The Design And Implementation Of Traffic Detection And Dynamic Scheduling System Based On SDN Architecture

There is huge traffic transmission between servers in the data center.According to the characteristics of traffic,it can be divided into throughput-sensitive elephant flow and delaysensitive mice flow.The two kinds of traffic compete for resources such as queue cache and link bandwidth,which requires dynamic scheduling.Different from the traditional network architecture,software defined network(SDN)can centrally obtain global network resources and facilitate the formulation of dynamic scheduling strategies.Therefore,in order to improve the quality of service for elephant flow and mice flow,this thesis designs and implements a traffic detection and dynamic scheduling system based on SDN.The main work is as follows:(1)This thesis proposes an elephant flow scheduling algorithm,which is called PERDDQN(prioritized experience replay double deep q-network)for short,to generate an elephant flow path from client to server for the detected elephant flow.The elephant flow path decision is approximated as a Markov decision process.Based on the DQN(deep q-network)algorithm,the bias of action value is corrected by adding the double DQN mechanism,and the priority experience playback mechanism is added to help learn important samples.Experiments show that the PER-DDQN elephant flow scheduling algorithm proposed in this thesis is superior to ECMP algorithm,Hedera algorithm and DQN algorithm in throughput performance.(2)This thesis proposes a mice flow scheduling algorithm,called PSO-ALM(particle swarm optimization adding lazy mechanism)for short,to generate distribution paths for clustered mice flows.Through the global mice flow scheduling model,the global link load balancing is realized;Using the Open Flow group table function,make full use of the equivalent shortest paths from the source edge switch to the destination edge switch;On the basis of particle swarm optimization(PSO)algorithm,by adding lazy particle mechanism,the total income of bandwidth allocation of each path is maximized.Experiments show that PSO-ALM mice flow scheduling algorithm proposed in this thesis is better than ECMP algorithm and Mice Trap algorithm in time delay performance,and PSO-ALM algorithm is more suitable for the mice flow scheduling model than ant colony algorithm.(3)Based on the above research,a traffic detection and dynamic scheduling system based on SDN is designed and implemented.The advance flow definition is carried out through the active discovery host method,the traffic visualization and elephant flow detection are realized through the sFlow-RT technology,and the equivalent shortest paths between the client and the server are obtained through the proposed Floyd-DFS method.To sum up,this thesis implements a data center internal traffic scheduling system that comprehensively considers elephant flow throughput and mice flow delay.The test results show that the system designed and implemented in this thesis is effective,and can comprehensively consider the elephant flow throughput and mice flow delay.