Dynamic load balancing of parallel road traffic simulation

The objective of this research was to investigate, develop and evaluate dynamic load-balancing strategies for parallel execution of microscopic road traffic simulations. Urban road traffic simulation presents irregular, and dynamically varying distributed computational load for a parallel processor system. The dynamic nature of road traffic simulation systems lead to uneven load distribution during simulation, even for a system that starts off with even load distributions. Load balancing is a potential way of achieving improved performance by reallocating work from highly loaded processors to lightly loaded processors leading to a reduction in the overall computational time. In dynamic load balancing, workloads are adjusted continually or periodically throughout the computation.
In this thesis load balancing strategies were evaluated and some load balancing policies developed. A load index and a profitability determination algorithms were developed. These were used to enhance two load balancing algorithms. One of the algorithms exhibits local communications and distributed load evaluation between the neighbour partitions (diffusion algorithm) and the other algorithm exhibits both local and global communications while the decision making is centralized (MaS algorithm). The enhanced algorithms were implemented and synthesized with a research parallel traffic simulation. The performance of the research parallel traffic simulator, optimized with the two modified dynamic load balancing strategies were studied.