Machine learning assisted development of IT equipment compact models for data centers energy planning

Abstract

In most data centers, performance reliability is often ensured by setting the amount of airflow provided by the cooling units to substantially exceed that which is needed by the IT equipment. This overly conservative strategy requires additional energy expenditure, which inevitably results in a huge amount of energy being wasted by the cooling system. To eliminate adopting such wasteful policies, conducting proper management of airflow, temperature, and energy is critical. To that end, this work proposes a novel approach to developing a compact IT equipment model at off-design conditions. This model is designed to support thermal and energy management functions in data centers. The benefit of this model is that it can accurately predict not only the IT equipment power consumption, but also the amount of flowrate required for the equipment and the air temperature leaving the equipment. While the compact model’s power consumption was derived as a function of CPU utilization, its flowrate demand and exhaust temperature were obtained from a dynamic detailed CFD model. Results from the compact model were validated with experiments where the maximum mismatch was found to be 5.7% in the outlet temperature field and 11.4% in flowrate. Compared to a state-of-the-art IT equipment compact model, the developed model was found to reduce the prediction error of the equipment’s flowrate and outlet air temperature by up to 5.2% and 9.3 % that of the state-of-the-art IT equipment compact model, respectively.