高性能集群系统分析、设计与应用

【作者】 田俊刚；

【导师】 宋东；

【作者基本信息】 西北工业大学 ， 通信与信息系统， 2004， 硕士

【摘要】 高性能集群是一组相互独立的服务器通过高速网络连接构成的高性能计算系统。相对于其它高性能计算机，它的优点是可扩展性好，具有高可用性和很高的性价比。 本文通过对高性能集群技术的分析和对集群互连网络的测试比较，选用SCI可扩展一致性接口技术设计了一个高性能集群，并实现了气象并行计算在高性能集群上的应用。 全文由四部分构成：高性能集群技术、高速互连网络的分析、高性能集群的设计和高性能集群在气象数值预报计算中的应用。 论文首先阐述了高性能计算机的需求和发展现状，研究了高性能集群的体系结构和软、硬件组成，然后详细分析了对集群性能起重要作用的几种高速互连网络的结构和组成，并通过测试对它们的性能指标做了比较。在此基础上，根据气象预报计算的具体需求，采用16节点和SCI网络，设计了高性能集群。同时对该集群网络(SCI)的可用性进行了深入的分析。最后为验证集群的性能，实现了气象数值预报中短期中尺度模式MM5、通用气候模式CCM3和区域气候模式RegCM3在该集群上的调试计算。 综上所述，本文完成了一个高性能集群的详细设计。同时在集群互连技术的分析比较中提供了大量有用的数据，为今后进一步设计大规模的集群奠定了基础。更多还原

【Abstract】 High performance cluster is a group of single servers that connected together to form a high performance computing system through the high-speed network. Compared with other high performance computers, the great virtue of the cluster is excellent scalability, high reliability and challenging performance value rate.Based on the analysis of the high performance cluster and the comparison of the cluster’s network, the thesis chooses the Scalable Coherent Interface (SCI) to design the cluster, and then realizes the weather parallel application on it.In the first part, the thesis describes the international trend of high performance cluster and researches the system structure of cluster. Secondly, several high-speed networks of the cluster are analyzed, including the structures and the components of them. Through the analysis, the result of performance comparison test of these networks is provided. According to the need of weather forecast computing, the thesis adopts 16 nodes and SCI to construct the cluster. Finally, the thesis realizes the application of weather numerical value forecast in the cluster, such as Mesoscale Model 5(MM5), Common Climate Model (CCM3) and Regional Climate Model 3(RegCM3).In conclusion, a high performance cluster system is designed and analyzed, which provides lots of useful data for advanced research.更多还原