【作者】 何飞跃；

【导师】 庞丽萍；

【作者基本信息】 华中科技大学 ， 计算机软件与理论， 2004， 硕士

【摘要】 随着高性能微处理器、高速网络的出现和对计算能力需求的增大,以廉价硬件和软件支撑的集群系统越来越被广泛地使用,引起集群技术的迅猛发展。集群文件系统是集群的一个重要组成部分,作为一种集群体系结构上的并行文件系统,它为用户提供一个虚拟化大容量存储器的统一访问接口和高I/O带宽。由于集群文件系统的文件数据分散存储在各个结点上,文件的定位需要借助元数据来完成,元数据的管理成为管理数据的一个关键。为了提高元数据管理的可靠性,需要具有容错能力的元数据管理系统。为此,我们针对集群文件系统的元数据管理,设计了一个双元服务器系统。该系统内部由两台元数据服务器组成,通过对元数据的镜像产生副本,保证元数据的可靠性;通过主服务器失效后从服务器接管服务来屏蔽故障,保证元数据服务的连续性。系统具有集中管理方式控制简单、易于实现和维护等优点,克服了其单一失效点的缺陷,同时又避免了分布式管理的一致性维护设计与开销。在Linux内核空间实现元数据镜像技术、故障检测技术、IP接管技术和恢复技术,具有对应用程序透明性的特点。系统的最终目的是将其结构推广到多机情况下,进一步提高容错能力,实现高可用性。为了提高元数据服务器的处理效率,提出了一种寄生式元数据存储管理方法。并行文件系统的元数据寄生在本地文件系统内核中,通过增加系统调用实现对寄生元数据的操作,保证对现有系统的兼容性。将该方法应用于PVFS(Parallel Virtual File System)的元数据管理,元数据操作性能提高大约5～8倍。更多还原

【Abstract】 With the emergence of high-performance microprocessor, high-speed network and the increscent demand of computing power, cluster of computer with cheap hardware and software is more and more used by users, which brings the quick development of cluster technologies. Cluster file system is an important component of a cluster. As a parallel file system on the cluster architecture, cluster file system provides a unified interface of accessing great capability storage and high I/O bandwidth for the users. Owing to the file data spreading around the cluster, locating a file needs the help of metadata, so managing metadata becomes a key of managing data.In order to improve the reliability of metadata managing, a fault-tolerant metadata management system is wanted. So we design a dual metadata server system for the cluster file system. In this system, there are two metadata servers and the reliability of metadata is provided by metadata mirroring, and the continuous metadata service is provided by failover: after the master server fails, the backup server takes over the metadata service. This system has the merits of the center metadata management: simple control, easy to implement and maintain, but overcomes the shortcoming of single point of failures. At the same time, it avoids the design and cost of keeping consistency of the distributing metadata management. We implement metadata mirroring, failure detecting, IP takeover and recovery in Linux kernel space, which is transparent to the applications. The last aim of this system is to extend its architecture to the multi-host to get greater fault-tolerant ability and come to high availability. In order to improve the efficiency of metadata server, an autoecious metadata storing and managing method is presented. Metadata of parallel file system are located in the kernel of the local file system and operate it by adding system call, which provides the compatibility of original system. When <WP=6>we apply this method on the metadata management of PVFS(Parallel Virtual File System), the test result indicates that the improved performance is about 5~8 times greater than the original method of it.更多还原