A failure recovery mechanism for distributed metadata servers in DCFS2

Distributed metadata servers are required for cluster file system's scalability. However, how to distribute the file system metadata among multiple metadata servers and how to make the file system reliable in case of server failures are two difficult problem. We present a journal-based failure-recovery mechanism for distributed metadata servers in the dawning cluster file system-DCFS2. The DCFS2 metadata protocol exploits a modified two-phase commit protocol which ensures consistent metadata updates on multiple metadata servers even in case of one server's failure. We focus on the logging policy and concurrent control policy for metadata updates, and the failure recovery policy. The DCFS2 metadata protocol is compared with the two phase commit protocol and some virtues are shown. Some results of performance experiments on our system are also presented.