Evolutionary power modeling for energy efficiency in CPU-GPU based systems

Supercomputers have reached a massive energy consumption due to computational demand, so there is an urgent necessity to keep them on a more scalable curve. In the last years, there has been a rising interest in reducing the power consumption of these systems. Recently research works focus on the adjustment of their power states by reducing clock frequency, applying power capping, and on the analysis of the thermal impact on static consumption. These techniques rely on power models to predict the power consumption of the infrastructure. However, the power consumption in these complex systems involves a vast number of interacting variables of different nature that may include non-linear dependencies. So, extracting the relationships between the most representative parameters and the power consumption requires an enormous effort and knowledge about the problem. We propose an automatic method based on Grammatical Evolution to obtain a model that minimizes the power prediction error of a supercomputer node that incorporates both CPU and GPU devices. We monitor the system during runtime using performance counters and frequency, temperature and power measurements. This evolutionary technique provides both Feature Engineering and Symbolic Regression to infer accurate models, which only depend on the most suitable variables, with little designers expertise requirements and effort. Our work improves the possibilities of deriving proactive energy-efficient policies in supercomputers that are simultaneously aware of complex considerations of different nature