ITI Student's Energy-Saving Technique Wins Award

Shuyi Chen

Information Trust Institute student Shuyi Chen has won the Best Student Paper award at the GreenMetrics 2009 Workshop for his paper "Blackbox Prediction of the Impact of DVFS on End-to-End Performance of Multitier Systems."

The objective of the GreenMetrics workshop is to study how improvements to or new uses of information and communication technology (ICT) can contribute towards activities intended to minimize climate change. In particular, it focuses on ideas for using ICT to reduce the environmental impact of non-ICT processes and for reducing the environmental impact of ICT itself, which has been estimated to be responsible for 2% of the global carbon footprint.

Chen's paper presented a metric he created with his collaborators that would allow developers to determine the potential performance impacts of using "dynamic voltage and frequency scaling," or DVFS, in a server setting. "DVFS is an energy-saving technique that has been extremely successful on laptop computers," explains Chen. "It holds great promise for data center environments too, but in that context, use of DVFS may have important ramifications on application response times." For example, studies have shown that users of e-commerce sites demand fast response times and are quick to give up and leave a slow site, so an online business might not be able to adopt an energy-savings strategy that degrades a system's end-to-end responsiveness.

Therefore, there has been a need for a way to predict the impact of DVFS's changes in processor frequency on a system's response time. Chen's paper introduces an approach called the "frequency gradient" that fills that need. Frequency gradients provide a simple metric that is easy to measure online with very little knowledge of the target system.

Specifically, the paper makes four technical contributions. It describes the development of a metric and runtime measurement system to predict the impact of CPU frequency changes on the end-to-end response time of a multitier system. It then explains how to enable the modeling of nonlinear response time vs. CPU frequency relationships by formulating gradients as a linear combination of nonlinear basis functions that are based on high-level, application-independent knowledge of system behavior. It goes on to show how to enable the chaining of gradients at different levels of granularity. Finally, it demonstrates that predictions made using frequency gradients are accurate regardless of application architecture, communication patterns, and configuration settings.

Shuyi Chen is a Ph.D. candidate in the Department of Computer Science at Illinois. The co-authors of his award-winning paper include Kaustubh R. Joshi, Matti A. Hiltunen, and Richard D. Schlichting of AT&T Labs, and Chen's advisor Professor William H. Sanders, who is the Director of ITI, Acting Director of the Coordinated Science Laboratory, and a Donald Biggar Willett Professor of Engineering in the Department of Electrical and Computer Engineering at Illinois.

GreenMetrics 2009 was held in Seattle, Washington on June 15, 2009 in conjunction with the SIGMETRICS/Performance 2009 joint conference, which combined the flagship conference of the Association for Computing Machinery's special interest group for computer systems performance evaluation and the flagship conference of the International Federation for Information Processing working group on performance modeling and analysis.