Considerations in Workload Characterization for Parallel Access Volumes

In this paper we consider issues that may arise in the I/O workload characterization for Parallel Access Volumes. Limiting the characterization of I/O device service time to the mean and the standard deviation is sufficient to correctly predict the average I/O time in a realistic model of a single-exposure device. Things are less simple, however, when one considers devices with multiple parallel servers. We use two examples to illustrate the need for a more detailed workload characterization with multiple servers. We also briefly consider methods for matching real-life workload distributions to phase-type distributions, used to enable numerical solutions of queues with multiple servers, and we point out potential issues related to their use in modeling the performance of Parallel Access Volumes.