Research And Application Of Deterministic Network Load Balancing System Based On SDN

The traffic in a large-scale backbone network has complex and diverse characteristics,and if all traffic is forwarded according to the same forwarding rules,traffic with delay-sensitive characteristics in the network will not reach its destination within the specified time.In addition,the overall traffic in the backbone network links is huge,and network congestion can easily occur when the traffic is unevenly distributed among the links,thus causing network link failures,which seriously affect the performance of the network and leads to the waste of network resources.Software Defined Network(SDN)plays an important role in the problem of fast traffic forwarding due to its centralized control and management features.To address the transfer delay problem in large-scale backbone networks and the load balancing problem among backbone network links,this thesis studies the load-balancing related technologies in software-defined networks.With the advantage of centralized control and management of software-defined networks,a load-balancing method based on queues and reinforcement learning is proposed through the cooperation between edge devices and backbone networks.In this thesis,we design and implement a load-balancing system based on a software-defined network.The research contents and contributions of this thesis are as follows:(1)Queue-based gating shaping algorithmTo ensure fast traffic forwarding and follow the principle that different types of traffic are forwarded according to different rules,this thesis designs a queue-based gate-shaping algorithm in edge devices and sets three queues with different forwarding rates.The algorithm distinguishes the traffic sensitivity to delay according to the size of the quality of service value in the packet header,and the traffic with relatively sensitive delay will be forwarded in the queue with high forwarding rate in priority;for latency-insensitive traffic,it will be forwarded in the low-speed queue.Thus,the algorithm ensures fast forwarding of delay-sensitive traffic and reduces the delay when traffic is forwarded at the edge devices.(2)Load balancing algorithm based on reinforcement learningTo ensure the transmission rate of traffic in the backbone network link,the utilization of network resources and network reliability,this thesis uses link load balancing in the backbone network link.Specifically,this thesis uses reinforcement learning to train the network model,select the current optimal path and then guide the forwarding of traffic in the form of a downlink flow table.In the reinforcement learning algorithm,Agent interacts with the environment and sets up appropriate action sets,state sets,and reward rules according to the experimental environment to generate experiences and store them in the experience set.A well-trained network model can perform the forwarding of traffic in a variety of situations.(3)Validity verification of algorithm and system designIn this thesis,the proposed algorithm is compared with ECMP algorithm and Hedera algorithm,and the effectiveness of the method in this thesis is verified by comparing three metrics: average link utilization,average throughput and average delay.In order to facilitate network managers to visually view the current network situation,this thesis designs and implements a load balancing system based on software-defined networks.The system can visually display the network state information and data plane topology,including the flow table query and distribution functions and the implementation of load balancing modules.Each module has passed the test.