Cluster File System Server Key Technologies

As a key component of cluster system, cluster file systems must provide high performance and scalable file service to solve the I/O bottleneck of cluster system. Research in cluster file system is important to high performance computing. With the design and implementation of DCFS (Dawning Cluster File System), this dissertation deeply discusses the key technologies of a cluster file system and studies problems in designing the servers of a cluster file system. The contributions of this dissertation include:1. This dissertation firstly compares architectures of cluster file systems and presents a new architecture called multi file system volumes which is scalable, manageable and flexible. This dissertation analyses the structure of storage servers,presents a model of a network disk stripe group and discusses the factors that impacts the read/write bandwidth of a network disk stripe group. A metadata placement policy is proposed for user to select suitable granularity.2. This dissertation studies two key issues of directory operations: directory cache and large directory optimization. The research showed that LOOKUP and READDIR directory cache on DCFS metadata servers have different access characteristics with that on client nodes. A new directory cache management method is presented according to these access characteristics. Experiments showed that this method has higher directory cache hit ratio than those which exploit LRU, LFU and FBR cache replacement algorithms. This dissertation introduces a large directory optimization algorithm LMEH (Limited Multi-level Extendible Hash) which is presented by author and a member of our research group. The experiments indicated that the throughput of LMEH for large directory is 1.97 times as Linear algorithm on average.3. This dissertation designs a metadata consistency protocol according to the metadata placement policy and cache management in DCFS. This protocol ensures the consistency of metadata in a cluster file system and has acceptable cost.4. Design and implement DCFS on the Dawning 4000L Cluster. Present methods to evaluate a cluster file system and define scalability of read/write and metadata operations. Compared with PVFS, DCFS exhibits higher aggregate bandwidth except read bandwidth for small files. The read bandwidth of DCFS for small files is 19% less than PVFS and the bandwidth of DCFS is 44.4% higher than PVFS in other experiments on average. The metadata throughput of DCFS is 7.391 times as PVFS on average. And the overall response time of DCFS is 18.2% as PVFS in the synthesis workload experiment.