Throughput-Optimal Topology Design for Cross-Silo Federated Learning

Abstract

Federated learning usually employs a client-server architecture where an orchestrator iteratively aggregates model updates from remote clients and pushes them back a refined model. This approach may be inefficient in cross-silo settings, as close-by data silos with high-speed access links may exchange information faster than with the orchestrator, and the orchestrator may become a communication bottleneck. In this paper we define the problem of topology design for cross-silo federated learning using the theory of max-plus linear systems to compute the system throughput---number of communication rounds per time unit. We also propose practical algorithms that, under the knowledge of measurable network characteristics, find a topology with the largest throughput or with provable throughput guarantees. In realistic Internet networks with 10 Gbps access links for silos, our algorithms speed up training by a factor 9 and 1.5 in comparison to the master-slave architecture and to state-of-the-art MATCHA, respectively. Speedups are even larger with slower access links.

41 pages, NeurIPS 2020

References in corpus (6)

• Federated Optimization: Distributed Machine Learning for On-Device Intelligence
• Decentralized Stochastic Optimization and Gossip Algorithms with Compressed Communication
• Expanding the Reach of Federated Learning by Reducing Client Resource Requirements
• Peer-to-peer Federated Learning on Graphs
• SlowMo: Improving Communication-Efficient Distributed SGD with Slow Momentum
• Decentralized gradient methods: does topology matter?