Bandwidth-Aware Decentralized Federated Learning in Wired Networks

Abstract

Decentralized federated learning (DFL) has emerged as a promising alternative to centralized federated learning by eliminating communication bottlenecks at a single server. However, prior studies on DFL typically assume wireless-network-like settings and measure communication cost simply as the number of transfer destinations per worker, without considering the actual transfer time governed by network bandwidth. In practice, cross-silo FL applications such as inter-datacenter and inter-enterprise settings require wired network infrastructure, where concurrent model transfers traverse shared physical edges and cause bandwidth congestion. To address this challenge, we formulate a unified optimization problem for DFL over wired networks that jointly considers data heterogeneity, application-layer topology, and physical-layer bandwidth bottlenecks. We introduce max-min fair bandwidth allocation for model transfers and propose a subtractive greedy algorithm that starts from a full-mesh topology and iteratively removes transfers that contribute least to convergence relative to the bandwidth bottleneck they cause. Experiments on two wired network topologies with 20 workers confirm that our method achieves higher test accuracy than baseline methods under the same training time budget.