Node Selection Toward Faster Convergence for Federated Learning on Non-IID Data