Abstract
Over the last decade or so, increases in hardware capability have not directly translated into increases in software performance. The emergence of multi-core and many-core architecture has caused difficulty for users who, without proficiency in software development, then turn to frameworks and toolkits for solutions. With the wide range of complex systems investigated throughout numerous disciplines, agent-based modelling (ABM) has become a popular way to study such systems. The bottom-up approach employed using ABM’s allows system behaviours to emerge through lower-level interactions, which is a characteristic of complex systems and may otherwise be lost when using less computationally demanding approaches that employ averaging (mean-field approaches). However, even many of these modelling frameworks are not suitable for large-scale problems as they are not able to efficiently use the parallel resources now available and therefore lack scalability.A number of coarse- and fine-grained optimisation techniques have been developed, using portable standards (MPI and OpenCL), to produce scalable solutions to help overcome these issues. A leading ABM framework (FLAMEGPU) is compared with the more direct approach of using General Purpose Graphics Processing Unit (GPGPU) in a selection of case studies to evaluate performance in terms of speed-up and spatial scaling. Implementation of an existing complex system facing scalability issues (soil microbial model) has been identified and relevant optimisation techniques have been applied and quantified. Future directions and improvements to further enhance the spatial up-scaling are presented.
| Date of Award | 2015 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Ruth Falconer (Supervisor) & Adam T. Sampson (Supervisor) |