Diskbench: User-level Disk Feature Extraction Tool

Despite the increases in disk capacity and decreases in mechanical delays in recent years, the performance gap between magnetic disks and CPU continues to increase. To improve disk performance, operating systems and file systems must have detailed low-level information (e.g., zoning, bad-sector positions, and cache size) and high-level information (e.g. expected read and write performance for different access pattern) about the disks that they use. In this paper, we present Diskbench, our tool for extracting such information. Diskbench uses both interrogative and empirical methods for extracting various disk features. We present our extraction methods and results for several testbeds. From our empirical study, we conclude that intelligent data placement and access methods can be devised to improve disk performance, by exploiting low-level disk knowledge. Diskbench has benefitted our video storage research in the implementation of Semi-preemptible IO and guaranteed real-time scheduling.