Exposing Inter-Process Information for Efficient Parallel Discrete Event Simulation of Spatial Stochastic Systems.

We present a new efficient approach to the parallelization of discrete event simulators for multicore computers, which is based on exposing and disseminating essential information between processors. We aim specifically at simulation models with a spatial structure, where time intervals between successive events are highly variable and without lower bounds. In Parallel Discrete Event Simulation (PDES), the model is distributed onto parallel processes. A key challenge in PDES is that each process must continuously decide when to pause its local simulation in order to reduce the risk of expensive rollbacks caused by future \"delayed\"' incoming events from other processes. A process could make such decisions optimally if it would know the timestamps of future incoming events. Unfortunately, this information is often not available in PDES algorithms. We present an approach to designing efficient PDES algorithms, in which an existing natural parallelization of PDES is restructured in order to expose and disseminate more precise information about future incoming events to each LP. We have implemented our approach in a parallel simulator for spatially extended Markovian processes, intended for simulating, e.g., chemical reactions, biological and epidemiological processes. On 32 cores, our implementation exhibits speedup that significantly outweighs the overhead incurred by the refinement. We also show that our resulting simulator is superior in performance to existing simulators for comparable models, achieving for 32 cores an average speedup of 20 relative to an efficient sequential implementation.