Self-adaptive, on-line reclustering of complex object data

A likely trend in the development of future CAD, CASE and office information systems will be the use of object-oriented database systems to manage their internal data stores. The entities that these applications will retrieve, such as electronic parts and their connections or customer service records, are typically large complex objects composed of many interconnected heterogeneous objects, not thousands of tuples. These applications may exhibit widely shifting usage patterns due to their interactive mode of operation. Such a class of applications would demand clustering methods that are appropriate for clustering large complex objects and that can adapt on-line to the shifting usage patterns. While most object-oriented clustering methods allow grouping of heterogeneous objects, they are usually static and can only be changed off-line. We present one possible architecture for performing complex object reclustering in an on-line manner that is adaptive to changing usage patterns. Our architecture involves the decomposition of a clustering method into concurrently operating components that each handle one of the fundamental tasks involved in reclustering, namely statistics collection, cluster analysis, and reorganization. We present the results of an experiment performed to evaluate its behavior. These results show that the average miss rate for object accesses can be effectively reduced using a combination of rules that we have developed for deciding when cluster analyses and reorganizations should be performed.