Storing Trees on Disk Drives

Tree-structured data are abundant today, ranging from Bioinformatics suffix-tree alignments, to multi-resolution video, to directory-file hierarchies, to XML. The storage techniques employed by systems that manage tree-structured data greatly affect their performance. Current approaches either map the tree data to an underlying relational database system, or use the abstraction provided by a general-purpose object storage manager, or simply use flat files. These storage schemes, however, ignore the tree structure of the data as well as the characteristics of disk drives. Relational databases are structured tables and flat files are unstructured. On the other hand, disk drives store information in circular tracks that are accessed with mechanical seek and rotational overhead. The performance of disk drives greatly depends on the I/O access pattern (orders of magnitude difference between sequential and random access times). To the best of our knowledge, there exists no data layout strategy that accounts for the structural mismatch between tree-structured data and disk drive storage. We propose a new storage technique, tree-structured placement, that explicitly accounts for the mismatch between treestructured data and disk drive characteristics, so that common navigation operations (parent-to-child and node-to-nextsibling) are efficient. This technique uses the recently proposed idea of semi-sequential disk access [2] to place the tree structure. It also presents optimizations that reduce the on-disk space fragmentation and average random seek-times. Experimental evaluation using the DiskSim disk simulator [1] suggests as much as 80% reduction in query IO times compared to the default sequential layout of tree-structured data.