The effect of collisions on the performance of reduction filters

The optimization of general queries in a distributed database management system is an important research topic. The problem is to select the best sequence of database operations that will process the query and minimize costs. Our algorithm, which is based on reduction filters, can process general queries consisting of an arbitrary number of relations and join attributes. Each query is represented by a graph and an adjacency list. Each relation is usually only processed once-to minimize data transfers. However, if a filter changes during use then certain relations must be processed again-a queue is used to record this information. This work has one serious limitation-we used a "perfect" hash function which means that collisions are not possible. Experiments showed that our algorithm could reduce data transfers very substantially. However, in a database collisions in the reduction filter are a reality and a problem. Such collisions mean that the hash-semijoin cannot achieve the same amount of reduction as a traditional semijoin would. We investigate the effect that collisions have on the performance of our algorithm.