Optimal load distribution for multiple classes of applications on heterogeneous servers with variable speeds. (17th April 2018)

• Record Type: Journal Article
• Authors: Li, Keqin
• Other Names: Heilig Leonard guestEditor.; Lalla‐Ruiz Eduardo guestEditor.; Voß Stefan guestEditor.; Buyya Rajkumar guestEditor.
• Is Part Of: Software, practice & experience. Volume 48:Number 10(2018)
• Page Start: 1805
• Page End: 1819
• Publication Date: 2018-04-17
• Subjects: average response time -- heterogeneous server with variable speed -- multiple classes of applications -- optimal load distribution -- optimal server speed setting -- power consumption -- task‐type–dependent server speed management; Computer software -- Periodicals; Computer programming -- Periodicals; Computer programs -- Periodicals; 005.3
• DOI: 10.1002/spe.2584
• Languages: English
• ISSNs: 0038-0644

Abstract

Summary: Performance and power are 2 significant issues in cloud computing. It is a critical issue on how to provide the best quality of service by consuming certain available power resource. For a given application environment and a given group of servers, optimal load distribution and optimal server speed setting can be an effective way to deal with the power‐performance tradeoff. The technique of variable and task‐type–dependent server speed management can be explored to optimize the server performance and to minimize the power consumption of a server with mixed applications. In this paper, we consider the problem of optimal load distribution for multiple classes of applications on heterogeneous servers with variable speeds. Given several classes of applications characterized by their arrival rates and expected execution requirements, several heterogeneous servers characterized by their power consumption parameters, and certain power supply, our problem is formulated as a multivariable optimization problem, ie, finding an optimal load distribution and an optimal server speed setting, such that the average task response time is minimized. To study the problem analytically, each server is treated as an M/G/1 queueing system with mixed classes of tasks such that both the average response time and the average power consumption can be calculated analytically. We define a power constrained performance optimization problem and develop a numerical algorithm to solve our optimization problem by solving a system of nonlinear equations. We also demonstrate numerical examples to show the effectiveness of our model and method. To the best of our knowledge, such analytical study of optimal load distribution and optimal server speed setting for multiple classes of applications on heterogeneous servers with variable speeds has not been available in the existing literature.