Abstract : Video streaming has become one of the primary contributors to global Internet traffic, posing significant challenges to network infrastructures. Software-Defined Networking (SDN) offers a promising solution for managing such dynamic and bandwidth intensive services by enabling centralized control and realtime adaptability. However, decoupled decision making fails to account for the interplay between server workload and link congestion, often leading to suboptimal resource allocation. To address this issue, this paper presents a Graph Neural Network (GNN)-based Client-Server-Link (G-CSL) prediction model designed to optimize video streaming performance in SDN environments. G-CSL utilizes a machine learning model in conjunction with a GNN-based link estimation model. The machine learning predicts the video streaming server CPU utilization, which serves as input to the GNN model as node embeddings for link estimation between the client and server. For load forecasting, two machine learning and two deep learning models are evaluated, with Random Forest (RF) outperforming its counterpart. For the link estimation task, both traditional and GNN-based models are considered. GraphSAGE outperforms its counterparts by accurately estimating the existence of a link between the client and the video streaming server. A lightweight neighbor score heuristic then assigns each request to the least loaded server over the highest confidence path, maximizing a composite utility of computational headroom and bandwidth. An ablation study of the GraphSAGE model is presented highlighting the importance of architectural components, including batch normalization, bilinear decoders, temporal features, and threshold-based edge masking, in enhancing model robustness. The proposed model is evaluated under realistic video streaming scenarios involving 10,000 requests and compared with baselines. Experimental results show that G-CSL has achieved a 61\% reduction in request drop rate, maintains an average delay of 22 ms per request, and improves system utility by 23\%, demonstrating its effectiveness in balancing computational and bandwidth resources.
Index terms : Graph Neural Network, Load Prediction, Machine Learning, Software-Defned Network, Video Streaming