Straggler-Aware Resource Allocation in Semi-Decentralized Federated Learning for Large-Scale Models Over OTNs

We propose a straggler-aware resource allocation in semi-decentralized federated learning for large-scale machine-learning models over OTNs. Simulations show it achieves only 2.55% gap to MILP while reducing runtime by 99.78%, and outperforms the SoTA heuristic with 7% higher task success rate and 41% less runtime.