Research And Implement Of The Metadata's Fault Tolerance In Cluster File System

Nowadays, high performance computers and high performance servers make use of large scaled cluster architecture gradually. Cluster file system has become a key point in current research about the cluster, and the PVFS(Parallel Virtual File System) can afford a virtual large storage's uniform access interface and high I/O bandwidth. Owing to the file data spreading around the cluster, locating a file needs the help of metadata, so managing metadata becomes a sticking point of managing data. Especially, research on architecture and crucial technologies of metadata fault tolerance based on cluster system is very important. To solve the bottleneck in PVFS's metadata management, a high availability cluster file system's metadata fault tolerance system MDFTS(MetaData Fault Tolerance System) which bases on PVFS, has the character of HA in metadata management, and system fault diagnosis and checkpoint recovery is subsequently done. For the coherence in metadata, a serverless architecture model based on all the data and metadata is adopted because they can be storaged in all place and can be dynamicly transferred; The associative mode of the disc metadata's journal structure and the memory metadata's journal structure can turn away the timing waste in scanning the enormous file system by fsck; Afterwards a intercalating checkpoint algorithm and a rollback recovery algorithm based on the metadata fault tolerance in cluster file system is implemented; A fetched stochastic process model is setup to analyze availability of this system, it can reduce the inspecting fault time and then they support high availability. With modifying metadata's data structure and system calls of data and metadata operation, the fault tolerance system implement in the application layer of operation system, so all the data nodes and metadata nodes can reciprocally team up and unite schedule. The test results indicate that the metadata fault tolerance system can examine different failure in the fault injection, and it can do the system and application's switch and come back.