In mobile edge computing (MEC), edge servers and mobile terminals use federated learning distributed architecture to build a deep model, so that terminals can cooperate in training without sharing data. Distributed training requires network virtualization to provide high bandwidth and low latency characteristics to support large-scale parallel computing. Traditional virtual network embedding (VNE) relies on a static network topology, which lacks flexibility and incurs high resource costs during model training. To improve the efficiency of embedding distributed training tasks, we propose a novel Node Selection and Dynamic Topology resource allocation scheme for VNE of distributed training, NSDT-VNE, based on reconfigurable network topology. This algorithm divides the underlying network into static and dynamic topologies, enhancing low latency for small flows while providing high bandwidth for large flows as needed. Additionally, we introduce a two-phase coordinated alternating optimization algorithm that optimizes embedding decisions at both computational and topological levels, ensuring optimal node selection. Overall, NSDT-VNE follows demand-aware network design principles, allowing continuous optimization of the underlying topology. Compared to state-of-the-art heuristic and reinforcement learning-based virtual network algorithms, NSDT-VNE achieves superior performance, with request acceptance rates improving by 6.67% to 25.68% and embedding revenue increasing by approximately 7% to 32%.