Research On Traffic Routing In Data Center Networks

In recent years, there are more and more computationally intensive applications such as indexing Web content, online map, and scientific computing. They are based on cloud computing. As important infrastructures of cloud computing, data centers have received a lot of attention from academia and industry. Data center network is the core part of the data center. It interconnects a large number of servers and storage devices and it provides a variety of services and applications for various private and corporation systems.With the rapid growth of the communication traffic inside the data center network, the bandwidth requirement of the data center network is increasing rapidly. Conventional data center networks are typically constructed based on tree topology by using electronic packet switches. However, they suffer the problems such as high cost, high power consumption, poor scalability and bandwidth bottleneck. In order to address these issues, several data center network architectures have been proposed recently. Clos network topology is introduced into data center network design to solve the low bisectional bandwidth problem. Besides, the researchers introduce the optical technology into the design of the data center network architecture and hybrid optical and electrical data center network architectures were proposed. In the data center, the flow size distribution is not uniform. The traffic follows a heavy tailed distribution and it is bursty. Conventional Internet routing mechanisms are not able to adapt to the new data center network architectures and data traffic patterns. To address the above problems, we need to design new routing schemes for data center networks.For the CLOS Data center network architectures, we propose a Software-defined Hybrid Routing(SHR) scheme. The SHR differentiates data flows by statistical calculations. The elephant flows utilize the adaptive routing algorithm while the mice flows use oblivious routing algorithm. SHR extends the OpenFlow scheme by offloading some basic functions such as flow statistical detection and mice flow forwarding to switches to reduce the switch-controller interaction overhead. Performance evaluations of SHR are carried out using the fat-tree and VL2 network topology. Results show that SHR can effectively increase network throughput and reduce the flow dropping rate as well as packet delay compared with the traditional ECMP algorithm.On the other hand, for the HWRE data center architecture, we propose a new software-defined traffic routing mechanism. The controller collects the data traffic and construct the ToR-to-ToR traffic matrix periodically. It calculates the forwarding probability of each equal cost path. The flows are forwarded according to the forwarding probability. The electrical switching link is used to handle bursty traffic. Synthetic flow model is used to analyze and evaluate the performance of this mechanism. The results show that the proposed mechanism can efficiently improve network throughput and reduce flow loss rate in the network compared with traditional Random Early Detection algorithm.The two routing schemes proposed in the thesis improve the performance of data center networks. They provide technical foundations for future data center network design and applications.