Traffic simulation is a valuable tool for the design and evaluation of road networks. Over the years, the level of detail to which urban and freeway traffic can be simulated has increased steadily, shifting from a merely qualitative macroscopic perspective to a very detailed microscopic view, where the behavior of individual vehicles is emulated realistically. With the improvement of behavioral models, however, the computational complexity has also steadily increased, as more and more aspects of real-life traffic have to be considered by the simulation environment. Despite the constant increase in computing power of modern personal computers, microscopic simulation stays computationally expensive, limiting the maximum network size than can be simulated on a single-processor computer in reasonable time. Parallelization can distribute the computing load from a single computer system to a cluster of several computing nodes. To this end, the exisiting simulation framework had to be adapted to allow for a distributed approach. As the simulation is ultimately targeted to be executed in real-time, incorporating real traffic data, only a spatial partition of the simulation was considered, meaning the road network has to be partitioned into subnets of comparable complexity, to ensure a homogenous load balancing. The partition process must also ensure, that the division between subnets does only occur in regions, where no strong interaction between the separated road segments occurs (i.e. not in the direct vicinity of junctions). In this paper, we describe a new microscopic reasoning voting strategy, and discuss in how far the increasing computational costs of these more complex behaviors lend themselves to a parallelized approach. We show the parallel architecture employed, the communication between computing units using MPIJava, and the benefits and pitfalls of adapting a single computer application to be used on a multi-node computing cluster.


Zitierform konnte nicht geladen werden.


Nutzung und Vervielfältigung: