Research On Key Issues In Large-Scale Cluster File Systems

The cluster architecture has been matured as the main-stream architecuture for high-performance computers. The requirements of high-end computing applications challenge current I/O system of clusters. With the rapid improvements of processing power of CPUs and performance of interconnection networks, the I/O system of clusters becomes the bottleneck that limits the actual performance of clustered applications. As a key technology for solving the I/O problem of clusters, cluster file system is a very important research direction.The cluster file systems for large-scale clusters should have multiple data paths and multiple metadata paths in architecture. Metadata transactions are decoupled from file read and write operations. Large-scale network storages are used to provide multiple file data paths. A group of metadata servers are used to provide multiple metadata paths.We study a few key issues for cluster file systems of such architecture, including metadata distribution issue, metadata consistency and fast recovery issue, and issues in regards to PB-scale cluster file system. The main contributions of our work in this dissertation include:(1) Presenting an efficient storage space management policy for PB-scale storage space——a Bitmap-Extent hybrid policy. Unlike traditional file systems based on a block device, we present a logical storage space organization policy that decoupling physical storage space management from file system implementation. This policy enables file system to expand its volume and storage devices. Moreover, we present a hybrid storage space management policy that takes both advantages of bitmap policy and extent list policy. This policy substantially improves the efficiency of storage space management.(2) Presenting a dynamic metadata distribution policy called dynamic dir-grain policy (DDG). Metadata distribution determines metadata processing performance of decentralized metadata management with a group of metadata servers. The DDG policy is targeted at improve the holistic metadata processing performance, and takes both balance degree of metadata and file system hierarchy into consideration. The DDG policy distributes metadata according to distribution granularity which is a directory hierarchy defined by a triple. Our performance test results show that this policy is potentially more efficient than the other two types of policies under the workload that simulates real environments.(3) Presenting a distributed metadata processing policy that based on a simplified...