Selective preemption strategies for parallel job scheduling

Service control with the preemptive parallel job scheduler Scojo-PECT

“…The checkpoint-and migration-based approach in [3] obtained benefits (though not considering preemption cost) if preempting jobs after a maximum of 1 hour runtime, provided that they are over a certain size limit. The approach in [15] demonstrates improvements in both average and worst-case response times for all job classes with suspension/resumption on the same resources. This approach imposes limits on relative sizes between preemptor and preemptees, while using the ratio of the different jobs' estimated relative response times to trigger preemptions.…”

“…Other options are to preempt jobs to disk in coarse-grain time sharing (switching between jobs in time intervals like tens of minutes or hours rather than of seconds or minutes as in gang scheduling) [7] or to only remove jobs from the system when necessary to fit other jobs into the machine [3,15]. Note that fine-grain or coarse-grain time sharing is safer as it avoids problems with "stranded" jobs, i.e., jobs which experience excessive wait times until they get access to the same resources again.…”

Integrated approach towards adaptive state-tracking job migration for maximising performance benefit

“…To overcome these problems, the adaptive job allocation [1], preemption scheme [2], and state-tracking migration [3,4] based scheduling mechanisms have been proposed as supplementary techniques to reduce the reallocation delays mentioned above. However, additional migration costs imposed by frequent job state traces and data movements have not been significantly investigated with various migration policies.…”

“…However, additional migration costs imposed by frequent job state traces and data movements have not been significantly investigated with various migration policies. Therefore, the effects of preemptive migration still remain uncertain, resulting in that the delayed execution frequently occurs due to the difference of jobs' importance [2]. Moreover, it has been found that makespan may not always be a main criterion to be optimised for the high-throughput computing [5].…”

“…This scheme is especially effective when a small size of jobs is dominant in the given workloads. Moreover, to avoid the delayed job execution we also adopt a selective preemption strategy [2] based on STMP (SP-STMP), so that a migrant job could compete with previously executed or reserved jobs at a backup resource. Therefore, to guarantee selective use of the benefit from each strategy, we propose that a job allocator adaptively integrates STMP with the two strategies (AI-STMP).…”