Logical, internal, and physical reference behavior in CODASYL database systems

This work investigates one aspect of the performance of CODASYL database systems: the data reference behavior. We introduce a model of database traversals at three levels: the logical, internal, and physical levels. The mapping between the logical and internal levels is defined by the internal schema, whereas the mapping between the internal and the physical levels depends on cluster properties of the database. Our model explains the physical reference behavior for a given sequence of DML statements at the logical level. Software has been implemented to monitor references in two selected CODASYL DBMS applications. In a series of experiments the physical reference behavior was observed for varying internal schemas and cluster properties of the database. The measurements were limited to retrieval transactions, so that a variety of queries could be analyzed for the same well-known state of the database. Also, all databases were relatively small in order to allow fast reloading with varying internal schema parameters. In all cases, the database transactions showed less locality of reference than do programs under virtual memory operating systems; some databases showed no locality at all. No evidence of physical sequentiality was found. This suggests that standard page replacement strategies are not optimal for CODASYL database buffer management; instead, replacement decisions in a database buffer should be based on specific knowledge available from higher system layers.