分布式文件系统名字空间管理-计算机系统结构专业论文.docx

华中科 技大学硕士学位论 华 中 科 技 大 学 硕 士 学 位 论 文 II II Abstract With the rapid development of internet, more and more materials need to be stored in the disk. The amount of global information grow more than 30 percent over a year, and the data that needs to be stored to the disk grow more than 114 percent over a year. Traditional file system like ext2, ext3 and net file system (NFS) can’t feed our needs, At this moment, distribute file system come into our sight. Distribute file system can store 1 PB to dozens of PB data, it can feed us well. Traditional file system typically uses 3 party architecture, and nameserver is the brain of all nodes. Nameserver typically faces C10K problem , so how to design the nameserver to adapt it to high concurrency problem becomes more and more important. This paper design and implement the distribute file system named RaccoonFS. The paper also analyzes and solves the problem that nameserver behaves bad after high concurrency because of the lock, it proposes a new namespace management method. RacoonFS uses B+ tree to manage namespace metadata, and implement copy on write and multiversion concurrency control on the B+ tree, so the read operations and write operations separate from each other and write operations can do concurrently. The things above contribute much to the performance of RaccoonFS. In the test part, we choose three kinds of object to test : the hadoop distribute file system, the B+ tree namespace management that has implemented copy on write , the B+ tree namespace management that has implemented copy on write and multiversion concurrency control. The result show that the B+ tree namespace management that has implemented copy on write and multiversion concurrency control behaves better than the B+ tree namespace management that has implement copy on write , and the B+ tree namespace management that has implement copy on write behaves better than hadoop distribute file system. The test also shows that copy on write and multiversion