Accelerating Agent-Based Modeling Using Graphics Processing Units

Abstract

Agent-based modeling is a disaggregated simulation approach for the exploration of complex spatial dynamics in geographic systems. The use of agent-based models for investigating social-ecological complexity in geographic systems is, however, severely hampered by the computational intensity of agent-based models. Graphics Processing Units (GPUs) are cutting-edge many-core parallel computing platforms that hold great potential in addressing this computational intensity. It is thus necessary to identify aspects that are fundamental in guiding the transformation of agent-based models into GPU environments. The objective of this paper is to identify and discuss the fundamental aspects that need to be considered when using GPUs to accelerate agent-based models. Specifically, these aspects include random number generation, parallelization of agent-based interactions, analysis of agent and environment patterns, and evaluation of computing performance. By linking with these aspects, I used a case study of modeling spatial opinion exchange to illustrate the massively parallel computing power of GPUs for accelerating agent-based modeling. Experimental results suggest that these aspects provide valuable guidance for transforming agent-based models into GPUs to best exploit massively parallel computing power. Further, these aspects are of vital importance for bridging the gap between advancement in GPUs and their applications for resolving spatiotemporal problems using agent-based modeling.