ELSA-FL: Energy and Load Serving Aware Federated Learning

Abstract

Federated Learning (FL) enables collaborative model training across distributed devices. A major concern in FL is how to operate it smoothly in resource-constrained environments, where this process must perform under strict operational constraints, such as cost or energy budgets. An often-overlooked aspect is the effect of introducing a new FL process into a running environment, where the additional workload competes for limited, shared resources, thereby disrupting already-running workloads that rely on stable, low-latency performance. This challenge is most pronounced in continuous FL, where inference requests and training overlap, potentially leading to offloading of model serving. Offloading can add costs, raise privacy concerns, and increase latency. Consequently, mitigating the number of inference requests offloaded is beneficial and should be considered when selecting devices for training. We propose ELSA-FL, a novel method that autonomously learns an effective client-selection policy under a strict budget constraint. Our approach aims not only to improve FL performance within the given energy budget but also to reduce the latency of parallel inference requests by preferentially selecting clients with lower background load. The objective is to enable a more efficient and practical deployment of FL in real-world, multi-task environments. Our results demonstrate that the proposed method consistently offloads significantly fewer requests than other RL and heuristic methods, while minimally affecting FL performance. Therefore, our method ensures good performance of the FL process, while interrupting ongoing processes as little as possible.