Edge-to-cloud Latency Aware User Association in Wireless Hierarchical Federated Learning

In this paper, we propose a backbone-aware user association algorithm for heterogeneous hierarchical federated learning. We consider the scenario in which mobile devices have different computation and communication capabilities, while edge servers have different model uploading delays to the cloud server. To find an optimal user association, we formulate a combinatorial optimization problem that takes into consideration mobile-to-edge delays and edge-to-cloud delays. To reduce the computational complexity, we put forward the backbone-aware greedy algorithm. In addition, we prove that it is not always optimal for a mobile device to connect to the edge server with the minimum mobile-to-edge delay. Furthermore, we propose using dynamic bandwidth allocation after assigning users to edge servers to further reduce the latency. We also use simulation results to show the advantages of the proposed approach.