The impact of GNN architecture on the robustness of edge routes in scenarios of single node types

Abstract

The paper considers the problem of intelligent routing (path finding) in software-defined networks (SDN) using graph neural networks in order to increase the efficiency of network resource use and adapt to dynamic changes in the network state (for example, channel congestion or delays).

Topicality. Modern software-defined networks (SDN) are faced with increasing traffic volumes and increased quality of service (QoS) requirements. Traditional routing algorithms (e.g. Dijkstra) are static and inefficient in conditions of high load dynamics or sudden topology changes (hardware failures). This leads to channel congestion, increased latency, and packet loss. The use of machine learning methods, in particular graph neural networks (GNN) and reinforcement learning (RL), opens up new opportunities for creating adaptive intelligent agents capable of optimizing routing in real time, which makes this research timely and important for the development of telecommunication systems.

Goal. The aim of the research is to improve the efficiency and reliability of data transmission in SDN networks by developing and comparative analysis of intelligent routing methods. The main focus is on the study of graph neural network architectures (GCN, GAT) and the Q-Learning algorithm to provide adaptive traffic management under conditions of variable load and network node failures.

Research methods. The methodological basis of the work is based on the integrated application of graph theory to formalize the network topology, deep learning methods to process node features, and reinforcement learning algorithms to make routing decisions. Experimental verification of the proposed approaches was carried out by emulating a software-configured network in the Mininet environment under the control of the Ryu controller. The software implementation included the development of models based on convolutional networks (GCN) and attention networks (GAT) using deep learning libraries, as well as the implementation of the Deep Q-Learning agent. The effectiveness of the algorithms was assessed by comparative analysis of key quality of service metrics - throughput, average latency, and packet loss percentage - in scenarios of gradual load growth and emergency topology change due to equipment failure.

The results. The study found that the integration of machine learning methods allows for significant improvements in data transmission parameters compared to the classic Dijkstra algorithm, especially in high traffic conditions, where intelligent agents provide lower latency and connection stability. Critical analysis of fault tolerance revealed significant differences between the studied architectures: the GCN model demonstrated limited adaptability with a routing failure rate of 30%, while the GAT architecture showed better flexibility, generating optimal paths in half of the cases. The highest efficiency was confirmed by the Q-Learning method, which, thanks to dynamic interaction with the environment, ensured the construction of ideal routes in 85% of experiments and minimized packet loss to 5–7% even in critical situations, which proves the superiority of Reinforcement Learning approaches over Supervised Learning methods in adaptive network control tasks.