Haste Makes Waste: A Simple Approach for Scaling Graph Neural Networks