Dr. Andreas Weißel, Prof. Dr. Frank Bellosa

Abstract:

Power management is recognized as an important research area for mobile devices, embedded systems and general purpose systems. There are several methods for reducing the energy consumption of individual components and of the whole system. But known methods often have the shortcoming that energy savings cause performance reduction. Therefore power management has to adapt to the applications and users performance demands. However there is a trade-off between the users performance demands and energy savings. This work presents an approach to automatically identify the performance demands of applications at runtime and this way guiding the power management policies.

This approach extends the student thesis of Matthias Faerber to system wide power management and several shortcomings are addressed. In the previous work the currently running application is identified; Here, application usage profiles are classified because one application can have different usage characteristics depending on its current job. Furthermore the heuristic classification done by Faerber is replaced by a theoretically sound classification and training algorithm based on Classification And Regression Trees.

The presented approach gives the user the possibility to specify his personal minimal performance demands for different application usage profiles. Then multiple resource usage characteristics are retrieved from the CPU, the wireless network interface card and the hard disk for each application. Upon those usage characteristics it is possible to identify the usage profile of the currently running application and to dynamically apply the adequate user-defined power management setting. To account usage statistics for each application, the abstraction of resource containers is used. The mapping of resource usage characteristics to the appropriate power management setting is done by a classification demon. This classifier is trained by supervised learning with the Classification And Regression Tree algorithm.

The proposed approach for adaptive power management was evaluated on an iPAQ running a modified Linux kernel and the classification demon. Several applications that are typical for such a mobile platform were tested. A rate of correct classifications of the user-specific performance demands and the corresponding power management settings of approximately 98% was achieved.

BibTex:

@diplomathesis{weinlein05applspecpm,
  author = {Thomas Weinlein},
  title = {Application-Specific Energy Management in Operating Systems},
  type = {Diploma Thesis},
  address = {Operating System Group, University of Erlangen, Germany},
  month = jan # "~17",
  year = 2005,
  url = {http://www4.informatik.uni-erlangen.de/DA/pdf/DA-I4-2005-01-Weinlein.pdf}
}