Monte Carlo simulation with the GATE software using grid computing

Abstract

Monte Carlo simulations needing many replicates to obtain good statistical results can be easily executed in parallel using the "Multiple Replications In Parallel" approach. However, several precautions have to be taken in the generation of the parallel streams of pseudo-random numbers. In this paper, we present the distribution of Monte Carlo simulations performed with the GATE software using local clusters and grid computing. We obtained very convincing results with this large medical application, thanks to the EGEE Grid (Enabling Grid for EsciencE), achieving in one week computations that could have taken more than 3 years of processing on a single computer. This work has been achieved thanks to a generic object-oriented toolbox called DistMe which we designed to automate this kind of parallelization for Monte Carlo simulations. This toolbox, written in Java is freely available on SourceForge and helped to ensure a rigorous distribution of pseudo-random number streams. It is based on the use of a documented XML format for random numbers generators statuses.

Dynamic Partitioning of GATE Monte-Carlo Simulations on EGEE

“…Instead of sequentially simulating a large number (up to several billions) of particles that may take several days or even weeks to complete, smaller groups (bags) of particles can be simulated independently and eventually merged. This process is valid only if the sub-simulations are statistically independent, which is guaranteed by using the random number generator integrated in GATE [26]. Previous work has been done on the parallelization of Monte-Carlo simulations in a Grid environment.…”