Highly concurrent multidimensional index for spatial database management systems

In this paper, we present highly concurrent R-link algorithms which use a novel (cheap and efficient) technique to detect node-split. A transaction S-latches one node at a time during tree traversal when fetching (window or exact-match query) or deleting spatial objects, while a transaction inserting spatial object latch-couples and U-latches two pages at a time. Insertion of new object and the triggered tree-structure-modifications (node-split, linking a child to its parent, increase-tree-height) are performed in one pass over the R-link tree. To simplify recovery, each tree-structure-modification is executed as a single atomic action involving pages on a single level of the tree, and logged using a single redo-only log record. The algorithms keep the R-link tree balanced during normal processing and after transaction aborts (or system failure) to guarantee logarithmic search path. Our approach can work for a large class of spatial index trees that are based on R-tree.