【作者】 李铮；

【导师】 薛质；

【作者基本信息】 上海交通大学 ， 通信与信息系统， 2012， 硕士

【摘要】 商业和科学研究的各个领域对计算能力的需求正在不断增长。传统的基于向量处理器和SMP的高性能计算机由于难以扩展、价格昂贵等原因始终局限在科学研究等专业领域。Linux系统和网络技术的发展为集群技术的发展提供了很好的条件。基于Linux的高性能集群具有良好的可扩展性,兼容性以及很高的性价比,使得其不仅在专业领域,也在小型高性能计算机的应用领域里成为了最为重要的研究对象本文构建的基于Linux的高性能集群包括管理节点、计算节点、网络互联设备、NFS文件共享系统、MPI并行编程库、BLAS线性函数库等多个软硬件模块。在上述的高性能集群的平台下,用C+MPI编程的形式实现了矩阵的并行乘法运算。矩阵相乘运算是在科学工程计算领域调用最多的运算之一,对矩阵相乘的实现具有测试和应用的双重意义。通过不同规模、不同处理器数量下,比较程序的运算时间验证了基于Linux高性能集群的并行性。HPL是验证高性能计算机性能最通用的Benchmark,也是高性能计算机TOP500的比较标准。它通过求解线性稠密方程组的运算测得高性能计算机的浮点运算能力。对HPL的优化是优化高性能计算机性能的一个重要课题。由于集群系统的复杂性,而且HPL提供很多参数供使用者调试,并没有一种通用的平台无关的调试方法。以往的优化一般是基于经验的穷举法,耗时很长而且很难涵盖所有的参数组合。本文通过提出一种基于参数影响因子排序的优化方案,寻找HPL参数调优的最佳路径。通过该方法对8节点、32核心、主频为2.67GHz的高性能平台进行优化,获得66%的效率。更多还原

【Abstract】 The demand of computing power in both commercial and scientificareas is growing rapidly. The development of traditional high-performancearchitecture, like vector-based high-performance computer processors andSMP, is limited in scientific research and some other professional fields.The development of Linux system and network technology provides agood condition for the development of cluster technology. Linux-basedhigh-performance cluster is widely used in many areas because of its goodscalability, good compatibility and low cost and becomes one of the mostimportant objects of study.A Linux-based high-performance cluster is built and it includes manyhardware and software modules such as management node, compute nodes,networking equipments, NFS file sharing system, MPI parallelprogramming library, BLAS etc.A matrix multiplication is parallel programmed in the language ofC+MPI under the high-performance cluster platform. Matrix multiplicationis one of the most widely used functions in the field of science andengineering computing. After running the parallel application withdifferent problem sizes and different number of processors and comparingthe test results, the parallelism of the Linux-based cluster is verified.HPL is the most popular method to measure the floating pointperformance of the high-performance computers, and is also the standardBenchmark for TOP500 super computers. Optimization of HPL is animportant step to improve the performance of high-performance computers.Since the cluster system is complicated and HPL provides a lot ofparameters for the users to tune, there is no general and platform independent method to optimize. The traditional methods are generallyexperience-based and exhaustive. They are time-consuming and areimpossible to cover all the possible combinations. This paper presents amethod based on the sort of impact factors of parameters and find out thebest tuning path of the parameters. A small Linux-based high-performancecluster which contains 8-node, 32 cores clocked at 2.67GHz is optimizedaccording to this method and an efficiency of 66% is obtained.更多还原