Prof. Dr. Frank Bellosa, Andreas Merkel

Abstract: 

The energy consumption of servers has become a major factor in their total costs of ownership. In 2005 the average server in the US consumed energy for US$ 264. However, there has been few research on predicting the energy efficiency of servers.

I argue that energy efficiency cannot be expressed as a single metric because a server's energy consumption is mostly determined by the workload. In this thesis I describe a system predicting the power consumption from an abstract description (profile) of a previously recorded workload. With this profile a company specifies the real workload of one of their servers. By replaying it on several machines it can be measured which machine is the most energy efficient one for the company's workload.

If this process is applied when a new server is bought, it leads to a much more energy efficient infrastructure. Consequently my system can be used to reduce the company's energy costs and save the environment.

I created an exemplarily implementation of this system for Linux 2.6 and the Pentium 4, but it can easily be ported to other platforms. It showed a prediction inaccuracy of < 4 % for test cases in my scope, making it a proper tool for it's purpose. For intentionally unsuitable test cases the prediction error was 12 %.

BibTex:

@diplomathesis{moltmann07workloadcharact,
  author = {Philip P. Moltmann},
  title = {Workload Characterization for Predicting Energy Efficiency of Servers},
  type = {Diploma Thesis},
  address = {System Architecture Group, University of Karlsruhe, Germany},
  month = sep # "~28",
  year = 2007,
  url = {http://i30www.ira.uka.de/teaching/theses/pasttheses/}
}