Declustering using fractals

A method for achieving declustering for Cartesian product files on M units is proposed. The focus is on range queries, as opposed to partial match queries that older declustering methods have examined. The method uses a distance-preserving mapping, the Hilbert curve, to impose a linear ordering on the multidimensional points (buckets); then, it traverses the buckets according to this ordering, assigning buckets to disks in a round-robin fashion. Because of the good distance-preserving properties of the Hilbert curve, the end result is that each disk contains buckets that are far away in the linear ordering, and, most probably, far away in the k-d address space. This is exactly the goal of declustering. Experiments show that these intuitive arguments lead to good performance: the proposed method performs at least as well as or better than older declustering schemes