Federated learning has emerged as a prominent technology in edge computing, as it allows for large-scale distributed learning while ensuring data privacy. However, high-latency edge nodes can create straggling links, significantly degrading communication efficiency. To tackle this problem, coding techniques have been widely explored. In an extended two-tier federated learning system, edge nodes transmit locally coded gradients to helper nodes, which handle aggregation and data routing. The choice of aggregation method critically affects downlink communication efficiency, while helper nodes operate under limited resources. To address these challenges, we propose an efficient equivalent aggregation method that guarantees bounded downlink communication cost as the number of edge nodes scales. Furthermore, we develop a layered coding aggregation scheme based on maximum distance separable codes, enabling a tunable tradeoff in communication costs between edge and helper nodes through a design parameter \(\alpha \) .