Coordinating Multiple Autonomic Managers to Achieve Specified Power-Performance Tradeoffs

• Authors: Jeffrey O. Kephart;Hoi Chan;Rajarshi Das;David W. Levine;Gerald Tesauro;Freeman Rawson;Charles Lefurgy
• Venue: ICAC '07 Proceedings of the Fourth International Conference on Autonomic Computing
• Year: 2007

Getting multiple autonomic managers to work together towards a common goal is a significant architectural and algorithmic challenge, as noted in the ICAC 2006 panel discussion regarding "Can we build effective multi-vendor autonomic systems?" We address this challenge in a real small-scale system that processes web transactions. An administrator uses a utility function to define a set of power and performance objectives. Rather than creating a central controller to manage performance and power simultaneously, we use two existing IBM products, one that manages performance and one that manages power by controlling clock frequency. We demonstrate that, with good architectural and algorithmic choices established through trial and error, the two managers can indeed work together to act in accordance with a flexible set of power-performance objectives and tradeoffs, resulting in power savings of approximately 10%. Key elements of our approach include a) a feedback controller that establishes a power cap (a limit on consumed power) by manipulating clock frequency and b) reinforcement learning, which adaptively learns models of the dependence of performance and power consumption on workload intensity and the powercap.