【作者】 刘群； 冯丹；

【Author】 Liu Qun~1 and Feng Dan~2 1(Center of Network and Computing,Huazhong University of Science and Technology,Wuhan 430074) 2(School of Computer Science & Technology,Huazhong University of Science and Technology,Wuhan 430074)

【机构】 华中科技大学网络与计算中心； 华中科技大学计算机科学与技术学院；

【摘要】 在大规模分布式存储系统中,元数据高性能服务和扩展性已成为一个重要的研究热点.在元数据服务器(metadata server,MDS)中,将元数据分解为目录对象和文件对象.目录对象为定位性元数据,提供文件所在位置和访问控制;文件对象为描述性元数据,描述文件的数据特性.每个MDS负责所有目录对象和自身的文件对象,同时,以目录对象ID和文件名为关键字的Hash值作为局部元数据查找表的索引,通过Bloom Filter算法将每个MDS的局部元数据查找表压缩成一个摘要,这样既可利用MDS中Cache,提高Cache的命中率,减少磁盘I/O次数,动态扩展MDS,又能够实现快速的元数据查找.更多还原

【Abstract】 In large distributed storage systems,it is critical to research on high performance metadata service and scalability.In the metadata server(MDS),metadata is divided into directory object and file object.The former is locating metadata,providing the file location and access control;the latter is descriptive metadata,depicting the data of file.Each MDS is responsible for the whole directory objects and its own file objects.Meanwhile,the hash value of key including directory object ID and file name is used as an index to the local metadata lookup table(LMLT).Through Bloom Filter algorithm,the LMLT of each MDS can be compressed into a brief.Not only does it make full use of the cache,improving the hit rate of cache,reducing the number of disk I/O,and extending MDS dynamically,but also achieves rapid metadata finding.更多还原