Domain decomposition models for parallel Monte Carlo transport

We present a strategy for parallelizing computations that use the transport method. It combines spatial domain decomposition with domain replication to realize the scaling benefits of replication while allowing for problems whose computational mesh will not fit in a single processor's memory. The mesh is decomposed into a small number of spatial domains - typically fewer domains than there are processors - and heuristics are used to estimate the computational effort required to generate the solution in each subdomain using Monte Carlo. That work estimate determines the number of times a subdomain is replicated relative to the others. Timing of runs for two problems show that the new method scales better than traditional domain decomposition.