The team examined the issue of tuning the performance of query evaluation without fundamentally changing the database management systems. The issue is important in the performance of the ADL testbed. The basic idea is to precompute a selected set of queries and store and index the answers, which are called ``materialized views''. When a user query comes, the query is rewritten (in a way transparent to the user) so that the materialized views are used instead of performing joins. The feasibility of using materialized views to improve performance is studied through experiments on the gazetteer (relational) database in which the largest relation has 6 million records. In particular, two materialized views were created according to the system profile and a set of frequent queries were selected and evaluated with and without materialized views. The results show a performance gain of 5 to 40 times using the materialized views on thes queries with a storage overhead of one and a half times of the original database. These results suggest the possibility of making a fully populated ADL operational.
Based on the findings from the experiments, an initial architecture to utilize the materialized view based performance tuning method in ADL is developed. The main components include algorithms to design materialized views, to translate user queries into equivalent queries that use materialized views, and to adaptively create materialized views at run time. Some of these results were reported in [5].