Deep Graph Neural Networks for Intelligent Traffic Routing in Software-Defined Networks

Modern networks are becoming more and more complex and scaled to make sure that that it can provide smart, real-time traffic management for best performance, reliability, and scale. In Software Defined Networking (SDN), we have a centralized control, so we have global network visibility and the programmable routing decisions. However, the routing algorithms for conventional networks often lack adaptability to change in network topology, deterioration of network conditions and the change in traffic demand. In this work, we introduce a new traffic routing frame work leveraging Deep Graph Neural Networks (GNNs) in order to enable intelligent decision making in SDNs. The GNN models the network as a graph where the dynamic attributes of nodes and links are bandwidth, latency, and traffic load respectively; and it learns spatial and contextual dependencies over the topology to infer a better routing path. In order to be scalable and adaptive, our approach is able to handle large scale real time routing decisions with high accuracy and low latency. We demonstrate extensive evaluations on simulated and real world network datasets where our method performs better than traditional shortest path and heuristic based algorithms in terms of throughput, delay reduction as well as with respect to adaptivity to topology changes. The contributions of this research are a viable way to embed intelligence into next generation SDN controllers through graph based deep learning models.