Scalable, repeatable, and contention-free parallelization of traffic simulation

Abstract

Tripod is a project funded by ARPA-E and partly carried on by the Intelligent Transportation Systems (ITS) Lab at MIT that aims to promote more energy efficient travel options by offering commuters incentives to make smart travel choices. These incentives depend on the current network state, and the ability to estimate the state of a given road network in real time is crucial. It relies on the DynaMIT system to determine what these incentives ought to be in order to optimize traffic flow on the network. Developed by the ITS lab, DynaMIT uses simulation to compute the current network state, predict its state in the future and, by extension, compute the incentives to travelers that optimize the global energy gain. While DynaMIT is able to do this effectively within smaller areas, it is unable to simulate traffic for the Greater Boston Area, or GBA, due to the scale of the network. The goal of this thesis is to scale the DynaMIT system so that it is less affected by network sizes. First, we outline a custom, lightweight profiling tool that is able to better track down the problems with scalability; next, we build off of previous work to address design errors that slow serial execution time; and finally, we implement a novel way to parallelize traffic simulation that avoids the race conditions and concurrency issues generally associated with such systems.