ALTM: Adaptive learning-based thermal model for temperature predictions in data centers

Abstract

To design effective control schemes for energy efficiency in data centers, it is crucial to have a thermal model of the system. Constructing thermal models of data centers for temperature prediction is extremely challenging, due to inherent complexity. Computational fluid dynamics (CFD) simulations or physical heat transfer equations are conventionally used to construct such thermal models. More recent approaches combine physical heat transfer rules and data-driven methods in an effort to obtain more accurate models. Our proposed adaptive learning-based thermal model (ALTM) is fast, adapts to thermal changes in the data center environment, and does not require prior knowledge of heat transfer rules between data center entities. Unlike other methods, ALTM is a holistic thermal model that predicts temperature of critical zones using data center operational variables as inputs. The operational variables are the controllable parameters and easily obtained measurements from IT and cooling units. A key use case for ALTM is that it can be effectively used for thermal-aware workload schedulers or cooling system controllers. Our results confirm the accuracy and adaptability of the model.