Research On The Key Technology Of Bandwidth Allocation In Data Center Backbone Networks

Data center has become an important and fundamental infrastructure of modern information and communication. By highly integration of vast amounts of computing, storage, network and other physical resources forming virtual pool of different resources, DC uses virtualization technology to realize the high-efficient sharing of resource. In order to improve the performance and reliability of DC, they are usually located in far away from each other geographically all over the world, interconnected with each other through the high speed backbone networks. On account of the network’s huge construction cost, and also the data packet loss is unacceptable, it is very important to make efficient and reasonable use of the resources of DC backbone network’s bandwidth to ensure the data stream transmission QoS guarantee. However, traffic engineering technology is typically deployed to schedule traffic for the purpose of load balance, and the bandwidth is shared in the best effort way. In this way, the bandwidth utilization is highly improved, but the QoS guarantee level is limited for the lack of fairness consideration. In the existing DC backbone network, main problems of bandwidth allocation are as follows:(1)The architecture level: DCs are interconnected in the distributed way, leading to that it is difficult even impossible to obtain the global flow view, meanwhile the optimization of bandwidth resource allocation is difficult too.(2) The node processing level: It is usually to obtain bandwidth requirements real-timely based on flow measurement technology and bandwidth is reserved accordingly. The cost is high and the relevance is lacked. And also, it is too coarse to al ocate bandwidth for the aggregate flow and lack of fairness.This study focuses on the research on the key technology of DC backbone network bandwidth allocation. First of all, we propose a traffic prediction based bandwidth requirements obtaining mechanism of elephant flow on the basis of the Software Defined Network(SDN) centralized control architecture, obtaining flow information timely and accurately. The forwarding plane and control plane separation mode of SDN make it easily get the global view of the network, and the bandwidth requirements of the stable elephant flow can be predicted using the traffic’s historical information, reducing the obtaining error of bandwidth requirements. Secondly, for the problem of the too coarse granularity of current bandwidth allocation policy and the lack of fairness, this study proposes a bandwidth allocation policy of elephant flow based on cooperative game theory, balancing the tradeoff between the efficiency and fa irness of bandwidth allocation. Finally, for the bursty mouse flow, to adapt to the traffic patterns dynamically, we propose a bandwidth allocation policy of mouse flow based on the Richards model, achieving the goal to fast match the bandwidth requirements. The researches in the dissertation are as follows:1. A bandwidth requirements obtaining mechanism of elephant flow based on traffic predictionThere is a small amount of elephant flow which occuping most bandwidth resources in DC backbone network. It is relativly stable and can be predicted according to the flow view, namely historical information of the bandwidth requirements. After obtaining the prediction results, we can allocate bandwidth accordingly. Due to the complex characteristics of network traffic, such as self-similarity and heterogeneity, we design a bandwidth requirements obtaining mechanism combining the long and short term prediction algorithm to improve the prediction accuracy. Theoretical analysis and simulation results show that this mechanism can improve the accuracy of bandwidth al ocation effectively when compared with other mechanisms.2. A bandwidth al ocation policy of elephant flow based on cooperative game theoryFor the bandwidth allocation problems of elephant flow, we model the problem of multiple flows’ contending the share bandwidth as a cooperative game, according to the obtained bandwidth requirements through prediction algorithms proposed above. Different flows compete for bandwidth and the overall utility is maximized. The optimal bandwidth a llocation policy, called Cooperation Game based Bandwidth Allocation(CGBA), is obtained through searching the Nash bargaining solution(NBS) of the game using the centralized bandwidth allocation algorithm which we design. Theoretical analysis and simulation results show that NBS is the best solution to allocate bandwidth, balancing the tradeoff between the minimum bandwidth guarantee and the fairness. The method proposed in this study not only provides high QoS guarantee level, but also improves the overall utilization of bandwidth resources.3. A bandwidth al ocation policy of mouse flow based on the Richards modelFor the bandwidth allocation problems of mouse flow, we prpose a closed- loop feedback control system to allocate bandwidth real-timely, taking account of its highly dynamic characteristics. We can make a smooth response to bursty mouse flow, designing the link feedback control factor based Richards model as the feedback of the system. According to the time delay sensitivity of flows, different flow rate growth curves can be derived of depending on different allometric parameters, which meet the QoS requirements of delay sensitive flows. The simulation results show that the policy proposed in this study can effectively guarantee the low latency requirements of delay sensitive flows, providing high efficiency, at the same time ensuring high fairness of bandwidth al ocation.