It's not fair - evaluating efficient disk scheduling

Storage system designers prefer to limit the maximum queue length at individual disks to only a few outstanding requests, to avoid possible request starvation. In this paper, we evaluate the benefits and performance implications of allowing disks to queue more requests. We show that the average response time in the storage subsystem is reduced when queuing more requests and optimizing (based on seek and/or position time) request scheduling at the disk. We argue that the disk, as the only service center in a storage subsystem, is able to best utilize its resources via scheduling when it has the most complete view of the load it is about to process. The benefits of longer queues at the disks are even more obvious when the system operates under transient overload conditions.