The Performance Impact of Advance Reservation Meta-scheduling

As supercomputing resources become more available, users will require resources managed by several local schedulers. For example, a user may request 100 processors, a telescope, network bandwidth and a graphics display in order to perform an experiment. In order to gain access to all of these resources (some of which may be in di erent geographical and administrative domains), current systems require meta-jobs like this to run during locked down periods when the resources are only available for meta job use. It is more convenient and eÆcient if the user is able to make a reservation at the soonest time when all of these resources are available. Low utilization during lock down periods can also be eliminated when meta-jobs are interleaved with existing local usage. System administrators are reluctant to allow reservations external to locked down periods because of the impact reservations may have on utilization and the Quality of Service that the center is able to provide to its normal users. This research quanti es the impact of advance reservations on supercomputing center metrics. It also outlines the algorithms that must be used to schedule meta-jobs. The Maui scheduler is used to examine metascheduling using trace les from existing supercomputing centers. These results indicate that advance reservations can improve the response time of supercomputing centers for meta-jobs, while not signi cantly impacting overall system performance. The appropriate balance between meta-jobs and local jobs is also speci ed using experimental results.