Power-aware systems

Abstract

In this thesis, we formalize the notion of power-aware systems and present a methodology to systematically enhance power-awareness. We define a power-aware system as one which scales its power consumption with changes in its operating scenario with a view to maximizing its energy efficiency. Operating scenarios are primarily characterized by five dimensions - input statistics, output quality requirements, tolerable latency (and/or throughput constraints), internal state and environmental conditions. We quantify the power-awareness of a system by equating it to the energy efficiency with which it can track changes along these dimensions. This is done by comparing the system's energy consumption in a scenario to that of a dedicated system constructed to execute only that scenario as energy efficiently as possible. We then propose a systematic technique that enhances the power-awareness of a system by composing ensembles of point systems. This technique is applied to multipliers, register-files, digital filters and variable-voltage processors demonstrating increases in battery-lifetimes of 60%-200%. In the second half of this thesis we apply power-awareness concepts to data-gathering wireless networks. We derive fundamental bounds on the lifetime of networks and demonstrate the tightness of these bounds using a combination of analytical arguments and simulation. Finally, we show that achieving a high degree of power-awareness in a wireless sensor network is equivalent to optimally or near-optimally solving the role-assignment problem. Provably optimal role assignment strategies using linear programming are presented. Hence, optimal strategies can be determined in a time that is polynomial in the number of nodes. As a result of applying power-awareness formalisms, the energy efficiency, and hence the lifetime of data gathering networks increases significantly over power-unaware schemes.