Research On Large-Scale Artificial Society Computational Accelerating Methods On CPU/GPU Heterogeneous Parallel Systems

Artificial society is an effective way for social science research, and its core methods are Agent-based modeling, simulation and experimental analysis. With the developing requirement of applications, the scale of artificial society raises gradually, Agent model and behavior rules is getting complex, and inter-relationship among networks is becoming bigger and bigger. In order to meet efficiency and timeliness requirement of computational experiment, the execution efficiency of large-scale artificial society then becomes the burning question. With the development of heterogeneous computational devices like GPU, heterogeneous parallel system of CPU/GPU can get tremendous computational performance under rational power consumption, which has become direction of the high performance computers and also applied in the simulation for high performance. Facing the ever-increasing scale and complexity of artificial society, to make use of the CPU/GPU heterogeneous parallel systems into optimize of simulation execution has become an effective ways to raising the performance of artificial societyAmong the framework of single process multi threads for conservative parallel simulation engine, the paper proposes a three layer-based engine management framework for CPU/GPU heterogeneous parallel simulation engine, and aiming at the characteristic of large-scale Agent simulation, we design and optimize the data structure and event scheduling algorithm of GPU-based simulation kernel. Because of the diverse application domain among artificial society, the paper proposes a method of specific domain-oriented GPU simulation computational service component, which can make up the limit of simulation kernel that show relative low performance for specific domain and make full use of the computational performance of heterogeneous device. Main research works of the paper as follow:1) A CPU/GPU-based accelerating computational method and implementation framework for large-scale artificial society are proposed. The GPU accelerating computational method jointly applied the different ways of using GPU which are GPU co-processor and general processor into large-scale Agent simulation computational acceleration. General processor of GPU deals with the model driven computation of Agent simulation together with CPU in the heterogeneous simulation engine. At the same time, GPU co-processor, which works as a computational component among the simulation engine, strip the large-scale current computation from specific artificial society application model and load it into GPU to implement parallel computation. Under three layers engine management-based simulation engine framework, heterogeneous simulation kernel and computational component can be flexible extend.2) A large-scale Agent simulation-oriented GPU simulation kernel is realized. The paper designs and optimizes the structure of Agent model and event queue according to the characteristic of Agent simulation and GPU execution, and based on the data structure, the synchronized conservative parallel policy-based GPU simulation kernel scheduling algorithm is proposed. Focusing on the current event operation during the process of GPU simulation kernel, a GPU event output algorithm is proposed and in order to promote the performance of the synchronized conservative parallel simulation algorithm, a GPU-based large scale time reduce algorithm is put forward. In the Game Of Life test Model, GPU simulation kernel get speedup of 11.2 compared to a single CPU.3) A GPU-based computational service component is proposed, the paper realizes an example of GPU-based social relation net searching component, and compared to CPU-based searching algorithm, the GPU accelerating method can not only reduce the memory assumption but also get a large performance promotion when facing large number of concurrent query operation.