This paper addresses the challenge of efficient resource utilization in federated learning, where data is distributed across multiple edge nodes and trained using gradient-descent-based approaches. The authors analyze the convergence bound of distributed gradient descent theoretically and propose a control algorithm to dynamically adjust the frequency of global aggregation to minimize the loss function under a given resource budget. The algorithm estimates parameters such as resource consumption and loss function characteristics in real-time and adapts the global aggregation frequency accordingly. Extensive experiments on real datasets and in simulated environments demonstrate that the proposed approach achieves near-optimal performance across different data distributions and system configurations.This paper addresses the challenge of efficient resource utilization in federated learning, where data is distributed across multiple edge nodes and trained using gradient-descent-based approaches. The authors analyze the convergence bound of distributed gradient descent theoretically and propose a control algorithm to dynamically adjust the frequency of global aggregation to minimize the loss function under a given resource budget. The algorithm estimates parameters such as resource consumption and loss function characteristics in real-time and adapts the global aggregation frequency accordingly. Extensive experiments on real datasets and in simulated environments demonstrate that the proposed approach achieves near-optimal performance across different data distributions and system configurations.