Performance optimization for agent-based traffic simulation by dynamic agent assignment

It is indispensable to make full use of parallel and distributed systems with increasing demands for large-scale traffic simulation, but problems remain about insufficient scalability due to costs of synchronization by load unbalancing among compute nodes. To tackle this problem, we propose performance optimization method for traffic simulations to underlying road networks preprocessed by graph contraction introducing dynamic re-assignment vehicles and cross points to threads and nodes based on time-series traffic congestion. By applying the optimization and running the simulation of the real-world Dublin city on 16 compute nodes of TSUBAME 2.5, the simulation performance has improved by 4 times with the proposed graph contraction method. We compared the effect of optimizations between the agent assignment method and existing adaptive synchronization method with comparison to regular 1 synchronization per step.