Exposing Inter-process Information for Efficient PDES of Spatial Stochastic Systems on Multicores

摘要: We present a new approach for efficient process synchronization in parallel discrete event simulation on multicore computers. aim specifically at of spatially extended stochastic system models where time intervals between successive inter-process events are highly variable and without lower bounds: This includes governed by the mesoscopic Reaction-Diffusion Master Equation (RDME). A central part our is mechanism optimism control, which each disseminates accurate information about timestamps its future outgoing interprocess to neighbours. gives precise basis deciding when pause local processing reduce risk expensive rollbacks caused “delayed” incoming events. apply natural parallelization Next Subvolume Method (NSM) simulating systems obeying RDME. Since this does not expose events, we restructure it such information, resulting algorithm called Refined Parallel NSM (Refined PNSM). have implemented PNSM simulator spatial Markovian processes. On 32 cores, achieves an efficiency ranging 43--95% large models, average 37% small compared sequential any code parallelization. It shown that gain restructuring naive into more than outweighs overhead. also show superior performance existing simulators multicores comparable models.