面向大规模人工社会的异构并行仿真引擎设计  

CPU/GPU Based Heterogeneous Parallel Simulation Engine on Large Scale Artificial Society

作　　者：李祯[1] 邱晓刚[1] 郭刚[1] 陈彬[1] 

机构地区：[1]国防科学技术大学信息系统与管理学院,长沙410073

出　　处：《系统仿真学报》2014年第10期2285-2292,共8页Journal of System Simulation

基　　金：自然科学基金(91024030;91324013)

摘　　要：大规模人工社会仿真是研究社会科学的有效手段,日益发展的高性能计算资源,如GPU等为发展大规模人工社会仿真研究带来了新的机遇。基于保守同步时间策略提出了一种CPU/GPU异构协同的仿真引擎架构,CPU与GPU协同独立地完成模型的调度和执行。详细阐述了GPU并行仿真内核设计方法,针对大规模人工社会仿真的模型特点,对数据结构,调度算法等进行了重点分析。通过3层引擎管理架构,保证了各CPU与GPU之间的时间同步和消息通信的完成。最后以GameOfLife测试模型对GPU并行仿真内核的运行效率进行了评估,获得了相对CPU最高11.2倍的性能加速。Large scale artificial society is an effective ways to study the social science, and the developing high performance computing resources, such as GPU, and new chance of developing large scale artificial society simulation research was brought. Based on conservative time synchronization algorithm, a CPU/GPU based heterogeneous collaborative simulation engine architecture was proposed, in which CPU and GPU completed execution and scheduling of model collaboratively and independently. The designing method of the GPU-based parallel simulation kernel was described in detail. Aiming at the characteristic of large scale artificial society simulation models, the analysis on data organization and scheduling algorithm was focused. Through three-layer engine management architecture, the time synchronization and message communication between CPU and GPU were realized. The efficiency of GPU-based parallel simulation kernel was analyzed with GameOfLife model, and up to 11.2x speedup was obtained compared to CPU.

关 键 词：人工社会 大规模Agent仿真 CPU/GPU 保守并行仿真 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]