Abstract Dr. Behrooz Parhami

 Robustness Attributes of Interconnection Networks for Parallel Processing

Large-scale parallel processors, with many thousands or perhaps even millions of nodes and links, are prone to malfunctions in their constituent parts. Thus, even under a best-case scenario of prompt malfunction detection to prevent data contamination, such systems tend to lose processing and communication resources over time. Whether they can survive such inevitable losses is a function of the way computational tasks and their attendant information exchanges are organized and on certain intrinsic properties of the interconnection topology. This talk begins with an overview of robustness features, as they pertain to interconnection architectures. Next, a number of well-known interconnection structures are viewed from the robustness angle. Finally, it is shown how large-scale hierarchical or multilevel networks can be synthesized for robustness, while keeping implementation cost, power dissipation, and routing overhead in check.