Increased awareness of power concerns and recent approaches to improve energy efficiency and heat dissipation in computer systems have not immediately translated into rich, adaptable, power-aware applications. This lack of effectiveness is particularly evident in power-aware real-time computing, where efforts to bridge the gap between hardware and software have generally been limited to developing new schedulers that operate by modulating a relatively high-level hardware control such as voltage scaling and frequency scaling. The problem with this approach is that the hardware controls operate at too gross a level to be effective in dynamic, poorly-modeled application scenarios.
This project develops a feedback-based, control-theoretic approach to optimize and balance, at runtime, power consumption and real-time performance. This approach not only integrates hardware-level power-management techniques with real-time and quality-of-service (QoS) constraints, but also integrates a range of hardware power-management techniques at different granularities into a unified framework that allows better task-level control of power management. This enables finer control of power/performance tradeoffs by the operating system, and permits applications to dynamically adjust the system's and chip's power-saving configuration on a per-task basis to meet specific performance requirements while reducing power consumption. With the new functionality provided by this project, a richer class of power-aware applications can be developed to provide valuable user experiences in next-generation, real-time environments. The project specifically focuses the requirements of web server farms where energy and cooling costs are substantial and where the web services provide differentiated levels of QoS, thus motivating a power-aware framework that balances performance, QoS requirements, and energy costs.
