Improving Parallelisation of a Monte Carlo Radiotherapy Simulation Using MPI

In this paper we improve upon a farmer-worker parallelisation of a Monte Carlo Radiotherapy Simulation by changing its communication model. Using clusters on the Amazon EC2 cloud, and the BEAMnrc system for Monte Carlo simulations as our case study, we demonstrate experimentally the performance bottlenecks faced when using a network file system for farmer-worker communications. We then suggest an alternative approach, which uses a message passing protocol. We find that the message passing approach is able to deliver a steady performance improvement as we scale up, while also avoiding application failure due to I/O thrashing and oversubscription.