Optimizing Agent-Based Simulations for the GPU

Agent-based simulation (ABS) is a powerful tool to analyze the behavior of agents interacting with each other. Many simulations contain a large number of agents or require multiple executions with different parameters, making the simulation-based experimentation slow. In recent years, graphics processing units (GPUs) have proven to be a powerful platform for accelerating ABS. In this paper, we propose several novel techniques to further improve the efficiency of GPU-based ABS. Our techniques reduce the total amount of computation in ABS, improve memory access patterns, and exploit temporal locality to minimize redundant calculations. Furthermore, the optimizations can be implemented as a layer on top of existing simulators, as we demonstrate by modifying the popular GPU-based simulator FLAME. We perform experiments using the Boids and Social Force simulations, and show that our techniques improve overall performance by up to 3.2x and 2.1x respectively compared to the original FLAME.