The Research On Scalable Metadata Service Of Parallel File System For High Performance Computing Systems

With the increasing of the computational abilities of supercomputers, problem sizeand complexity targeted by applications has scaled, which required higher performance ofI/O subsystems. While the throughput of single metadata server limited the performanceof the parallel file system in high concurrent access and high-frequency filecreating/deleting scenarios. A high scalable metadata service based on metadata delegation,applied to I/O forwarding architecture, is proposed. Depending on job scheduling system,it distributes the requests for metadata from the file system to multiple metadatadelegations to accelerate metadata operations.Parallel I/O is a typical scenario in high performance computing. It can becharacterized to two categories file per process mode and shared file mode. The formerrenders performance on metadata, and we mainly focus on this I/O scenario. The keycontribution of this dissertation is implementing Meta-Data Delegation Service(MDDS) inLustre file system and proposing a scalable distributed metadata management scheme.MDDS is based on Lustre Cluster MetaData(CMD) design. We use loose coupling to keepthe high availability of the cluster, organize the MDDS namespace by directory subtree toavoid the complex and inefficiency of distributed atomic operations introduced bycross-node operations, and developed a metadata migration mechanism to avoid objectsdata moving between data servers. Also, the system is allowed to add and remove aMDDS server dynamically. Two job-scheduling strategies have been proposed---one jobscheduled on single MDDS and jobs sharing multiple MDDSes. The former is suitable fortraditional job’s I/O access patterns, and can avoid competition of metadata operationsamong jobs; while the latter is able to distribute the metadata’s operations to multipleMDDSes to achieve load balancing of the requests for metadata inside a job.We analyzed the performance of MDDS on116storage servers, and simulated theapplications’ metadata access load in I/O forward architecture to evaluate the performanceof the two schedulers. The initial experimental results show that quasi-linear scalable metadata performance is achieved by MDDS, and even show better scalability than LustreCMD in large-scale cluster. The two job-scheduling strategies distribute the applications’metadata access load effectively, and overcome performance bottlenecks in accessing filemetadata in HPC.