Theoretical Support for Efficient Network Discovery and Reconfiguration Techniques

University Of Southern California, Los Angeles CA

Investigators

Abstract

This research focuses on the development of a sound theoretical framework for the design of efficient network discovery and reconfiguration techniques that are deadlock-free and generically applicable to state-of-the-art interconnection networks used in future multiprocessor servers, network-based computing clusters and distributed storage systems. The theory should have wide applicability and serve as an authoritative basis on which deadlock freedom of reconfiguration processes for arbitrary networks can be formally proved. As part of the theoretical framework, a general methodology will be developed that will allow the power of the theory to be realized in the straightforward design of new and more efficient techniques. The techniques derived from the theory and methodology will provide alternative ways of increasing network reliability, availability, and serviceability (RAS) in the presence of unconstrained fault patterns as well as in the presence of voluntary changes to a network configuration due to resource re-partitioning, communication-aware algorithm/process re-mapping, data migration, etc. The techniques will not impede the injection, transmission, or delivery of user packets during discovery and reconfiguration processes but, rather, will remain relatively transparent to the user and/or system administrator. Proof of concepts will be pursued by simulating the developed techniques and applying them to state-of-the-art switch-based system-area and storage-area networks amenable to their implementation. This research ultimately could result in new opportunities for dependable network architectures to off-load some of the QoS handling to lower network layers while, at the same time, providing enriched RAS features and higher performance.