Abstract: The execution of data intensive Grid applications still raises several questions regarding job scheduling, data migration and replication. The optimisation techniques applied by these services significantly determine how fast a job can be executed and how quickly the user can get the execution results. In this paper new strategies are presented for scheduling the execution of data intensive, parameter-sweep applications. By taking into account the way applications access their data, the extended Grid middleware can achieve lower response times and earlier execution results. Therefore, the proposed improved Grid middleware (1) monitors the execution of jobs and gathers resource access information, (2) analyses the compiled information and generates a description of the behaviour of the job, and (3) uses this newly generated behaviour description to run scheduling algorithm(s). The proposed job behaviour description allows new schedulers to estimate the job completion time more precisely, therefore it leads to better scheduling decisions. Besides, the proposed monitoring technique and analyser tool enable the user to generate the required job behaviour description automatically.
Keywords: Clusters, data grids, job scheduling, execution time prediction, parameter-sweep applications
