Abstract: Federated Learning (FL) enables devices to train model locally using their data, thereby protecting data privacy. However, in synchronous FL implementations, using all local data for model training may cause a relatively long computation latency, and certain data have a rather limited impact on enhancing model accuracy. Moreover, excessive participation of devices and unreasonable bandwidth allocation will further increase communication latency. Rational strategies for bandwidth allocation, device selection, and dataset streamlining can accelerate FL training while ensuring the model accuracy. Therefore, an issue of data selection, bandwidth allocation, and device selection is put forth based on data importance, aiming to select the data that can most improve model accuracy while minimizing the latency of FL under devices’ energy constraint. To solve this problem, an algorithm is designed based on Lagrange duality decomposition and subgradient method. Specifically, since the dual problem of the original problem is a convex problem, the solution of the dual problem can be obtained by iteratively multiplying the solution using the subgradient method, and then the solution of the dual problem can be used as the suboptimal solution of the original problem. The solution of dual function values is implemented by dividing it into two subproblems, namely bandwidth allocation subproblem and device and data selection subproblem, and then iteratively solving them alternately. The bandwidth allocation subproblem is solved using the KKT conditions, while the device and data selection subproblem is addressed via exhaustive search. Experimental results demonstrate that our algorithm can reduce the total latency by at least 8.32 percentage points and at most 51.07 percentage points for different device total numbers, device power thresholds, and other situations, on average, compared with baseline algorithms. In terms of model accuracy, this algorithm can ensure no loss in model accuracy and even outperform some baseline algorithms by approximately 1~3 percentage points, on average.