Group-based optimization for parallel job scheduling in clusters via heuristic search

Job scheduling for parallel processing typically makes scheduling decisions on a per job basis due to the dynamic arrival of jobs. Such decision making provides limited options to find globally best schedules. Most research uses off-line optimization which is not realistic. We propose an optimization on the basis of limited-size dynamic job grouping per priority class. We apply heuristic domain-knowledge-based hi-level search and branch-and-bound methods to heavy workload traces to capture good schedules. Special plan-based conservative backfilling and shifting policies are used to augment the search. Our objective is to minimize average relative response times for long and medium job classes, while keeping utilization high. The scheduling algorithm is extended from the SCOJO-PECT coarse-grain pre-emptive time-sharing scheduler. The proposed scheduler was evaluated using real traces and Lublin-Feitelson synthetic workload model. The comparisons were made with the conservative SCOJO-PECT scheduler. The results are promising--the average relative response times were improved by 18-32 while still able to contain the loss of utilization within 2