Using the Distiller to direct the development of self-configuration software

Many storage systems have become so complex that that the system administrator's salary represents almost half of the total cost of ownership. One approach to reducing this cost is to develop storage systems that can configure and manage themselves. Unfortunately, our ability to develop such software has been hindered by a limited understanding of how workloads and storage systems interact. In earlier work we presented the design of the Distiller - our tool that automates the process of finding a workload's key performance-affecting attributes. In this paper, we distill three production workloads and show that the values of the chosen attributes contain information that will help self-configuring disk array to choose a reasonable prefetch length and RAID stripe unit size. We also discuss how the chosen attributes may help direct the development of algorithms that compute near-optimal prefetch lengths and stripe unit sizes.