On the Practicality of Differential Privacy in Federated Learning by Tuning Iteration Times