基于GPU的离散模拟统一计算软件框架研究

项目名称： 基于GPU的离散模拟统一计算软件框架研究

项目编号： No.11272312

项目类型： 面上项目

立项/批准年度： 2013

项目学科： 数理科学和化学

项目作者： 王小伟

作者单位： 中国科学院过程工程研究所

项目金额： 78万元

中文摘要： 离散模拟是一种主要的计算机模拟方法，它将模拟系统离散为大量相互作用的粒子，通过动力学计算描述每个粒子的行为，从而直接或间接复现系统的行为。它涵盖了物理系统微观、介观和宏观的不同尺度，是一类直观而且普适的计算方法，在科学研究和工程技术中得到广泛的应用。离散模拟虽然应用极其多样，却可能用统一的软硬件平台来实现并行计算。本项目基于在科学计算领域已崭露头脚的GPU，研究多种主流离散方法的大规模计算实现，对这些模拟方法进行分析抽象，提炼其共性的算法和特征，首次提出一个基于GPU平台的高效、统一的大规模离散模拟软件框架。采用面向对象的设计方法，研究确定该框架的层次结构及主要数据结构。计算框架支持CUDA、OpenMP和MPI三种主流并行编程环境，可用于分子动力学、格子玻尔兹曼方法、光滑粒子流体力学、离散元方法等在GPU集群上的大规模耦合应用计算，并且具有较高的计算效率和较好的通用性、易用性和可扩展性。

中文关键词： 离散模拟；图形处理器；计算软件框架；并行计算；

英文摘要： Discrete simulation is one kind of mainstream computational methods, which discretizes the simulation systems into lots of interacting particles and describes the particles with dynamics to reproduce the complex system directly or through statistics. It is an intuitionistic and generic approach in that it covers several length scales of physical system, including microscale, mesoscale and macroscale. Although it has a wide application, several discrete methods can be parallelized with uniform software platform on the same hardware for large-scale simulation. In this project, a uniform framework for parallel computation of discrete method based on GPU is firstly put forward to be studied and implemented. The hierarchy and main data structure of this software framework are designed with the object-oriented method based on the analysis and abstract of the common algorithm and characteristic and consideration of speciality of different discrete methods. CUDA, OpenMP and MPI is supported and coupled in this software framework and at least four kinds of discrete methods are dealt with, which are molecular dynamics (MD), lattice Boltzmann method (LBM), smooth particle hydrodynamics (SPH) and discrete element method (DEM). Large-scale computational applicateions are performed with this software framework to show its h

英文关键词： discrete simulation；graphic processing unit；computational software；parallel computation；