Research On Optical Interconnected Data Center Networks For Cloud Computing

Rapid development of cloud computing applications poses new challenges to cloud computing data centers(DCs). Data center networks(DCNs) interconnecting a large number of servers in DCs have become one hot research topic in both academia and industry.Traditional DCNs are based on electrical packet switches with hierarchical tree topology, but they suffer the problems such as high power consumption, high latency, poor scalability and bandwidth bottlenecks. The improved electrical switching DCNs increase link redundancy compared with traditional DCNs, and they solve the problem of low bisection bandwidth but they still suffer the problems such as high power consumption and high latency. Hybrid DCNs and all optical DCNs introduce Micro Electric Mechanic System(MEMS) based optical circuit switches to take place of electrical switches, but they introduce long-time network configuration and complex network management. Therefore, designing a new data center network architecture to improve network performance is a research problem. This paper designs a new data center network architecture from two levels, inter-rack data center network and intra-rack data center network.The inter-rack network connects the top of rack(To R) switches. This thesis proposes a hybrid optical wavelength routing and electrical packet switching based inter-rack network(HWRE). Software defined network technology is used to decouple the data plane and control plane of the network. The data plane is a hybrid network with an Arrayed Waveguide Grating Router(AWGR) and an electrical packet switch, which provides reliable high-speed switching channels for inter-rack traffic in DCNs without any configuration.The intra-rack network replacing To R switches connects servers in the edge layer. This thesis presents a software defined passive optical intra-rack network based on optical couplers in DCs. The data plane utilizes passive optical couplers to interconnect servers in the same rack, and the intra-rack traffic are split by optical couplers and broadcast to all other servers in the same rack. To solve the collision problem in the same rack, we propose a software defined MAC(SD-MAC) mechanism. This mechanism utilizes the Max-Min Fair Share bandwidth allocation algorithm to enable fair and efficient bandwidth allocations for servers in the rack.The performance of the proposed network and protocol is evaluated in two aspects, inter-rack network and intra-rack network. First we theoretically derive the simulation model, and then evaluate the network performance through network simulations. Inter-rack network performance simulation results show that the proposed inter-rack network HWRE has the best end-to-end packet delay performance compared with other typical DCNs. The light source plan 3 of HWRE with multi-carrier light source(MCLS) has the highest network throughput and the lowest wavelength request blocking probability among the three light source plans. Intra-rack network performance simulation results show that intra-rack network cost and power consumption are much lower than those of the electrical To R switch solution, and the intra-rack network has lower packet delay compared with the electrical To R switch solution. The network saturates only when intra-rack traffic ratio is over 80% and the offered load is very high. The Max-Min Fair Share bandwidth allocation algorithm can achieve better network throughout than the traditional fixed TDM allocation method.The simulation results show that the proposed optical interconnected data center network has advantages in terms of cost, power consumption and network performance. This research lays a solid technical foundation for the development of future cloud computing data center networks.