Full-System Power Analysis and Modeling for Server Environments

(c) 2006 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other users, including reprinting/ republishing this material for advertising or promotional purposes, creating new collective works for resale or redistribution to servers or lists, or reuse of any copyrighted components of this work in other works.The increasing costs of power delivery and cooling,\ud as well as the trend toward higher-density computer systems,\ud have created a growing demand for better power management in\ud server environments. Despite the increasing interest in this issue,\ud little work has been done in quantitatively understanding power\ud consumption trends and developing simple yet accurate models to\ud predict full-system power. We study the component-level power\ud breakdown and variation, as well as temporal workload-specific\ud power consumption of an instrumented power-optimized blade\ud server. Using this analysis, we examine the validity of prior adhoc\ud approaches to understanding power breakdown and quantify\ud several interesting trends important for power modeling and\ud management in the future. We also introduce Mantis, a nonintrusive\ud method for modeling full-system power consumption\ud and providing real-time power prediction. Mantis uses a onetime\ud calibration phase to generate a model by correlating AC\ud power measurements with user-level system utilization metrics.\ud We experimentally validate the model on two server systems\ud with drastically different power footprints and characteristics\ud (a low-end blade and high-end compute-optimized server) using\ud a variety of workloads. Mantis provides power estimates with\ud high accuracy for both overall and temporal power consumption,\ud making it a valuable tool for power-aware scheduling and\ud analysis