Parallel Simulation of Large-Scale Artificial Society on CPU/GPU Mixed Architecture

Abstract

Parallel simulations focus on conservative or optimistic algorithms to guarantee state consistency and causal order of messages between logical processes (LPs). It is usually hard for application domain users to develop complicated models for parallel simulations. For simplicity in large-scale artificial society, a modified DEVS component model is advocated in time-stepped parallel simulation with two-phase synchronization. A two-tier parallel simulation engine is designed on CPU/GPU mixed architecture with support of MPI and OpenCL. One-sided communication is selected for reflection of remote components and message passing between LPs. For cooperation between CPU and GPU, a size of 512 work items in each group is recommended. The parallel simulation engine is implemented in a micro kernel manner. An artificial society based on agent, container, grid and relation models are used to test the performance on an ordinary computer and a cluster with varied scales. The maximum speedup reaches 46 and 114 on the computer and the cluster respectively with about half a million agents.