SHF: Small: Power Husbanding via Architectural Techniques (PHAT)

University Of Wisconsin-Madison, Madison WI

Investigators

Abstract

Power consumption has become a dominant design constraint for computer systems. The ITRS Roadmap states: "power management is now the primary issue across most application segments." Worst-case power provisioning has given way to power capping, which throttles performance when necessary. Unfortunately, power-capping is analogous to brown-outs, which penalize all customers of an electrical utility, rather than those who can best tolerate loss of power. The thesis of this research is that computer systems should treat power as a precious resource to be husbanded: To direct and manage with frugality; to use or employ to good purpose and the best advantage. The research will investigate effective power husbanding via architectural techniques that appropriately shift power between different hardware components, much as good budgeting applies money where it is most effective. Power husbanding solutions include (a) low-overhead online mechanisms to estimate resource use, (b) low-cost techniques for dynamically changing a resource's operational level, and (c) effective policies for reaching power efficient operating configurations. Research will extend previous accomplishments on reconfigurable caches to enable online control at fine granularities, extend previous and develop new core architectures to enable proportional scaling of performance and power, and adapt and develop new memory systems which often accounting for 30% of system power. Power husbanding seeks to make computation more power efficient, which has important national and global implications. The PIs will continue to advance the state-of-the-art through students, courses, talks, industrial affiliates, commercial influence, and open-source simulation infrastructure.