Dynamic voltage and frequency scaling (DVFS) is a well-known technique for gaining energy savings on desktop and laptop computers. However, its use in server settings requires careful consideration of any potential impacts on end-to-end service performance of hosted applications. In this paper, we develop a simple metric called the “frequency gradient” that allows prediction of the impact of changes in processor frequency on the end-to-end transaction response times of multitier applications. We show how frequency gradients can be measured on a running system in a push-button manner without any prior knowledge of application semantics, structure, or configuration settings. Using experimental results, we demonstrate that the frequency gradients provide accurate predictions, and enable end-to-end performance-aware DVFS for mulitier applications.
ASJC Scopus subject areas
- Hardware and Architecture
- Computer Networks and Communications