Abstract:

The future of information technology (IT) is often represented as a vision of ubiquitous computing with large numbers of self-sufficient devices intelligently interacting via globally connected networks. The move towards such a vision is already noticeable in the trends of compaction and consolidation that have led to systems of shrinking dimensions with increasing functionality. Unfortunately, the large amounts of power required to enable multiple device features has simultaneously resulted in heat dissipation levels that are orders of magnitude higher than previous generations. Correspondingly, the active management of high-power chips, systems and data centers requires a complex cooling stack which is both capital-intensive and energy-intensive. The economic and environmental ramifications of such an infrastructure motivate the need to dynamically balance power consumption and cooling considerations from the chip to the data center. The enabling of a multi-scale resource allocation system to address this need will require a common thermodynamic platform upon which diverse infrastructure demands can be evaluated. This dissertation proposes such a foundation using the thermodynamic metric of exergy.

Rooted in the second law of thermodynamics, the proposed exergy-based approach attaches a "quality" to a given quantity of energy. Classical thermodynamics dictates that transport of energy across an inefficient system leads to a loss of quality (exergy) proportional to the inefficiency. Thus, using an exergy analysis, it becomes possible to comparatively assess the loss in performance resulting from various components of the cooling stack at the chip, system and data center levels. By normalizing this measure of thermodynamic performance per unit compute performance, a scalable figure-of-merit is derived to assess the efficiency of information processing at each level. Those components along the compute path which are least efficient are thus identified, and by preventing further allocation of resources to these suboptimal devices, the optimal end-to-end power delivery and cooling path is devised.

Thus, at a minimum, this dissertation presents the analytical framework required to design energy-efficient thermal management systems for chips, systems and data centers. At best, the proposed exergy-based approach enables an evaluation engine for the end-to-end management of resources from the chip to the data center and beyond.