Multi-agent simulation on multiple GPUs

Multi-agent simulation is widely used in many areas including biology, economic, political, and environmental science to study complex systems. Unfortunately, it is computationally expensive. In this paper, we shall explore the implementation of a general multi-agent simulation system on a system with multiple GPUs acting as accelerators. In particular, we have ported the popular Java multi-agent simulation framework MASON to a nVidia CUDA-based multi-GPU setting. We evaluated our implementation over different numbers and types of nVidia GPUs. For our evaluation, we ported three models in the original version of MASON. On the well-known Boids model, we achieved a speedup of $187\times$. Using a fictional model, we showed that speedup of up to $468\times$ is possible. In the paper, we shall also describe the detailed internals of our system, and the various issues we encountered and how they were solved.