BibTeX
@MISC{Gandhi_performancemodeling,
author = {Anshul Gandhi},
title = {Performance modeling for data center power management},
year = {}
}
Share
 |
 |
 |
 |
||
OpenURL
Abstract
Data centers play an important role in today’s IT infrastructure. Government organizations, hospitals, financial trading firms, and major IT companies such as Google, Amazon, IBM, and HP, all rely on data centers for their daily business activities. However, the enormous energy consumption in data centers makes them very expensive to operate. On the one hand, it is desirable to limit the number of servers to reduce power consumption, but on the other hand, obtaining good response times requires having many servers available. This is one of many examples of the power-performance tradeoff faced by data centers today. Thus, an important concern is how to efficiently manage the power-performance tradeoff in data centers. In this thesis, we propose to design and implement power-management policies for data centers that optimize the power-performance tradeoff. Specifically, we propose to address several important, yet unanswered questions in data center power management, including: (i) How many servers are needed to handle the incoming load? (ii) When should servers be turned on/turned off/left idle/put to sleep? (iii) At what frequencies should servers be run? (iv) What policy should be used to route jobs to servers? In order to answer the above questions, we follow a two-pronged approach consisting of: 1. Performance Modeling: Performance modeling is a useful tool for analyzing the behavior of large computer systems, and it has been traditionally used to predict and improve system performance. However, power necessitates the development of new models and novel analysis involving multiple CPU operating frequencies, multiple server states (busy, idle, sleep and off) and the various setup costs involved in transitioning between server states. Thus, we propose to come up with new queueing-theoretic models that will allow us to analyze the various power-performance tradeoffs in data centers.
Keyphrases
data center thesis committee power-performance tradeoff many server performance modeling important role enormous energy consumption various setup cost unanswered question major company power-management policy data center today multiple cpu operating frequency good response time government organization daily business activity novel analysis system performance server state large computer system important concern new queueing-theoretic model various power-performance tradeoff power consumption new model two-pronged approach data center power management useful tool many example financial trading firm multiple server state
