Constrained Power and Performance Optimization for Embedded Systems

University Of California-Irvine, Irvine CA

Investigators

Abstract

Constrained Power and Performance Optimization for Embedded Systems The goal of this research is to explore techniques for power management of embedded systems with provable bounds on power efficiency as well as adverse effects on latency due to power management. The current state of the art in system-level power management is limited to shutting parts of a system after a certain period of idle time, thus ignoring the application timing constraints, runtime traffic and application usage information. This work will develop power-performance control "knobs" that will allow us to make effective online decisions. Specific applications will include power-aware resource scheduling in RTOS, timing-aware power optimizations and tradeoffs between power savings and application quality of service (e.g., missed deadlines). Our technical focus is on solving two key problems: (a) latency-constrained power optimization, i.e., minimization of system-level power consumption with constrains on the effect of system latency due to power management; and (b) power-constrained performance optimization, e.g., system-level task implementation and scheduling within a given power budget. As a first step, we focus on analytic bounds on the effectiveness of online power management algorithms and their efficiency by developing bounds on latency increases due to power management. We introduce the notion of a competitive ratio as a quantitative measure of how well a given power management algorithm performs against an optimum power consumption profile. Next, we incorporate these analytic bounds in a broader system-level timing and power simulation engine which enables an accurate performance simulation while minimizing the details related to actual system functionality. Our experimental evaluation is through RTOS implementation of new power management services through coordinated scheduling and resource shutdown.