一种高性能hdfs存储平台的分析与实现-analysis and implementation of a high performance hdfs storage platform.docxVIP

摘 要近年来，云计算成为了互联网行业最热的话题之一。由于其高可靠性，高吞 吐量及高可扩展性， Hadoop 分布式文件系统 HDFS 成为了主流的云计算平台的 底层存储系统，政府、高校、科研机构或者企事业单位均先后使用 HDFS 作为其 集群的存储平台。然而，尽管 HDFS 有着良好的表现，其在性能上仍然存在明显 的问题。比如 NameNode 的单点失效，数据存储空间浪费，小文件存储效率低等。 现阶段的国内外研究往往仅针对上述单个问题作出分析与研究，没有系统地将 HDFS 的性能问题进行整合与改进。本文针对这些问题，设计与实现了一个高性能的 HDFS 存储平台。平台采用内 存与硬盘的双重元数据备份机制，在保证 NameNode 高可用性的同时，降低了集 群因 NameNode 宕机失效所需的恢复时间；在数据的存储上，平台采用基于 Erasure Code 的编码存储技术，对文件分块后的 block 数据进行编码存储，降低了集群的 副本数量，提高了存储空间的利用率；平台也对 HDFS 存储小文件效率低的问题 作出了改进，采用 Hadoop 提供的 HAR 文件格式，将小文件打包上传，减轻了 NameNode 的元数据存储压力，也同时提高了小文件的读写速度。文章最后通过一系列的实验来验证高性能 HDFS 存储平台在可用性，存储效率 上的优势，满足了系统的设计要求。关键词：云计算HDFS 高可用性 Erasure Code 小文件存储AbstractIn recent years, cloud computing has become one of the hottest topics in the Internet industry. Because of its high reliability, high throughput and high scalability, the open-source project of the Apache Foundation Hadoop Distributed File System, HDFS,has become the mainstream underlying storage system in cloud computingplatforms，andGovernment, universities, research institutions or enterprisessuccessively use HDFS as its clusters’ storage platform. However, despite HDFS has a good performance, it still has obvious problems.Such as the single point failure of NameNode,the wasted in data storage and the low efficiency of small files storing.Now,researches often aims at only one of these problems,therefore ,they don’t integrate and improve the problems of HDFS systematically.To solve these problems, we design and realize a high-performance HDFS storage platform. The platform introduce both memory and hard disk metadata backup mechanism, as a result, it ensures the high availability of NameNode and reduce the cluster’s recovery time caused by the NameNode downtime failure at the same time. In the data storage, the platform introduce encoding storage technology based on Erasure Code, it encodes and storages the block data after the file chunked, reducing the number of replication of the cluster and improving the storage space utilization. The pla