Publications
Obtaining identical results on varying numbers of processors in domain decomposed particle Monte Carlo simulations
Domain decomposed Monte Carlo codes, like other domain-decomposed codes, are difficult to debug. Domain decomposition is prone to error, and interactions between the domain decomposition code and the rest of the algorithm often produces subtle bugs. These bugs are particularly difficult to find in a Monte Carlo algorithm, in which the results have statistical noise. Variations in the results due to statistical noise can mask errors when comparing the results to other simulations or analytic results. If a code can get the same result on one domain as on many, debugging the whole code is easier. This reproducibility property is also desirable when comparing results done on different numbers of processors and domains. We describe how reproducibility, to machine precision, is obtained on different numbers of domains in an Implicit Monte Carlo photonics code.