As part of the Computer Architecture Group at Northeastern University, we aim to develop scalable high performance storage systems for data-intensive workloads. This work targets Beowulf clusters and attempts to obtain parallelism in both processing and I/O access. We are supported by the CenSSIS Engineering Research Center.

As the performance gap between storage subsystem and the other components in the system continues growing, single I/O cannot provide enough bandwidth to meet the requirement of high performance computing. Parallel I/O technologies provide simultaneous disks accesses by creating multiple I/O channels between CPUs and disks. In order to effectively explore I/O parallelism, I/O needs to be parallelized at all system levels. Data contention at any level can severely limit the overall system performance.