Large-scale multi-agent mobility simulations on a GPU: towards high performance and scalability

摘要

This paper describes the ongoing development of GEMSim, a GPU-based mobility simulator that is systematically designed for generic large-scale networks and population samples. In order to fully exploit the benefits of the massively parallel architecture of GPU hardware, in GEMSim, the structure of the overall simulation loop, the organisation of memory transactions on GPU, data structures on both GPU and host, and the learning process are considered carefully. First results for a large-scale scenario of Switzerland are presented, and show that a whole simulation loop of GEMSim is more than 12 times faster than MATSim, and mobility simulations run up to 58 times faster in GEMSim compared to MATSim. Thus, this GPU-based mobility simulator makes practical advanced traffic simulation and forecasting tools more accessible to planners and decision makers.