Research And Implementation Of Load Balancing Based On GCN And DRL In SDN Data Center Network

Since the 20 th century,the rapid development of the Internet has led to the explosive growth of various network services,which produces lots of cloud services and users and traffic,finally result in larger Data Center Networks.In order to better manage the network,SDN is proposed and applied to the Data Center Networks.With the increase of traffic scale,the load balancing strategy becomes a key measure to improve network stability,which can avoid congestion by balancing traffic between different data links.Load balancing algorithms select proper routes by analyzing the network status,so that the traffic can be distributed on the network evenly.However,the classical load balancing algorithms such as Equal-Cost Multipath can not adapt to the increasing network and traffic scale.In recent years,the application of AI to load balancing algorithm has become an effective means.Due to the powerful information fitting and computing ability of neural network,the AI-based load balancing algorithms has significant advantages over the traditional algorithms,which can deal with more complex network states and make better routing decisions.Aiming at the load balancing problem in SDN Data Center Networks,thesis designs and implements a three-layer load balancing system based on SDN by summarizing and analyzing the existing work,which solves the load balancing problem in the SDN Data Center Networks by making routing decisions through the deep reinforcement learning model.The performance of our model were verified by comparison with algorithms such as ECMP and GDLB.Specifically,thesis combines deep reinforcement learning with GCN and RNN and applies them to the load balancing algorithm for the first time,innovatively designs a neural network named BRGCN,which can not only make decisions according to the state information of the physical network,but also capture the information of topological structure of the physical network and the timing information of traffic.Thesis also designs a deep reinforcement learning algorithm named DDDRQN and a new random continuous sampling strategy.DDDRQN uses multiple factors to evaluate rewards and calculates the variance of link utilization as the primary reward factor.By using random continuous sampling strategy,DDDRQN solves the problem of discontinuity of training data caused by random sampling,so it can train the neural network with RNN well.