Research On The Architectures Of Data Center Networks

Traditional data center networks employ three-level tree structure. Core level andaggregation level usually utilize high-priced switches to connect the ones at the edgelevel. However, more servers are added and connected to those switches with the datacenter network scaling. High oversubscription makes higher-level switches thebandwidth bottleneck easily. And they perform poorly in terms of fault tolerance anddisaster recovery because of the lack of sufficient redundant routing paths. But, we canbenefit a lot from reducing the data center network level number and making it as flatas possible. Better performance and considerably lower capitalized cost for laying out,operation and maintenance of data center networks can be obtained. Interconnectionnetworks have been popular and successful in the domains of parallel computersnetwork and supercomputer research. So, we consider migrating the interconnectionnetwork designs and solutions to the construction of data center networks.At first, the development of cloud computing is summarized and problemsbrought by cloud computing are listed when designing new data center networks. Then,various solutions from network equipment vendors and research efforts from academiaare introduced for flatting the data center networks. After that, a connection-oriented,virtual hierarchical Mesh-interconnected data center network is proposed and thecorresponding subnet agent selection rules aiming to reducing the size of routing tableis designed by us. We also analyze the impacts of different subnet agent locations onthe network deadlock avoidance and performance. And an efficient routing algorithmwith deadlock-free is proposed to guide packets’correct and fast transmission.Furthermore, the hierarchy mechanism for Mesh-of-Torus hybrid interconnectionis designed to support large-scale data center networks. Without using virtual channels,a load-balanced, deadlock-free and efficient routing algorithm for Torus is proposedbased on our unique port assignment algorithm. Then, server and link models aredefined in OPNET, and a series of simulations are carried. Simulation results make itclear that Mesh-of-Torus hybrid interconnection has lower end-to-end latency, highernetwork throughput than virtual hierarchical Mesh.