On the Convergence of Local Descent Methods in Federated Learning