Energy modeling of supercomputers and large-scale scientific applications

We derive models of time, power, and energy as a function of frequency and demonstrate that these models accurately reflect measurements of large, parallel, scientific applications. We then apply these models to explain how sensitive the various applications are to CPU frequency; to analyze the compute-boundedness (frequency sensitivity) of complete, production workloads running on three of the world's fastest supercomputers; to predict how changes in system power properties will impact applications; and to determine what CPU frequencies offer the best energy, EDP, and ED
²
P savings.