A regression approach to performance analysis for the Differential File Architecture

The Differential File Architecture (DFA) has been previously proposed as a physical database design which confines the volatility of a large database into a small, more manageable file. Although this approach affords many advantages, the use of dual files often results in two accesses to secondary storage. The response performance of the DFA is therefore a key issue, yet the only predictive performance model currently in existence assumes an ideal environment. This paper represents a preliminary step toward developing a general model of DFA performance, applicable to realistic settings in which the distribution of key accesses and the hashing distribution of keys to filter switches are non-uniform. In support of a regression approach to the problem, a metric is developed and tested for suitability as a predictor of non-uniformity effects.