In this paper, a graph neural net-based routing algorithm is designed which leverages global information from controller of a software-defined network to predict optimal path with minimum average delay between source and destination nodes in software-defined networks. Graph nets are used because of their generalization capability which allows the routing algorithm to scale across varying topologies, traffic schemes and changing conditions. A deep reinforcement learning framework is developed to train the Graph Neural Networks using prioritized experience replay from the experiences learnt by the controllers. The algorithm is tested on various small and large topologies in terms of packets successfully routed and average packet delay time. Experiments are performed to check robustness of routing algorithms to changes in network structure and effects of varying hyperparameters. The proposed algorithm shows impressive results when compared to q-routing and shortest path routing algorithm in terms of above experiments and is robust to varying graphical structure of the network.

在本文中，设计了一种基于图神经网络的路由算法，该算法利用来自软件定义网络控制器的全局信息来预测软件定义网络中源节点和目标节点之间具有最小平均延迟的最佳路径。使用图网络是因为它们的泛化能力允许路由算法在不同的拓扑、流量方案和不断变化的条件下进行扩展。开发了一个深度强化学习框架，以使用从控制器学习到的经验中的优先经验回放来训练图神经网络。就数据包成功路由和平均数据包延迟时间而言，该算法在各种小型和大型拓扑结构上进行了测试。执行实验以检查路由算法对网络结构变化和不同超参数影响的鲁棒性。与上述实验中的 q-routing 和最短路径路由算法相比，所提出的算法显示出令人印象深刻的结果，并且对网络的不同图形结构具有鲁棒性。