Scheduling jobs on parallel systems using a relaxed backfill strategy

Backfill is a technique in which lower priority jobs requiring fewer resources are initiated before one or more currently waiting higher priority jobs requiring as yet unavailable resources. Processors are frequently the resource involved and the purpose of backfilling is to increase system utilization and reduce average wait time. Generally, a scheduler backfills when the user-specified run times indicate that executing the lower priority jobs will not delay the anticipated initiation of the higher priority jobs. This paper explores the possibility of using a relaxed backfill strategy in which the lower priority jobs axe initiated as long as they do not delay the highest priority job too much. A simulator was developed to model this approach; it uses a parameter omega to control the length of the acceptable delay as a factor times the wait time of the highest priority job. Experiments were performed for a range of omega values with both user-estimated run times and actual run times using workload data from two parallel systems, a Cray T3E and an SGI Origin 3800. For these workloads, overall average job wait time typically decreases as omega increases and use of user-estimated run times is superior to use of actual run times. More experiments must be performed to determine the generality of these results.