Research On Intelligent Traffic Scheduling Technology For SDN

With the development and growth of the Internet,large enterprises,schools,organizations,and small individual users use the network all the time,the data center network has become an important part of the Internet infrastructure.The end-to-end high-bandwidth and low-delay mixed traffic generated by massive server communication in data centers adds difficulties to the balancing of data flow scheduling.The proposal of Software Defined Network architecture opens a new space for solving the problem of traffic scheduling load.The unique nature of SDN separates control and forwarding,realizes the traffic control strategy based on the centralized network view,so that the controller and the switch do not need to communicate frequently.Reduce bandwidth usage for controller communication.However,most of the existing traffic scheduling methods in data centers use static hash method to evenly distribute the data flow to multiple equivalent paths with the same load,which is not suitable for the transmission of elephant flows.At the same time,it increases the queue delay for the mouse flow behind the elephant flow,which is easy to block the network and cannot solve the conflict of network resources.In order to solve the mixed traffic scheduling problem of elephant flow and mouse flow,the following research work is carried out in this paper:(1)The Fat-Tree network topology is introduced,the left and right parts of the FatTree topology have the same bandwidth,the uplink bandwidth is equal to the downlink bandwidth,and the link bandwidth varies with the scale of the network,so it has high network throughput,and the fault tolerance of the network is increased by multiple parallel paths between host pairs in different transmission areas.(2)A traffic scheduling based on ant colony algorithm is proposed,which sets different pheromone concentration for different traffic types,and adjusts the pheromone evaporation factor of ants to avoid reducing the path concentration due to too large pheromone evaporation factor,increasing the search range and slowing down the convergence speed.If the pheromone concentration is too early,the global search ability will be reduced.Enter the local optimal solution in advance.A Mininet simulation platform was built,Ryu control was installed on the simulation platform,and different traffic sending modes were set to simulate network traffic to verify the operation effect of the algorithm.At the same time,compared with ECMP and Hedera traffic scheduling algorithm,the simulation results show that the algorithm has better network performance under the same conditions.It has higher average network throughput,lower average transmission delay,and effectively improves the link utilization.(3)Try to combine the deep reinforcement learning algorithm with SDN architecture,propose a traffic scheduling algorithm based on DQN,set the reward value according to the traffic characteristics,and introduce the DDQN algorithm to update the Q value to make full use of the autonomous training of deep reinforcement learning.The SFlow traffic monitoring software is installed on the Mininet simulation platform to monitor the network transmission traffic in real time and monitor the large flows,which reduces the number of interactions between Open Flow switch and controller.Finally,compared with the traditional traffic scheduling algorithm in average link throughput,average link delay and link utilization,the results show that the proposed algorithm makes full use of link resources,increases the average link throughput,reduces the average transmission delay,and has better network performance.