Federated learning for edge devices delay control in software defined wide area networks
TL;DR
An SDN-orchestrated architecture for delay control in wide-area networks, combining edge-based GRU Active Queue Management (AQM) with layer-wise federated averaging, indicates that federated averaging reduces rare congestion events without meaningful degradation of normal bottleneck operation.
Abstract
This paper presents an SDN-orchestrated architecture for delay control in wide-area networks, combining edge-based GRU Active Queue Management (AQM) with layer-wise federated averaging. The proposed approach extends AQM from isolated per-router control to network-coordinated control with a shared global objective. Unlike conventional AQM approaches that regulate queues independently, the proposed design retains decentralized packet-level control while coordinating normalized queue targets and sharing locally learned control models between routers. A single interpretable parameter defines the target queue occupancy and the resulting delay–bandwidth trade-off. Fluid Flow simulations across multiple topologies, traffic loads, and three operating presets show that the SDN-controlled configurations reduce bottleneck queues and round-trip times relative to FIFO and NLRED under heavy traffic. A matched deterministic ablation isolates the effect of federated averaging. It reduced several extreme queue-tail measures, while typical bottleneck target tracking remained broadly comparable. These results indicate that federated averaging reduces rare congestion events without meaningful degradation of normal bottleneck operation.