| Reducing power consumption is important because (1) It prolongs the operational time between recharging batteries in portable systems. (2) It decreases heat generation and improves the reliability of electronic components. (3) The concept of green computers requires better energy efficiency.; This research focuses on using operating systems (OS) to dynamically control the power states of IO devices and processors. When a device or a processor has low utilization, it enters a lower-power state. This is called dynamic power management (DPM). DPM may degrade performance. One example is to spin down the plates in hard disks to save power. Spinning up the plates takes several seconds; read or write requests have to wait. Thus, it is essential to predict future utilization accurately to save power and maintain satisfactory performance.; OS can be structured as layers, including device drivers, process managers, and schedulers. Device drivers are the closest to hardware and can observe requests sent from software; however, they have only limited information about programs for predicting their requests. Process managers are closer to application programs; hence, they have more information regarding to request generation and have better accuracy for predicting hardware utilization. Moreover, process schedulers determine the execution order of programs and control when requests are generated; hardware utilization can be calculated accurately.; This thesis investigates how to use the information from OS to improve power management on interactive systems like a laptop computer. It describes the principles for designing power-management policies and a framework for implementing policies. It presents how to use process managers and schedulers to perform power management on IO devices. A method is proposed to reduce the power consumption of processors; this method inserts data buffers in application programs. The buffers are filled when the processors run in higher-power states; then, the buffers are drained while the processors enter lower-power states. It achieves nearly optimal power saving with only a few power states. Buffering also improves the response time of sporadic jobs. A graph traversal technique is used for efficiently assigning power states. |
