Traffic Measurement And Scheduling Based On Flow Characteristics In SDN

With the rapid development of Internet technology,the scale of the network continues to extend,and the service types are increasingly complex.Software Defined Network,which is able to decouple the control plane and data-forwarding plane with strong programmability and the global network view,has been widely used in data center network.However,congestion can easily occur when the frequent communications between different network entities generate a large number of flows,having a serious impact on the whole network.Therefore,it is important to make reasonable scheduling strategies for flows.In this thesis,different scheduling strategies are proposed for the detected elephant flows and mice flows in the Software Defined Data Center Network,in order to increase the link utilization and reduce the transmission delay.The main contributions are as follows:An elephant flows scheduling strategy based on Deep Reinforcement Learning(DRL)is proposed to solve the problem that the optimization performance of conventional traffic scheduling is very sensitive to environmental changes.Firstly,the scheduling of elephant flows is modeled as a Markov Decision Process(MDP),and the optimization goal is to maximize the total transmission bandwidth of the entire network.Secondly,in order to better represent the state in the network,the remaining bandwidth of multiple paths between forwarding nodes is used as the input of MDP.Compared with the link remaining bandwidth,it is easier to extract the overall utilization between end-to-end paths in the network topology.Then,in order to reduce the crad of action space,the action is designed to select one from multiple available paths.Finally,in order to reduce the crad of state space caused by rapid environmental changes,an elephant flows scheduling strategy based on DRL is proposed.Simulation results show convergence under different superparameters and demonstrate the better performance of the proposed algorithm compared with other scheduling algorithms.A mice flows scheduling strategy based on improved particle swarm optimization(PSO)was proposed to reduce the processing delay and increase the convergence.Firstly,the scheduling of the mice flows is modeled as a multi-commodity flow problem,and an aggregation mecha-nism is given to transmit the aggregated mice flows over multipath.Then,PSO with low algorithm complexity is used to solve the problem,which reduces the processing delay and satisfies the delay-sensitivity requirement of mice flows.Finally,the particle search success rate is exploited to improve the convergence of the conventional PSO algorithm,which leads to a better solution,and the scheduling strategy of mice flows is obtained.The simulation results show that the proposed strategy can effectively improve the network throughput and link utilization with guaranteeing the average delay.