A Rate Matchingbased Approach to Dynamic Voltage Scaling.pdfVIP

A Rate Matching-based Approach to Dynamic Voltage Scaling David Biermann, Emin G¨un Sirer, Rajit Manohar Computer Systems Laboratory Cornell University Ithaca, NY 14853, U.S.A. Abstract converge to a single optimal operating point. This has We present a simple rate matching-based mechanism for led to a recent experimental study that thoroughly evalu- voltage adaptation in a microprocessor running a mul- ated many previous voltage scaling schemes to conclude tiprogrammed workload. The mechanism incorporates that “No heuristic policy that we examined achieved a set of architecture and operating system extensions [the optimal voltage and frequency].”1 [8] Part of the through which applications can communicate their actual reason why these heuristic approaches are limited is be- and desired progress to the operating system. Using this cause they are driven solely by system idle time and have feedback, the operating system uses a modified schedul- no application-specific information. ing algorithm to run all applications at a single, globally- Other work has examined how to select the optimal optimal voltage. We demonstrate that significant energy voltage given complete information about application savings are possible with a simple, practical set of exten- start times, deadlines, and computation needs [12, 20]. sions to the architecture and operating system. Given complete application knowledge, these omniscient schemes can optimally pick the operating voltage to min-