Client grouping and time-sharing scheduling for asynchronous federated learning in heterogeneous edge computing environment

Abstract

To overcome the three key challenges of federated learning in heterogeneous edge computing, i.e., edge heterogeneity, data Non-IID, and communication resource constraints, a grouping asynchronous federated learning (FedGA) mechanism was proposed.Edge nodes were divided into multiple groups, each of which performed global updated asynchronously with the global model, while edge nodes within a group communicate with the parameter server through time-sharing communication.Theoretical analysis established a quantitative relationship between the convergence bound of FedGA and the data distribution among the groups.A time-sharing scheduling magic mirror method (MMM) was proposed to optimize the completion time of a single round of model updating within a group.Based on both the theoretical analysis for FedGA and MMM, an effective grouping algorithm was designed for minimizing the overall training completion time.Experimental results demonstrate that the proposed FedGA and MMM can reduce model training time by 30.1%～87.4% compared to the existing state-of-the-art methods.

Keywords

• edge computing;federated learning;Non-IID;heterogeneity;convergence analysis