Felinet: Accelerating Federated Learning Convergence in Heterogeneous Edge Networks

The edge network has been introduced for providing computing capabilities to accelerate federated learning. However, the heterogeneity of edge networks increases the complexity of traffic scheduling, which can result in network congestion and decreased FL training efficiency. In this paper, we propose Felinet, a path-driven routing solution designed to effectively handle heterogeneous connections and unpredictable FL workloads. We evaluate Felinet through real-world network experiments, demonstrating a 26% improvement in network resource allocation compared to the benchmark routing solution.