A general approach to compression of hierarchical indexes

Abstract

Tree-structured indexes typically restrict the search domain level by level, which means that the search information can be encoded more and more compactly on the way down. This simple observation is here formulated as a general principle of index compression. Saving storage space is one advantage, but more important is reduction of disk accesses, because more entries can be packed into a page. The index fan-out can be increased, reducing the average height of the tree. The applicability of compression is studied for several popular one-and multidimensional indexes. Experiments with the well-known spatial index, R*-tree, show that with modest assumptions and simple coding, 30-40% reduction of disk accesses is obtainable for intersection queries. Compression of index entries can be used together with other index compaction techniques, such as quantization and pointer list compression.