Static and dynamic overprovisioning strategies for performance consistency in grids

It is not uncommon that grid users observe highly variable performance when they submit similar workloads at different times. From the users' point of view, such inconsistent performance is undesirable, and it leads to user dissatisfaction and confusion. We tackle this performance inconsistency problem using overprovisioning which is increasing the system capacity by a factor that we call the overprovisioning factor (K). Although overprovisioning is not cost effective, its simplicity makes it the preferred method for providing performance guarantees. Hence in this work, we present a realistic investigation of overprovisioning in grids. Towards this end, first we present a performance and cost evaluation of static and dynamic overprovisioning strategies. We find that the dynamic overprovisioning strategy, for which we use computing clouds, provides better consistency with lower costs compared to static strategies, and overprovisioning beyond a certain value of K (in our case K=2.5) incurs significant costs without significant consistency improvements. Then, we design and evaluate a feedback-controlled system to dynamically determine K to give performance guarantees to grid users. We show that our system determines K dynamically and provides significant improvements over the initial system, as high as 67%, in the number of jobs that meet the performance requirements.