Archived: 30 January 2004
University of Adelaide Honours thesis, November 2003.
Supervisors: Paul Coddington and Andrew Wendelborn
A grid is a distributed computing environment where a group of individuals and/or institutions can share computational resources. Resources that can be shared include computing cycles, data storage, software, licences and special scientific equipment. The success of grid computing will depend on the existence of grid-specific frameworks that provide core services, such as security, resource management, data management and information services, which one would need to construct a grid. The problem with most current resource management schemes is that they only deal with the computational requirements of applications. However, an increasing number of applications involve the geographically dispersed analysis of very large collections of measured or computed data. For example, the Belle experiment, situated at the KEK particle accelerator in Japan, involves a collaboration of approximately 400 people from 50 institutes worldwide and currently generates about 100TB of experimental data each year. The effective scheduling of jobs in such environments is a difficult task as the cost of data transfers may be quite significant. In this honours project, a new resource management scheme, which factors dynamic forecasts of the availability of the computational resources and performance of the network links into its cost models, has been implemented. Tests on the Belle testbed, consisting of nodes in Adelaide, Melbourne, Sydney and Canberra, have demonstrated that both factors must be considered in order to efficiently schedule data intensive tasks.