Joint dynamic placement and resource allocation for the Federated Learning aggregation function at the edge

Federated Learning (FL) is gaining momentum as a prominent solution to enable on-device training while preserving data privacy. In a typical FL architecture, the function of aggregating model updates from thousands of client devices, each locally training the shared model on its own data, is placed at a dedicated node within the edge domain. Such a placement decision remains static throughout the entire training process. However, the time-varying, heterogeneous, and limited availability of computing capabilities in edge environments, coupled with fluctuating populations of participating clients across training rounds, calls for dynamic per-round placement of the aggregation function. To further optimize the usage of heterogeneous resources while accounting for FL aggregation tasks, which may require different scales of computing capabilities, a fine-grained allocation of computing resources is needed across edge nodes. The objective of this work is to jointly optimize the FL aggregation function placement at each training round and the allocated computing resources in order to minimize the overall per-round training time while not exceeding the computing capabilities of each edge node. An optimization problem is formulated, and an efficient and effective heuristic algorithm is proposed, based on local search techniques. Experimental results demonstrate that the proposed solution significantly outperforms benchmark approaches, achieving a reduction in per-round training time and bandwidth consumption. Under the considered settings, the proposal yields up to a 27% gain in training time compared to methods that optimize aggregation placement only, without performing fine-grained resource allocation.