A generalized system performance model for object-oriented database applications

Although relational database systems have met many needs in traditional business applications, such technology is inadequate for non-traditional applications such as computer-aided design, computer-aided software engineering, and knowledge bases. Object-oriented database systems (OODB) enhance the data modeling power and performance of database management systems for these applications. Response time is an important issue facing OODB. However, standard measures of on-line transaction processing are irrelevant for OODB . Benchmarks compare alternative implementations of OODB system software, running a constant application workload. Few attempts have been made to characterize performance implications of OODB application design, given a fixed OODB and operating system platform. In this study, design features of the 007 Benchmark database application (Carey, DeWitt, and Naughton, 1993 ) were varied to explore the impact on response time to perform database operations Sensitivity to the degree of aggregation and to the degree of inheritance in the application were measured. Variability in response times also was measured, using a sequence of database operations to simulate a user transaction workload. Degree of aggregation was defined as the number of relationship objects processed during a database operation. Response time was linear with the degree of aggregation. The size of the database segment processed, compared to the size of available memory, affected the coefficients of the regression line. Degree of inheritance was defined as the Number of Children (Chidamber and Kemerer, 1994) in the application class definitions, and as the extent to which run-time polymorphism was implemented. In this study, increased inheritance caused a statistically significant increase in response time for the 007 Traversal 1 only, although this difference was not meaningful. In the simulated transaction workload of nine 007 operations, response times were highly variable. Response times per operation depended on the number of objects processed and the effect of preceding operations on memory contents. Operations that used disparate physical segments or had large working sets relative to the size of memory caused large increases in response time. Average response times and variability were reduced by removing these operations from the sequence (equivalent to scheduling these transactions at some time when the impact would be minimized)