Title:	PERFORMANCE COMPARISON OF SIZE-BASED SCHEDULING POLICIES IN CLUSTERED WEB SERVERS
Author(s):	Victoria Ungureanu , Benjamin Melamed , Michael Katehakis , Bin Zhou
ISBN:	972-98947-3-6
Editors:	Nuno Guimarães and Pedro Isaías
Year:	2004
Edition:	Single
Keywords:	Clustered Web server, scheduling policy, power-law distribution, simulation.
Type:	Full Paper
First Page:	1011
Last Page:	1017
Language:	English
Cover:
Full Contents:	click to dowload
Paper Abstract:	Web servers commonly use cluster-based architectures, which combine good performance and low cost. A cluster-based server consists of a front-end dispatcher and several back-end servers. The dispatcher receives incoming requests, and then assigns them to back-end servers for processing, which in turn serve the requests according to some discipline. It has been proven that giving priority to short job yields optimal performance for single (unclustered) servers. In this paper we are comparing the performance of two policies that apply this strategy to clustered Web servers. The first policy, called CDA (Class Dependent Assignment) calls for the dispatcher to assign short jobs as soon as they arrive in Round-Robin manner, while long jobs are assigned to back-end servers only when the servers become idle. Moreover, while a back-end server processes a long job, it will not be assigned any other jobs. The second policy, called SRPT (Shortest Remaining Processing Time), calls for a back-end server to schedule short jobs (or with short remaining processing time) with higher priority than long jobs. In effect, CDA uses the global information available to the dispatcher, while SRPT uses the local information available to back-end servers. Therefore, an experimental comparison between these two policies would contrast the performance metrics achieved when using local vs. global information, and it would quantify the performance differences. To gauge the performance of these policies, we exercised them on empirical data traces measured at Internet sites. The experimental results show that CDA outperforms SRPT by 40% with respect to the metrics of average waiting time and slow-down. These results confirm that using global information yields better performance than using local information and quantify the performance improvement.
	Go Back