Energy-Aware Queueing Models and Controls for Server Farms

Thumbnail Image

URL

Journal Title

Journal ISSN

Volume Title

Sähkötekniikan korkeakoulu | Master's thesis

Date

2014-06-16

Department

Major/Subject

Networking Technologies

Mcode

S3029

Degree programme

TLT - Master’s Programme in Communications Engineering

Language

en

Pages

82 + 7

Series

Abstract

Data centers are known to consume substantial amounts of energy. Together with the rising cost of energy, this has created a major concern. Server farms, being integral parts of data centers, waste energy while they are idle. Turning idle servers off may appear to eliminate this wastage. However, turning the server back on at the arrival of the next service request incurs a setup cost in the form of additional delays and energy consumption. Thus, a careful analysis is required to come up with the optimal server control policy. In this thesis, a queueing theoretic analysis of single server systems is carried out to determine optimal server control policies. Additionally, multiple server systems are also be studied through numerical methods. In this case, the task assignment policies that define how incoming requests are routed among the servers are also studied along with the control policies. The results of this study illustrate that the optimal control policy for a single server system leaves an idle server on or switches it off immediately when there is no request to serve. This is a general result that does not depend on service, setup and idling time distributions. However, in the case of multiserver systems, there is a plethora of choices for task assignment and server control policies. Our study indicates that the combination of the Join the Shortest Queue and Most Recently Busy task assignment policies can save up to 30% of the system cost if the control policy applied can wait for a specific amount of time before turning a server off. Moreover, a similar gain can be achieved by the simple Join the Shortest Queue task assignment policy when it is used along with a control policy that leaves an optimized number of servers on while switching the remaining servers off when they become idle.

Description

Supervisor

Aalto, Samuli

Thesis advisor

Lassila, Pasi

Keywords

server farms, queueing models, energy efficiency

Other note

Citation