A massive multilevel-parallel microscopic traffic simulator with gridlock detection and solving

Abstract

Traffic simulators based on microscopic models are a detailed approach to infrastructure and policy evaluation. They allow a closer to reality representation of the factors that influence traffic flow: individual driver and vehicle attributes, dynamic route decisions, lane changing and restrictions, driver cooperation. Such aspects add to the complexity and volume of computations, leading to slower simulation speeds compared to macroscopic models. Also, route restrictions can lead to gridlocks, a common problem in such simulations. In this paper, we propose a multi-level parallel architecture for the TrafficWeb microscopic traffic simulator. The solution combines random load allocation, for multithreaded processing, and distributed parallelization, through geographical domain decomposition. Adaptive load balancing is used for optimizing the distributed processing speed. Gridlock detection and solving are employed through efficient parallel and distributed algorithms, significantly decreasing their cost. Performance tests show an overall efficiency of 85% for the multilevel-parallel architecture, on a cluster with 5 nodes, each having 4 cores. This allows simulating metropolitan traffic 85 times faster than in real time.