| Cluster has become one of the most popular platforms for high-performance computing. As in traditional parallel computing systems, the I/O sub-system is a bottleneck to the overall system performance. One solution to alleviate the I/O bottleneck is to deploy a distributed storage system, which utilizes the aggregate bandwidth and capability of existing I/O resources on each cluster node, to provide high performance and scalable storage service for cluster computing platforms. The research of this dissertation concentrated on designing a high performance and high availability distributed storage system to improve I/O system performance. The system provided high performance by efficiently managing the aggregate cache space of a multi-level hierarchy, organizing file system data servers and iSCSI storage targets into a two level hierarchy with striping/parity techniques, and exploiting the potential of high speed network to reduce the RDMA registration cost. The system achieved high availability by overcoming single point of failure of metadata servers with a symmetric active/active metadata service. With a combination of various research approaches, including analysis using mathematical models, simulation using real world traces, prototype implementations of real systems running on Linux platforms, and experiments using real workloads, both high performance and high availability of a distributed storage system were achieved. The experimental results indicated that the average I/O response time was improved by up to 46% to 53% for various workloads, and the availability was increased to 99.98%; with less than 10% performance trade-off. |
