社会シミュレーションのための異種並列計算環境における再現性の確保

抄録

In this paper, we propose a computational framework to run reproducible simulation on heterogeneous computing environments for social simulation using random number. From simulation results using various parameter settings, we focus not only on average results but also a unique agent or trial. To find unique trials from a number of simulation results, we need to accelerate large-scale simulations and multi-trial simulations. However, we observe different simulation results using parallel computing because random number sequences vary by the number of parallelism. We must ensure the reproducibility of simulations when we try to analyze simulation results in detail. In this paper, we propose a computational framework to run reproducible simulations using unparalleled CPU, paralleled CPU and GPU. Our experimental results show that we get the same simulation results in heterogeneous computing environments.