Greedy scheduling with custom-made objectives

We present a methodology to automatically generate an online job scheduling method for a custom-made objective and real workloads. The scheduling problem comprises independent parallel jobs and parallel identical machines and occurs in Massively Parallel Processing systems and computational Grids. The system administrator defines the scheduling objective that may consider job properties and priorities of users or user groups. Our scheduling method combines a Greedy scheduling algorithm with the dynamic sorting of the waiting queue. This sorting algorithm uses a criterion that is modifiable by a set of parameters. Finding good parameter settings for the sorting criterion is viewed as a nonlinear optimization problem which is solved with the help of Evolution Strategies. We evaluate our scheduling method with real workload data and compare it to approximated optimal offline solutions and to the online results of the standard EASY backfill algorithm. Keywords Online job scheduling

• Type: Journal
• Year: 2010
• Where: ANOR
• Authors: Carsten Franke, Joachim Lepping, Uwe Schwiegelshohn
• Keywords: ANOR 2010, Greedy Scheduling Algorithm, Neural Networks, Online Job, Real Workload