基于Linux的小型集群的研究与实现

发布时间：2018-05-02 12:11

本文选题：并行计算 + 微机集群　；参考：《南京邮电大学》2013年硕士论文

【摘要】：采用传统的单个处理器进行运算已经无法满足人们对计算能力的需求，而基于向量处理机和对称多处理机的高性能计算机存在扩展性差、价格昂贵、整体性能较低的问题。针对上述问题，探讨了集群系统的相关理论、技术和方法，该系统易于实现，具有良好的可扩展性、可用性以及很高的性价比使其在商业和科学研究的各个领域里受到人们的青睐。本文使用普通以太网交换机和四台PC在Linux环境下完成集群系统的构建，该集群系统包括NFS文件共享系统、SSH远程登陆系统、MPI并行编程库以及PBS作业管理系统等软件模块。利用C+MPI编程模式，以计算PI值的串行算法和并行算法为例，通过比较不同节点数、不同处理器数的程序运算时间，，验证了该集群系统的并行性。算法的并行度容易受矩阵划分方法的影响，常见的矩阵划分方法主要有带状划分和棋盘划分，通过矩阵相乘的算例分析并比较了这两种划分方法的时间和加速比，实验结果表明，棋盘划分方法能开发更高的并行度。 HPL是测试集群系统浮点性能的最佳选择，通过调整HPL.dat中的矩阵规模、LU分解的分块大小、处理器网格大小以及参加计算的节点数等参数，总结了HPL测试参数的选取原则，优化了集群系统的性能，评测出集群系统的最优运算速度为13.61Gflops。
[Abstract]:The traditional single processor is no longer able to meet the demand for computing power, but the high performance computer based on vector processor and symmetric multiprocessor has the problems of poor expansibility, high price and low overall performance. Aiming at the above problems, the related theories, techniques and methods of cluster system are discussed. The system is easy to realize and has good expansibility. Availability and high cost-effectiveness make it popular in all fields of commercial and scientific research. In this paper, the cluster system is constructed in Linux environment by using ordinary Ethernet switch and four PCs. The cluster system includes NFS file sharing system, NFS remote landing system, MPI parallel programming library, PBS job management system and so on. Using C MPI programming mode, the parallelism of the cluster system is verified by comparing the program operation time of different nodes and different processors, taking the serial algorithm and parallel algorithm for calculating Pi value as examples. The parallelism of the algorithm is easy to be affected by matrix partitioning methods. The common matrix partitioning methods mainly include banded partition and chessboard partitioning. The time and speedup ratio of these two partitioning methods are analyzed and compared by an example of matrix multiplication. Experimental results show that the chessboard partition method can develop a higher degree of parallelism. HPL is the best choice for testing floating-point performance of cluster system. By adjusting the size of matrix in HPL.dat and the partition size of LU decomposition, the size of processor grid and the number of nodes participating in the calculation, the selection principle of HPL test parameters is summarized. The performance of cluster system is optimized and the optimal operation speed of cluster system is determined to be 13.61 Gflops.
【学位授予单位】：南京邮电大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP338;TP316.81

【参考文献】