Network Representation Learning And Routing Optimization Based On Software Defined Network

The rapid development of network technologies and the proliferation of network devices facilitate the development of network architecture in diversification,heterogeneity and scale.As one of the basic functionalities of the network,routing indicates the calculation of transmission paths of data packets.Network modeling is the basic of routing and is important for the improvement of network processing ability and efficiency.Traditional matrix-based modeling gradually lacks efficiency and accuracy in depicting a network,and thus limits the improvement of network performances.In this paper,we focus on the network representation learning(NRL)of a software defined network(SDN)and try to combine NRL with network routing optimization.We introduce knowledge graph architecture and embed network nodes,links and packets into low-dimensional vectors to improve the efficiency and accuracy of network modeling.Simulation experiments are carried out to compare our algorithms with shortest-path benchmark schemes.For static network topologies,represented by terrestrial networks,we design a knowledge graph embedding model to embed network nodes,links and user packets into low-dimensional vectors,which consist of topological characteristics.Then we propose a routing scheme based on vector similarities.Simulation results indicate that our algorithm has a similar average delay,a lower packet loss ratio and a higher average throughput comparing to shortest-path routing algorithms,with a good scalability.For dynamic network topologies,represented by satellite networks,we propose a modeling algorithm combining network snapshots and knowledge graph embedding model.In this model,dynamic topologies are transformed to a knowledge graph sequence consisting of low-dimensional vectors of network nodes and links.Similar routing schemes based on vector similarities are then proposed.Simulation results show that our algorithm has a lower packet loss ratio and a higher average throughput comparing to benchmark schemes.