Computer modelling and simulation of road networks are a vital tool used to evaluate, design and manage road network infrastructure.
Road network simulations are however computationally expensive, with simulation runtime imposing limits on the scale and quantity of simulations performed within a reasonable time frame.
This thesis examines the appropriateness of many-core processing architectures (such as GPUs) for the acceleration of microscopic and macroscopic road network simulation, and the potential impact on the choice of modelling approach.
Fine-grained agent-based microscopic simulations of individual vehicles are parallelised using GPUs, achieving high performance through a novel graph-based communication strategy for data-parallel simulations.
A minimal benchmark model and scalable road network are defined and used experimentally to evaluate performance compared to Aimsun, a commercial simulation tool for multi-core processors.
Performance improvements of up to 67x are demonstrated for large scale simulations.
High-level macroscopic simulations model network flow rather than individual vehicles.
Although less computationally demanding than microscopic models, simulation runtimes can still be significant, often due to the calculation of many shortest paths.
A novel Many-Source Shortest Path (MSSP) algorithm is proposed to concurrently find multiple shortest paths through sparse transport networks using GPUs.
This is embedded within a commercial multi-core CPU macroscopic simulation tool, SATURN, and the performance evaluated on large-scale real-world road networks, demonstrating assignment performance improvements of up to 8.6x when comparing multi-processor GPU and CPU implementations.
Finally, the impact of the performance improvements to both modelling techniques are evaluated using a common benchmark model and the relative improvements demonstrated by the benchmarking of each approach using different transport networks.
These results suggest that GPUs will allow modellers to shift towards using finer-grained simulations for a broader range of modelling tasks.
