| With the rapid development of digitalization,cloud computing,Internet of Things,artificial intelligence and other technologies continue to deepen,more and more computing resources are needed to process and store data,and the network traffic scale of data center,as a facility for data processing and providing computing resources,is also increasing.Traditional rule-based load balancing algorithms,which route traffic by polling and hashing,cannot dynamically sense the data center network status and cannot meet the increasingly large data center scale and traffic scale.Software-defined networking（SDN）can flexibly allocate network resources and adjust traffic routing in real time by separating the control plane and data plane to improve the performance,flexibility and security of data centers.Therefore,this thesis investigates the following two directions based on SDN architecture,combined with machine learning related algorithms:（1）A deep reinforcement learning-based load balancing algorithm HGCN algorithm is proposed for the routing decision of traffic in data centers.In this paper,the routing decision process is modeled as a Markov decision process,and the HGCN（Hybrid GraphConvolutional Network）algorithm is based on the DQN algorithm,and the combination of graph convolutional neural network（GCN）and recurrent neural network（GRU）is used on the neural network to extract data center network state information and timing information,respectively,and the link utilization,delay,and packet loss rate are taken as the rewards that affect the actions of the intelligence in the reinforcement learning algorithm design In the design of the reinforcement learning algorithm,the link utilization,delay,and packet loss rate are used as reward factors affecting the actions of the intelligences and are assigned different weights to make the HGCN model have balanced performance in load balancing and QoS metrics.The converged HGCN model has a 10%,13%,and 14%improvement in the average link utilization index representing load balancing compared with SPF,ECMP,and DQL algorithms in the conventional environment,and a 12%,20%,and 7%improvement in the high-pressure network environment.In terms of QoS-related delay and packet loss rate indicators,compared with ECMP,SPF and DQL,the packet loss rate of HGCN can be maintained at the default value of 10%,and the delay has improved by 11%,15%and 9%.（2）To evaluate the node load balancing status in data centers,a classification algorithm based on node importance,DGraphSage（Degree-based GraphSage）is proposed,which innovatively proposes an aggregation function based on node importance based on GraphSage,a graph neural network.When evaluating the load balance state of nodes,the state of neighboring nodes is aggregated,and the weight of nodes in the aggregation process is assigned based on the degree of nodes,which allows key nodes in the data center to have a higher weight in the aggregation process.In the training process of the algorithm,the homemade DataCenter dataset of the data center is used,and the port state and flow table state of the aggregation switch nodes are used as the feature representation of the nodes,so that DGraphSage can perform better in the classification of the data center topology graph.In the comparison experiments,the average improvement values ofmacro-F1 of DGraphSage algorithm on DataCenter,Citesseer and Cora datasets are 21.96%,8.91%and 15.66%.The average improvement values of micro-F1 are 19.09%,12.90%and 16.65%,respectively. |
