Improving the R*-tree Storage Allocation Algorithm

Geographic information systems (GISs) are one of the major applications of database systems. GISs are characterized by the huge amount of spatial data that has to be handled (in the order of thousands of terabytes, i.e. petabytes). This makes fast indexing techniques indispensable. The huge size of the databases causes indexing methods to be stored on slow secondary memory (i.e. disk), because they are too big to fit into primary memory. One of the most successful and applied indexing techniques is the R*-tree [Beckmann, 1990]. A lot of research has been conducted to improve its performance. However, one issue that has been overlooked is its storage on secondary memory. We present a hypothesis, that it is possible to exploit spatial relations between nodes to design and construct a (dynamic) R*-tree storage allocation algorithm that improves query processing by reducing seek time costs. We designed and implemented different types of storage allocation algorithms. The empirical results obtained from an algorithm that clusters the leaves of the R*-tree on disk (Dynamic Clustered Leaves) are used to sustain our hypothesis. Compared to algorithms that do not use spatial relations, a reduction of query time up to 64% can be achieved. However, insertions and deletions are up to 54% slower and the structure occupies around 60% more disk space.