High Performance Computing in Dynamic Traffic Simulation
Abstract: Dynamic traffic simulation enables to emulate the traffic congestion, which is a necessary aspect of traffic evolution. However, dynamic traffic simulations must compute in seconds if they are to be used in real-time traffic management systems. High performance computing (HPC) resources provide the power and computation to significantly speed up these simulations, thus enabling their use in instantaneous traffic control.
Macroscopic dynamic simulation uses continuum traffic-flow models that are based on traffic volume, density and speed. Though there have been many works that apply parallel computation in HPC for agent-based microscopic dynamic simulation, parallel macroscopic simulation has not been studied adequately. Hence, we devised a parallel strategy for the Berkeley Advanced Traffic Simulator (BeATS), which a simulation framework for macroscopic dynamic simulation. Given n cores, the parallel simulation begins by having the root core (processor 0) partitioning the network into n minimum-cut partitions using the METIS program. Core i loads the ith partition and computes the corresponding traffic states. The cores use graph-based MPI interface to communicate boundary information to other cores that have adjacent network partitions.
We implemented the parallel BeATS simulator on Cori supercomputer at NERSC (nersc.gov). The parallel BeATS simulator was tested on a synthetic grid network with 2500 nodes, 10000 links, and over 600 origin-destination pairs. Results showed linear speed-up as the number of compute cores grew from 1 up to 8 cores. The simulation time was reduced from 28 minutes to 25 seconds with 256 cores.
Back to Women in HPC: Diversifying the HPC Community Archive Listing
Back to Full Workshop Archive Listing