Power Management for Server Clusters

Stephan Sigwart

Advisor: Andreas Weißel, Dr.-Ing. F. Bellosa

Registered as Studienarbeit SA-I4-2004-10 , April 5 2004

[Abstract] [Full Paper (pdf) , 4496 kB]

Power consumption is becoming a more and more important issue when there are many computers deployed in a server farm. Especially with up-to-date processors becoming faster and more complex, a significant fraction of the operation costs in a data center, which is hosting many servers, is actually spent on power consumption and cooling infrastructure. While there is already a lot of research done on how to balance incoming requests in such a cluster for best performance, there is nowadays a growing need to incorporate the aspects of power consumption.

There already exist some publications covering this research topic, presenting various methods and strategies as possible solutions. However, often different parameters and setups are used for the analysis forming the basis of those methods. Usage of different data as test-input, miscellaneous system platforms and varying assumptions about energy consumptions at idle- and peak-load are typical examples for such differences. Additionally, missing explanations about heuristics in the underlying algorithms do not allow an easy comparison between the quality of the presented results. So it is not possible to check whether or not those results are transferable to another system platform or workload situation. For this purpose, it would be necessary to check whether the heuristics and parameters used in the methods are tuned to fit the particular analyzed environment.

This work provides a synopsis about important aspects in cluster power management and points out characteristics by means of which different power management policies can be classified and analyzed. A survey of existing approaches is given, listing the particular underlying assumptions as well as a classification of the respective policy proposed to reduce the power consumption.

In addition, a simulator is presented which is capable of representing both different policies and scenarios by adjusting its parameters. This allows e.g. to simulate testing scenarios presented in various existing approaches with the same system platform parameters, so that an overall comparison can be done. Additionally, the heuristics of an algorithm may be investigated further. The simulator is using real traces from server logs which are identical to those applied in some existing publications. An implementation of several representative strategies is presented as an example and also used for analysis of their algorithms, parameters and heuristics.