Paper

Alen Peacock. Dynamic Detection of Deterministic Disk Access Patterns. Master's thesis, Brigham Young University, 2001.

Abstract

As components within a system increase in performance at different rates, the components with the slowest rate of increase become bottlenecks for the entire system. Examination of bottlenecks and of strategies for improving their performance is vital to increasing overall system performance. In personal computing environments, hard disk drives are a good candidate for performance evaluation, since increases in processor speed and main memory speed far outstrip that of hard disks.

Many strategies for disk performance optimization have been proposed in the past. We focus our attention on detecting deterministic disk access patterns, or sequences of disk accesses which are demonstrably repeatable in normal computer usage, and we suggest several strategies which might be used to improve performance based on the deterministic sequences that are found. We implement a Linux-based disk trace collection tool and demonstrate that it has minimal impact on system performance. We explain a method and provide a system for finding deterministic patterns in real time, with analysis of the effects of such a system on overall system performance and discussion of the effects of tunable parameters to the system. We show that deterministic access patterns make up more than half of all disk accesses collected for this study, and discuss the sources of these patterns.