Joint Computational Resource Allocation and Layer Partitioning for Federated Learning

Despite its popularity, federated learning (FL) in heterogeneous networks faces two critical challenges, i.e., the straggler problem due to devices with limited capabilities and low resource utilization rate of the FL server. The straggler problem arises when devices with limited computational capabilities delay the convergence of the global model. On the other hand, the computational resources of the FL server are often underutilized, mainly due to its relatively simple involvement for model aggregation. To tackle the issues in diverse scenarios, we propose a new joint computational resource allocation and layer partitioning (JCRALP) scheme to improve the overall FL performance by leveraging the capabilities and resources of both FL server and all clients. In the scenario where system parameters regarding the computational capabilities of the clients and the task burden can be accurately measured, we propose an optimization-based approach that leverages our proposed multistep water-level equalization algorithm and the incremental ceiling adjustment algorithm. In the scenario where parameters cannot be measured accurately, we propose a reinforcement learning-based method using a modified twin delayed deep deterministic policy gradient algorithm. Extensive simulation results demonstrate that JCRALP efficiently and effectively mitigates the straggler problem and inclusively enables more client participation in FL. By including more datasets, the global model becomes more representative, while server computational resources are utilized more efficiently, significantly reducing convergence latency.