混合型异构集群的组建及其应用

发布时间：2019-04-16 16:28

【摘要】：自有第三类科学之称的计算机科学诞生以来，人们认识世界改造世界的步伐不断加快，也不断的推动计算机科学的进一步发展：并行技术在慢慢取代串行技术，高性能计算越来越受到欢迎，以及随之而来的高性能计算机、超级计算机。随着计算机软硬件技术的进一步发展，利用普通计算机来构建拥有计算性能高、通信效率高的电子系统——集群系统，同时由于它有着高性能计算机和超级计算机不具备的价格便宜、组建管理维护简单等优点，越来越受到中小型研究小组的欢迎，它们的计算系统通常有着单机性能不高、计算节点架构差异大等特点。本文以三种不同架构的计算节点为实验硬件，以集成化的集群系统为基本操作系统来组建这种电子系统，主要通过合理设计节点属性和任务调度管理系统的配置来实现不同架构节点的差异化控制和资源最大化利用，通过给监控系统添加图形处理器模块实现监控系统对部分带有图形处理器节点的监控。详细阐述了集群这种电子系统从设计、实施到维护的全过程，并总结了其中的经验教训，为有意向组建集群的人们提供参考。对这种电子系统进行需求分析，并设计实现相应的功能。对实验电子系统进行了各项测试，在HPL测试中，选取问题规模31000和分块大小256作为测试参数，，8节点并行测试的实测性能是理论值的30%，当问题规模增大能提升到42%，仍有很大提升空间；利用NAMD对图形处理器进行测试，在ApoA1作为测试基准的情况下单显卡实现了4倍的加速比。本文还进一步了探讨了这种电子系统在高性能计算之外的其他应用。简化了内外部网站构建维护难度，节省了服务费；内部协作平台的建设加强了课题组内部的沟通交流，实现了多种方式的共享互助；项目管理系统实现了实时追踪项目进度，多平台查看项目日志，动态项目成员分配任务等功能。
[Abstract]:Since the birth of computer science, known as the third category of science, people have come to realize that the pace of world transformation has been accelerating and promoting the further development of computer science: parallel technology is slowly replacing serial technology. High-performance computing is becoming more and more popular, as well as high-performance computers, supercomputers. With the further development of computer software and hardware technology, the cluster system, which has high computing performance and high communication efficiency, is constructed by using ordinary computers. At the same time, it is more and more popular with small and medium-sized research groups because it has the advantages of low price and simple management and maintenance, which are not provided by high-performance computers and supercomputers, and their computing systems usually have low performance on a single machine. Computing node architecture differences and other characteristics. In this paper, three kinds of computing nodes with different architectures are used as the experimental hardware, and the integrated cluster system is used as the basic operating system to construct this electronic system. Through the rational design of node attributes and task scheduling management system configuration to achieve the different architecture of node differentiation control and maximize the utilization of resources. By adding graphics processor module to the monitoring system, the monitoring system can monitor some nodes with GPU. In this paper, the whole process of designing, implementing and maintaining the electronic system of cluster is described in detail, and the experiences and lessons are summarized, so as to provide reference for the people who intend to set up the cluster. The requirements of this electronic system are analyzed and the corresponding functions are designed and realized. The experimental electronic system is tested. In the HPL test, the problem size 31000 and the block size 256 are selected as the test parameters. The measured performance of the 8-node parallel test is 30% of the theoretical value, and can be increased to 42% when the scale of the problem increases. There is still a lot of room for improvement; The graphics processor is tested with NAMD, and the speedup of single graphics card is 4 times higher when ApoA1 is used as the benchmark. The application of this kind of electronic system to high performance computing is also discussed in this paper. It simplifies the difficulty of construction and maintenance of internal and external websites and saves the service charge, and the construction of internal cooperation platform strengthens the communication and communication within the research group, and realizes various ways of sharing and mutual assistance. Project management system realizes real-time tracking project progress, multi-platform view project log, dynamic project member assignment tasks and so on.
【学位授予单位】：湖南大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP38

【参考文献】