Storage subsystems have long been the bottleneck of computer system performance. In recent years, developments in object-based storage systems and parallel file systems have demonstrated the ability to scale aggregate throughput for large data transfers in network storage systems. Yet, there are application workloads that are not amenable to parallel I/O classes and as such do not benefit from these recent advances. Such applications include those that are dominated by small data transfers and heavy file system transactional usage as well as I/O patterns that are characterized by reduction operations. The PIs present an approach to extract scalable performance by inserting active elements in the network. These active storage networks can draw on knowledge of overall data layout as well as the ability to process all data that is retrieved from the actual storage node. The PIs plan to incorporate reconfigurable components in the network to perform these operations in hardware.
