Dr. Andreas Weißel, Prof. Dr. Wolfgang Schröder-Preikschat

Abstract:

An important issue of mobile devices is energy consumption. There are different approaches on how to optimize the available energy. One method is to reduce the CPU frequency to save power. Other methods include scheduling strategies and if possible the complete shutdown of hardware components.

For power management, energy accounting, battery lifetime estimation and similar approaches it is necessary to measure how much energy was consumed before and after the changes. The measurement can be done with external hardware. In cases when there is no developer board available it can be difficult to measure the consumed energy, e.g. if you want to do tests on a mobile phone or organizer with embedded processors.

Therefore this thesis is an approach to provide energy consumption estimation without the need for external measurement hardware. Exemplary the Intel PXA 255 architecture will be used. It has two performance counters which make it possible to estimate the energy. The first part of this work is to get an energy consumption per event depending on the current processor and RAM frequency. The Intel PXA 255 processor supports different CPU and memory speeds. As a consequence there are different internal CPU voltage settings. So it is necessary to map different events to energy consumptions depending on the current CPU and memory frequencies.

There are only two performance counters and a variety of different applications with different usage of memory, processor and I/O. This thesis will show that two performance counters are not sufficient to estimate all types of programs in an accurate way but it is possible to train the energy weights for representing specific applications in an accurate way. It also evaluates the approach to switch the performance counters during runtime. This multiplexing technique gives a virtual view of more counters. The acquired results lead to the conclusion that the performance counter multiplexing is not sufficient to represent all types of applications at once but for a larger subset of applications compared to only two counters. For example it was possible to get an average estimation error below 2% with three tested real world applications.

BibTex:

@studythesis{fruth05intelpxa,
  author = {Florian Fruth},
  title = {Run-time Energy Characterization of the Intel PXA},
  type = {Study Thesis},
  address = {Operating System Group, University of Erlangen, Germany},
  month = apr # "~11",
  year = 2005,
  url = {http://www4.informatik.uni-erlangen.de/SA/pdf/SA-I4-2005-04-Fruth.pdf}
}