Routing And Resource Allocation In Data Center Networks

With the increasing of applications in the network and the development of “big data”, tranditional single server based computation cannot satisfy the users’ requirement and more and more users would like to outsource their tasks to the cloud or data center networks(DCNs). Furthermore, the emergence of pay-as-you-go charge model provides a plus to such trend. The increase of users leads to a significant traffic growth in the DCNs. To deal with such traffic growth, provide users a high performance network and maintain the services, the operator should optimize the network performance. In this dissertation, we study how to optimize the performance of DCNs through traffic routing and resource allocation. To this end, we mainly focus on the following four topics:1. Routing and scheduling in DCNsIn DCNs, a job is divided to be multiple tasks and executed on a number of hosts. Accordingly, the flows to the same job have semantics and are depend on each other. In this topic, we study how to optimize the completion of a group of flows that have semantics(called coflow). To minimize the average coflow completion time in the DCNs, we propose a system named RAPIER that can jointly optimize traffic routing and scheduling in the networks. In addition, we implement the prototype of RAPIER. Both the experiment on small-scale testbed and the C++ based simulation in large-scale networks show that RAPIER can greatly reduce the average coflow completion time the DCNs.2. Multi-objective optimization in DCN traffic engineeringWhen we optimize the traffic routing in the DCNs(traditionally, we call it is traffic engineering), we usually have multiple optimization objectives, such as load balance and energy efficiency. However, there is no traffic routing scheme that is best for both objectives, and hence we should consider how to derive a fair tradeoff between such two objectives. In this dissertation, we will take the tradeoff between load balance and energy efficiency as an example to present how to solve a multi-objective optimization problem in DCN traffic engineering in game theoretical perspective. To this end, we formulate a Nash bargaining based threat value game to negotiate the traffic routing in the DCNs. Both mathematic analysis and simulation show that our method can derive a fair tradeoff between load balance and energy efficiency. In addition, our work can provide insights to the general multi-objective optimization problem.3. Jointly optimizing network topology and virtual machine placementIn the DCNs with static topology, we usually find some part of the network is underutilized. To enhance the network utilization, some researchers propose dynamic topology to adapt to the traffic. On the other hand, the operator can determine the virtual machine placement scheme to optimize the network traffic scalability. In our work, we jointly optimize both network topology and virtual machine placement with the expectation to derive a better network performance. We first formulation such joint optimization problem as a mixed integer linear programming problem and decompose it as two subproblem based on Lagrange relaxation decomposition. After that, we design efficient heuristics for both subproblems. Extensive simulation results show that joint optimization scheme can greatly enhance the network traffic scalability compared with only optimizing network topology or virtual machine placement.4. Dynamic topology management in DCNsThough dynamic topology can enhance the network scalability and make network carry more traffic, convert the topology setting to a new one may result in large amount of traffic loss. More importantly, it may prevent the deadline aware flows from meeting their deadlines. For this reason, we propose a progressive topology reconfiguration method that can not only maintain the network connectivity during the topology reconfiguration and guarantee the performance of deadline aware flows, but also reduce the traffic loss caused by the topology reconfiguration.