【作者】 张少辉；

【导师】 马骏；

【作者基本信息】 河南大学 ， 应用数学， 2009， 硕士

【摘要】 网络技术的迅速发展为网络并行计算提供了很好的条件,基于局域网的机群并行计算和基于Internet的网格计算是其典型代表。机群并行计算以其易可扩展性、高性价比等优点成为当今高性能计算(High Performance Computing,HPC)领域的一个研究热点。在机群并行计算中,任务调度和动态负载平衡是进行网络并行计算的关键。在并行程序设计中,如果能准确衡量和预测结点负载,将会有效提高并行程序的执行效率,对其进行研究具有重要的理论和应用价值。论文首先介绍了并行计算技术和MPI(Message Passing Interface,消息传递接口)并行程序设计的相关内容,并对并行程序设计中出现的负载问题进行了研究。其次分析了负载的特性,为负载预测提供了依据。结合MPI并行编程标准在Windows系统上建立基于误差反向传播(error Back Proragation,BP)算法预测的负载平衡系统,给出了系统的规划方案和设计架构。采用自适应、主动的负载收集策略,建立负载收集模块,及时、准确地收集结点负载信息,为负载预测提供信息基础。采用BP神经网络预测结点的负载变化情况,并建立BP神经网络预测模型。最后在局域网内构建了基于MPI的并行计算平台,对系统模型进行了测试验证。实验结果表明,与MPI直接分配方式相比,基于BP算法负载预测设计的调度系统的性能有了一定的提高。更多还原

【Abstract】 The rapid development of network technology provides the proper conditions for the network parallel computing,and the typical representative is the Parallel computing based on LAN cluster and the grid computing based on the Internet.Parallel computing cluster for the excellent scalability and the high performance-price ratio has become a hot research topic in the High Performance Computing field at present.The task scheduling and dynamic load balancing are the key to network parallel computing.We can improve the efficiency of parallel program effectively if we can measure and predict node’s load accurately in the parallel programming,and conducts the research to have the important theory and the application value.In this paper,the content of parallel computing technology and the MPI(Message Passing Interface) parallel programming are introduced firstly,and the load balancing problem of parallel programming is researched.Secondly,the load characteristics are analyzed,which are the basis of load prediction.Establishing a load balancing system based on BP algorithm combined with the MPI parallel programming standard on the Windows system,and putting forward the plan and framework.Building the load collection module take advantage of the self-adaptive and initiative strategy of load collection,which can collect the node load information accurately and promptly,and providing the sources of data for the load prediction.Predicting the node’s load information changes using the BP neural network, and establishing the prediction model based on BP neural network.Finally,building a parallel computing platform based on MPI in LAN,and tesing the system model.Result shows that the load scheduling system based on the BP algorithm prediction improves performance certainly compared with the way of MPI direct allocation.更多还原