Speculative Scheduling of Parameter Sweep Applications using Job Behaviour Descriptions

Speculative Scheduling of Parameter Sweep Applications using Job Behaviour Descriptions

Attila Ulbert (Eötvös Loránd University, Hungary), László Csaba Lorincz (Eötvös Loránd University, Hungary), Tamás Kozsik (Eötvös Loránd University, Hungary) and Zoltán Horváth (Eötvös Loránd University, Hungary)
Copyright: © 2011 |Pages: 18
DOI: 10.4018/978-1-60960-603-9.ch005
OnDemand PDF Download:
$30.00
List Price: $37.50

Abstract

The execution of data intensive Grid applications raises several questions regarding job scheduling, data migration, and replication. This paper presents new scheduling algorithms using more sophisticated job behaviour descriptions that allow estimating job completion times more precisely thus improving scheduling decisions. Three approaches of providing input to the decision procedure are discussed: a) single job description, b) multiple job descriptions, and c) multiple job descriptions with mutation. The proposed Grid middleware components (1) monitor the execution of jobs and gather resource access information, (2) analyse the compiled information and generate a description of the behaviour of the job, (3) refine the already existing job description, and (4) use the refined behaviour description to schedule the submitted jobs.
Chapter Preview
Top

Our approach focuses on the resource access of jobs; the scheduling decisions are made based on the finishing time estimations exploiting the knowledge of the behaviour of jobs.

Nabrizyski et al. (Nabrizyski, Schopf, & Weglarz, 2003) gives an excellent overview of Grid resource management. Besides presenting a number of scheduling strategies (Ranganathan & Foster, 2003), in Chapter 16 W. Smith introduces new statistical prediction techniques for the execution times for applications. The first technique uses historical information of previous similar runs to form predictions. The similarity of runs are determined by categorising discrete characteristics of the submitted jobs. The second technique uses instance-based learning: a database of experiences is maintained and used to make predictions. Each experience consists of input and output features. The input feature is a simple job description (user name, job name, number of CPUs requested, requested operating system, etc.).

Complete Chapter List

Search this Book:
Reset