A Fine-grained Approach for Power Consumption Analysis and Prediction

Power consumption has became a critical concern in modern computing systems for various reasons including financial savings and environmental protection. With battery powered devices, we need to care about the available amount of energy since it is limited. For the case of supercomputers, as they imply a large aggregation of heavy CPU activities, we are exposed to a risk of overheating. As the design of current and future hardware is becoming more and more complex, energy prediction or estimation is as elusive as that of time performance. However, having a good prediction of power consumption is still an important request to the computer science community. Indeed, power consumption might become a common performance and cost metric in the near future. A good methodology for energy prediction could have a great impact on power-aware programming, compilation, or runtime monitoring. In this paper, we try to understand from measurements where and how power is consumed at the level of a computing node. We focus on a set of basic programming instructions, more precisely those related to CPU and memory. We propose an analytical prediction model based on the hypothesis that each basic instruction has an average energy cost that can be estimated on a given architecture through a series of micro-benchmarks. The considered energy cost per operation includes all of the overhead due to context of the loop where it is executed. Using these precalculated values, we derive an linear extrapolation model to predict the energy of a given algorithm expressed by means of atomic instructions. We then use three selected applications to check the accuracy of our prediction method by comparing our estimations with the corresponding measurements obtained using a multimeter. We show a 9.48\% energy prediction on sorting