Research On Fault Tolerant Technology Of New Generation High-speed Interconnection Networks

In the research and development process of high performance computer (HPC)systems, the increasing speed of multi-core processor makes the bottleneck issue of datatransmission between processors more severe. The utilization of high frequencytransmission technology can improve the quality of transmission performance, itreduces the inherent reliability among different components as well. Meanwhile, withthe expanding scales of HPC systems and the increase of HPC’s interconnectionnetwork components, the inherent reliability in systems becomes more unstable.Therefore, it becomes an essential issue in HPC system about how to promote thesystem’s reliability by utilizing the succinct and efficient fault-tolerant technique. Basedon previous researches on interconnection networks of TH-1A HPC system, theresearch group proposes several key techniques on the new generation high-speedinterconnection networks.This thesis focuses on the research of fault-tolerant technique in high-speedinterconnection networks, whose main ideas are summarized as follows:1. We analyze the topologies, routing algorithms and flow control strategies ininterconnection networks on current HPC interconnection networks and sum up thefault-tolerant techniques utilized in those systems.2. Propose the design of a micro architecture in terms of routers which offersadaptive routing algorithms support and is available to various kinds of topologies.3. Put forward two types of adaptive fault-tolerant routing algorithms in terms ofnetwork architecture based on channel sorting and channel escaping respectively;aiming to establish a high dimensional network topology combined with3D-Torus andfully connected networks.4. Study the topology discovery, routing algorithms and path distribution relatedalgorithms by analyzing the InfiniBand network management protocol.5. Implement and verify the router structure, topological structure, routingalgorithm and network management protocol in the interconnection network based onthe OMNeT++simulation platform and make corresponding simulation evaluation andperformance assessment.