Automatic configuration and selection of power management policies

Although hundreds of papers have been written about software-directed power management, commodity computer systems lag fifteen years behind the state-of-the-art. Many power management policies use heuristics to improve power consumption. However, different heuristics are required to minimize the energy consumption of different applications. For software-directed power management techniques to become pervasive in commodity computer systems, policies must achieve significant power savings for a large number of applications. This dissertation describes two policies that may be analytically reconfigured at run-time based upon the current workload. The first policy derives a statistically optimal buffer size for producer-consumer applications that automatically configures its behavior based upon hardware parameters and workload data. The second policy mathematically formulates file system layout as a constrained optimization problem and minimizes power consumption using a genetic algorithm. The third part of this dissertation presents a technique for choosing a "good enough" policy from a library of existing policies for I/O devices with variable workloads where an optimal solution is unknown. This technique, called the Homogeneous Architecture for Power Policy Integration (HAPPI), compares multiple policies simultaneously at run-time to select the best available policy for the current workload and hardware configuration. The methods in this dissertation automatically configure and select power policies, remove administrators and users from making power management decisions, and significantly lower the barrier to deploying power management on general purpose computers.