomposition on your array, you're free to use any MPI topology that you like. In order to fetch remote data; however, you will need to know which processes are neighbor to a given sub-array. The following will help you achieve optimal decomposition in the sense of reducing the ratio of surface (communication) to volume (work) and determine the neighbors to a rank.

import pnumpy

# create a domain decomposition for 24 procs, the topology is a 2x3x4 cube
decomp = pnumpy.CubeDecomp(nprocs = 24, dims = (2, 3, 4))

# find the rank one block down from rank 5
neighRk = decomp.getNeighborProc(5, offset = (0, 0, -1))

Note that getNeighborProc can return None (no neighbor). You can specify which axes wrap around with te additional periodic argument

neighRk = decomp.getNeighborProc(5, offset = (0, 0, -1), periodic = (True, False, True))

Also note that there are cases where no regular domain decomposition can be found, which happens anytime the number of processes does not match the product of the dimensions.