Database performance evaluation in an indexed file environment

Reviewer: Geoffry Stanton Howard

This paper develops and validates two analytical prediction models: one for access time and another for response time in an indexed file context. While the validity of the models proves to be excellent, their generalizability is not as broad as is implied by the title and abstract of the paper. This is because the models rely on Zipfian assumptions about the distribution of access keys, and these assumptions are valid only for files containing natural language text or keywords. The author constructed and tested the prediction models in a bibliographic searching environment, so it should not be assumed that the results are valid in the broader domain of general purpose database management systems that are implemented using indexed files. The author clearly notes these limitations in the text, but they should also have been reflected in the title and abstract. A more appropriate title might have been “Performance Prediction Models for Indexed Bibliographic Databases.” Within the bibliographic searching realm, the prediction models are quite accurate. Testing the prediction equations against actual searches in the MEDLINE database, an R-square value of 0.95 was obtained for predictions of access time, and 0.66 for response time predictions. These R-square values reflect the proportion of overall variance in the criterion variable that was explained by the analytical prediction model using regression analysis. A possible limitation of the results is that the models were validated using data collected in 1979 in an IBM mainframe environment. The parameters for the response time prediction model that were valid in 1979 may not be accurate in 1987. This would be due not only to absolute improvements in CPU speed since then, but to enhancement of operating system task scheduling and priority assignment techniques. The paper does not discuss the operating system environment in which the data were collected. Important applications for accurate models of response time in bibliographic systems are discussed. The design of optimum index structures, determination of the amount of information that should be indexed for each citation, overall database size, limits on the number of subscribers, and pricing structures can all be efficiently investigated using the validated models. The author has made a very significant contribution toward the development of more generalizable prediction models.