Holistic thermal-aware workload management and infrastructure control for heterogeneous data centers using machine learning

摘要

Two key contributors to the energy expenditure in data centers are information technology (IT) equipment and cooling infrastructures. The standard practice of data centers lacks a tight correlation between these two entities, resulting in considerable power wastage. Considering the cooling cost of different locations inside a data center (cooling heterogeneity) and various cooling capabilities of servers (server heterogeneity) has significant potential for saving power, yet has not been studied thoroughly in the literature. There is a necessity for state-of-the-art approaches to integrate the control of IT and cooling units. Moreover, the literature still lacks an accurate and fast thermal model for temperature prediction inside a data center. In this paper, innovative approaches to quantify data center thermal heterogeneities are presented. Using data center thermal models the cost of providing cold air at the front of servers can be (indirectly) calculated, and the capability of servers to be cooled is formulated. Our approach assigns jobs to locations that are efficient to cool (from the perspectives of both servers and cooling units) and tunes cooling unit parameters. The method, called holistic data center infrastructure control (HD1C), has the potential to save a considerable amount of power by exploiting synergies between the workload scheduler and operational parameters of cooling units.