From Reactive to Proactive: GRU-Based Dynamic Link Cost Prediction for QoS-Aware SDN Routing

Abstract

Software Defined Networking (SDN) enables centralized and flexible routing control by allowing the controller to dynamically compute link costs based on Quality of Service (QoS) metrics. However, many existing QoS-based routing approaches remain reactive, where routing adjustments are only performed after congestion thresholds have been exceeded, resulting in packet loss, increased latency, and routing instability. This paper proposes a proactive predictive dynamic weightage assignment approach that integrate a Gated Recurrent Unit (GRU)-based Prediction Engine into the SDN controller. The proposed model forecasts future network conditions using historical telemetry data collected from the Network Monitoring module and dynamically adjusts QoS metric weight assignments before congestion occurs, enabling high-priority traffic to be proactively redirected to more stable paths. The model is evaluated under multiple traffic conditions and compared against the reactive dynamic approach. Experimental results show significant improvements, where average packet loss is reduced from 2.4% to 0.8%, end-to-end latency decreases from 22 ms to 17 ms, and routing flaps are reduced from 14 to 5 per minute. Although convergence time increases slightly from 45 ms to 58 ms due to prediction overhead, the GRU model achieves strong forecasting accuracy with RMSE values below 0.06 across different traffic profiles. These results demonstrate that proactive predictive routing provides better network stability, lower congestion impact, and more reliable QoS performance for high-utilization SDN environments