Close to 1% of the world's electricity is consumed by computer servers. Given that the increased use of electricity raises costs and damages the environment, optimizing the world's computing infrastructure for power consumption is worthwhile. This thesis is one attempt at such an optimization. In particular, I began by building a cluster of 6 Intel Atom based low-power nodes to perform work analogous to data center clusters. Then, I installed a version of Hadoop modified with a novel power management system on the cluster. The power management system uses different algorithms to determine when to turn off idle nodes in the cluster.
Using the experimental cluster running a modified Hadoop installation, I performed a series of experiments. These tests assessed various strategies for choosing nodes to suspend across a variety of workloads. The experiments validated that turning off idle nodes can yield power savings. While my experimental procedure caused the apparent throughput to significantly decrease, I argue that using more realistic workloads would have yielded much better throughput with slightly reduced power consumption. Additionally, my analysis of the results, show that the percentage power savings in a larger, more realistically sized cluster would be higher than shown in my experiments.
| Identifer | oai:union.ndltd.org:CALPOLY/oai:digitalcommons.calpoly.edu:theses-1319 |
| Date | 01 June 2010 |
| Creators | Oppenheim, Brian Michael |
| Publisher | DigitalCommons@CalPoly |
| Source Sets | California Polytechnic State University |
| Detected Language | English |
| Type | text |
| Format | application/pdf |
| Source | Master's Theses |
Page generated in 0.0019 seconds