Parallel discrete event simulation on shared-memory multiprocessors

This paper describes the implementation and the performance study of three parallel discrete event simulation methods on a shared memory machine. These methods, which share a single user interface, include the Chandy-Misra paradigm with deadlock avoidance; the Time Warp approach with direct, aggressive, and lazy cancellation; and a hybrid approach, which exploits the parallelism available at each point in simulated time. In this study the authors also examine the impact of task-partitioning and of processor self-scheduling on the efficient implementation of the above methods. Two kinds of systems are simulated: a synchronous multiprocessor machine and an asynchronous toroid network with FCFS server nodes. The performance of the implemented methods is discussed, and conclusions are drawn from the obtained results