Dynamic Scheduling for Heterogeneous Federated Learning in Private 5G Edge Networks

Private 5G edge networks support secure and private service, spectrum flexibility, and edge intelligence. In this paper, we aim to design a dynamic scheduling policy to explore the spectrum flexibility for heterogeneous federated learning (FL) in private 5G edge networks. Particularly, FL is implemented with multiple communication rounds, in each of which the scheduled device receives the global model from the edge server, updates its local model, and sends the updated local model to the edge server for global aggregation. The heterogeneity in FL comes from unbalanced data sizes across devices and diverse device capabilities. In this regard, we start with the convergence analysis of FL to determine the role of unbalanced data sizes in the learning performance. Then, based on the fact that diverse device capabilities make the completion times of local updates asynchronous, we adopt the sequential transmission for global aggregation. On this basis, we formulate a heterogeneity-aware dynamic scheduling problem to minimize the global loss function, with the consideration of straggler and limited device energy issues. By solving the formulated problem, we propose a dynamic scheduling algorithm (DISCO), to make an intelligent decision on the set and order of scheduled devices in each communication round. Theoretical analysis reveals that under certain conditions, the learning performance and energy constraints can be guaranteed in the DISCO. Finally, we demonstrate the superiority of the DISCO through numerical and experimental results, respectively.