General-Purpose Computing on GPUsを用いた再現性のあるAgent-Based Simulationの高速化

In this study, we try to accelerate reproducible Agent-Based Simulation (ABS) using the General- Purpose Computation on GPUs (GPGPU). GPGPU is a technology that uses the calculation resource of GPU to calculate other than the image processing. CPU and GPU architecture and their programming technique are diffrent. Therefore it is difficult to run reproducible ABS on CPU and GPU. In this study, we propose two models to run reproducible ABS in CPU and GPU. The first is a stand-alone model that to parallelizes the independent ABS. The other is a distributed model to run large-scale ABS parallelizing decisions agents. These models can be ensured reproducibility even in the experiment of changing the number of parallelization. The results of the experiment show that the stand-alone model was able to obtain a 40-fold acceleration rate than CPU. The distributed model was able to obtain a 120-fold acceleration rate than CPU.