High Throughput Monte Carlo

We present a cost-eeective framework for conducting large scale Monte Carlo (MC) studies which exploits the natural parallelism of the MC method to harness the power of large, dynamic collections of computing resources. We describe the beneets of the dynamic master-worker (MW) parallel programming paradigm and how task parallel and job parallel MW applications t into our framework. We discuss the issues involved in supporting MC applications in this framework, including random number generation, resource management, remote le access, and checkpointing. We conclude with descriptions of selected customer experiences.