Data Base System Performance Prediction Using an Analytical Model (Invited Paper)

Much progress has been made recently in developing strategies for data base design at both the logical and physical levels. Various approaches, some built into automated design aids, produce designs that are known to be "good" (or even "optimal" in some sense). The measurement criteria by which the designs are judged, however, are difficult to relate to some of the performance measures of importance to computer system managers and data base system users. Such performance measures include device utilizations, transaction throughputs, and the distribution of responsetimes. In this paper, we suggest an overall framework for assessing and predicting the effect on resource consumption, throughputs, and response times of a variety of physical and logical data base design decisions that affect performance. We use ananalytical model based, at the lowest level, on queueing network models. Queueing network models have already proven useful in understanding and predicting performance in many actual computer systems (with and without data base components). At higher levels of the analytical model, we establish a sequence of data base system workload descriptions, each one dependent on more performance related design decisions. By analytical techniques, the workload description at one level and a set of design choices are transformed into the workload description at the next lower (more fully specified) level. By this approach, many data base design alternatives can be represented by changes at a single level of the layered model. The design alternatives can be assessed with respect to their effect on a variety of performance measures, including record accesses, block accesses, physical disk transfers, throughputs, and mean response times. The presence of other workload components running concurrently on the same hardware configuration can also be taken into account.