Systems Support for High Performance I/O on Shared Storage Clusters

Conventional solutions for I/O have attempted to provide hardware and software parallelism via RAIDs or parallel machines/supercomputers. However, the problems associated with cost, scalability, and/or accessibility of these environments make them unattractive for widespread usage. This research addresses this important deficiency in high-performance I/O support, by proposing a shared storage system using an off-the-shelf cluster of workstations, disks, and networks. The proposed research goes beyond current state-of-the-art in I/O support for clusters and examines a broad spectrum of issues related to I/O software on clusters, that include application-directed, compiler-directed, and runtime system-directed optimizations. These optimizations are crucial to reduce/hide the latencies to different levels of the I/O hierarchy which will help accelerate the deployment of clusters for I/O-intensive applications.