Creating GPU-Enabled Agent-Based Simulations Using a PDES Tool

By offloading some computation to graphical processing units (GPUs), agent-based simulation (ABS) can be accelerated up to thousands of times faster. To exploit the power of GPUs, modellers can use available simulation frameworks to auto-generated GPU codes without requiring any knowledge of GPU programming languages. However, such frameworks only support computation on the GPUs of a particular vendor. This paper proposes techniques, implemented in a synchronous parallel discrete event simulation (PDES) tool, to allow modellers to create ABS models, and to specify computation regions in the models for multiple vendor’s GPUs or CPUs. The technique comprises a set of meta-language tags and a compilation framework to convert user-defined GPU execution regions to OpenCL. A well-known cellular ABS models, the Conway’s Game of Life, have been implemented and evaluated on two platforms

i.e.

, the NVIDIA GeForce 240M LE and AMD Radeon HD6650M. The preliminary results demonstrate two findings: (a) the proposed technique allows the example ABS model to be executed on a PDES engine successfully; (b) the generated GPU-enabled ABS model can achieve fourteen times faster than its multicore version.