Accelerating large-scale distributed traffic simulation with adaptive synchronization method

In the era where traffic problems are critical for realizing smarter cities, large-scale and real-time traffic simulations are becoming important. To enable such a simulation in highly distributed environment such as supercomputers, we have built a microscopic traffic simulator called Megaffic on top of an X10-based distributed agent-based simulation framework. In previous work, we have found out that microscopic approach - by representing each vehicle as one agent - makes the synchronization serious bottleneck to realize a nearly scalability in distributed environment. In this paper, we propose a new approach that accelerates large-scale agent-based simulations by adaptively adjusting synchronization granularity. The tradeoff exists in that the precision of the simulation result might be lost to some extent, however we design our method in a way of not losing the precision as much as possible. In our experiment, we have used 192 CPU cores and the Tokyo road network data in a supercomputer and validated that our proposed method achieves at least 2.5 times speed-ups without sacrificing much precision with the comparison of the regular synchronization method.