power aware 32 - Requester-Aware Power Reduction...

Info iconThis preview shows pages 1–2. Sign up to view the full content.

View Full Document Right Arrow Icon
Requester-Aware Power Reduction Yung-Hsiang Lu, Luca Benini, Giovanni De Micheli CSL, Stanford University, USA. { luyung, nanni } @stanford.edu DEIS, Universit`a di Bologna, Italy. lbenini@deis.unibo.it Abstract Typically, power reduction is conducted by hardware techniques, such as varying clock frequencies and/or supply voltages. However, hardware devices consume power to serve the requests from software programs. Consequently, it is essential to consider software for power reduction. This paper proposes “requester-aware” power reduction through the collaboration with pro- grams. Experimental results show that this approach can save nearly 70% power with negligible performance degradation. 1. Introduction Power reduction has become a major goal in designing electronic systems. For portable systems, low power means longer battery life or lighter battery weight. For desktop computers or servers, low power reduces elec- tric bills and improves reliability. Many run-time power-reduction techniques have been proposed. Dynamic power management is a power- reduction technique that puts idle hardware devices into low-power sleeping states to reduce power consump- tion [3]. A device is idle if it has no request to serve; it can sleep to save power. When new requests arrive, the device wakes up entering a working state to serve these requests. Requests are generated by programs (also called requesters ). For example, when a program reads or writes a file on a hard disk, it generates IO requests on the hard disk. Similarly, when a program sends a packet through networks, it generates a request on the network card. Power state changes have overhead: additional energy consumption and delay [9]; therefore, a device should sleep only when the overhead can be justified by the amount of energy saved. Power managers (PM) deter- mine power states according to certain rules (also called policies ). In the past, power management was mainly implemented in two ways. In the first approach, power managers observe requests at devices to predict future workloads; they are implemented in hardware or device drivers without direct interaction with requesters [3]. On the other extreme, programs can directly control power states through Microsoft’s OnNow [10] application pro- gramming interface (API). We believe neither extreme is appropriate. This paper proposes one approach between the two extremes. It considers how requests are generated— by running pro- grams. Programs can specify what devices are needed before generating requests. Power states are affected, but not controlled, by individual programs. Making pro- grams aware of power management is suggested in [5]; however, no study has been devoted to examining ap- propriate interaction between power managers and pro- grams. This paper has three major contributions. First, we explain the advantages to distinguishing individual requesters. Second, we propose a performance-based API which allows programs to indicate their device re-
Background image of page 1

Info iconThis preview has intentionally blurred sections. Sign up to view the full version.

View Full DocumentRight Arrow Icon
Image of page 2
This is the end of the preview. Sign up to access the rest of the document.

Page1 / 6

power aware 32 - Requester-Aware Power Reduction...

This preview shows document pages 1 - 2. Sign up to view the full document.

View Full Document Right Arrow Icon
Ask a homework question - tutors are online