A Case Study of Efficient Social Network Simulation through General Processing on Graphics Processing Units (GPGPUs)

Abstract

Agent based simulation has been both a large area of study and a widely used tool for scientific research in past years. Current implementations run on standard CPUs, and with the requirement of processing ever growing data sets, higher computational speed is of the utmost importance. General processing on graphics processing units (GPGPU) is an emerging platform offering the possibility of increased speed for data sets and models that can be processed in parallel. Agent based simulation is one such candidate for performance gains in a GPGPU implementation. My research has focused on thoroughly investigating GPGPU?s suitability for providing researchers with a more efficient way of conducting agent based simulation research. Studies were done using two conventional models: two-dimensional diffusion and Conway’s Game of Life. I first created an optimized CPU diffusion model and, following a determination of accuracy, compared computational speed with an Open Graphics Library GPGPU implementation previously developed at Oak Ridge National Laboratory. Similarly, like studies were completed with the Game of Life. Following this strict CPU and GPGPU comparison, further comparisons and analyses were conducted with a widely used agent based simulation API, Repast. Evaluations involving Repast revolved around the premise that were GPGPU to be harnessed by researchers for agent based simulation, it must be competitive with currently used research technologies. Results obtained with both two-dimensional diffusion and the Game of Life show significant performance gains through GPGPU. For a plethora of data sizes, it has been found that the GPU processes the models in parallel at much greater rates than both optimized CPU code and Repast. Furthermore, as both sample size and the number of iterations through the model increase, the gap between GPU and CPU performance becomes even wider. These successful studies are to now be extended by investigating new models through GPGPU and exploring compatibility with necessary agent ba