Edge Resource Autoscaling for Hierarchical Federated Learning Over Public Edge Platforms

Abstract

Federated learning promises to empower ubiquitous end devices to collaboratively learn a shared model in a privacy-preserving manner. To reduce the enormous and expensive wide-area-network (WAN) traffic incurred by the traditional two-tiered cloud-device federated learning, hierarchical federated learning over cloud-edge-device has been proposed recently. With hierarchical federated learning, edge servers are leveraged as intermediaries to perform local model aggregations to reduce the model updates aggregated by the centralized cloud. Considering the emerging public edge platforms such as Aliyun Edge Node Service that rent edge servers to users in an on-demand manner, we present AutoEdge, an edge server autoscaling framework for hierarchical federated learning. The goal of AutoEdge is to autoscale edge servers against dynamical device participants in a cost-efficient manner. Achieving this goal is challenging since the underlying long-term optimization problem is NP-hard involves the future system information. To attack these challenges, AutoEdge first applies regularization technique to decompose the long-term problem into a set of solvable fractional subproblems. Then, adopting a randomized dependent rounding scheme, AutoEdge further rounds the fractional solutions to a near-optimal and feasible integral solution. AutoEdge achieves outstanding performance guarantee, as verified by both rigorous theoretical analysis and extensive trace-driven simulations.