The Analysis And Research Of HPC Interconnection Network Topology

With the development of the transistor integration and storage technology and the rising demand for computing performance,the compute nodes' scale of high performance computing(HPC)system is continuously increasing.Besides,the interconnection network has presented its wide impacts on high-performance computer system.Those influenced aspects include the cost,the throughput rate,the energy consumption,the fault tolerance of routing,as well as the physical deployment complexity.This dissertation is unfolding its analysis from three major facets — the topology structure,the routing algorithm,and the physical deployment of the topology,all of which are closely related to the interconnection network.The specific content is as follows.First,this dissertation presents the new family of hybrid topology.Traditionally,the advantages and disadvantages coexist in both the direct and the indirect network.Thus,the hybrid interconnection network,combining the respective strong points of direct and indirect network,can obtain better performance cost ratio.Through the transformation and optimization on the basis of that,the dissertation presents a new hybrid interconnection network whose performance cost ratio shows a further improvement.Second,this dissertation presents the High-radix interconnection network routing algorithm based on Dragonfly.In this dissertation,based on Dragonfly(a topology network),the routing algorithm put forward by Dragonfly is discussed in detail from aspects like flow control and the solution to deadlock.Backed by that,a new routing algorithm-Greenred,a combination of channel buffer sharing and the recovery strategy of deadlock,is proposed,its good performance has been verified through simulation.Third,this dissertation analyzes and valuates the deployment of High-radix interconnection network topology based on Dragonfly.Benefited from the progress of interconnect chip switching technology,the message delay decreases increasingly in the interconnect chip.Hence,the delay of the interconnecting link between network nodes of future high-performance computer system will occupy a big proportion.Through the investigation into the existing physical deployments of topology and with the help of their specification,this dissertation based on Dragonfly evaluates the cabinet lines and the grid type physical deployments' effects on the performance of network.