Neural network methods for fast and portable prediction of CPU power consumption

The need for energy efficient computing has established performance-per-watt as a first-class metric for evaluating HPC applications. Consequently, optimizations that target HPC systems and data centers are required to dynamically monitor system power consumption in order to be effective. Although newer architectures are making power sensors available on the chip, the general state of power measurement tools across different architectures remains deficient. This paper describes a neural-network based model for fine-grain, accurate and low-cost power estimation. The main novelty of the proposed approach is its portability. The methodology can be adopted to predict power consumption not only on a range of current processors but future architectures as well. This portability is achieved by taking advantage of performance monitoring units (PMU) available on current systems and applying a carefully constructed sequence of feature selection techniques. We evaluate our models along several dimensions on multiple platforms. The experimental results show that the constructed models are able to predict power consumption with high accuracy at a low overhead. The results also provide key insight as to the number of features necessary to achieve reasonable prediction accuracy.