A combined method for maintaining large indices in multiprocessor multidisk environments

Abstract

Consider the problem of maintaining large indices (or secondary memory indices) in a multiprocessor multidisk environment in which each processor has a dedicated secondary memory (one disk or more). The processors either reside in the same site and communicate via shared memory, or reside in different sites and communicate via a local broadcast network. The straightforward method (SFM) for maintaining such an index, which is commonly called declustering, is to partition the index records equally among the processors, each of which maintains its part of the index in a local B/sup +/-tree. In prior work (Inform. Processing Lett., vol. 34, pp. 313-321, May 1990), we have presented another method, called the "totally distributed B/sup +/-tree" (TDB) method, in which all processors together implement a "wide" B/sup +/-tree. There are settings in which the second method is better than the first method, and vice versa. In this paper, we present a new method, called the combined distribution method (CDM), that combines the ideas underlying SFM and TDB. In tightly coupled environments, CDM outperforms both SFM and TDB in almost all practical settings (in many settings by more than 30%). This is shown by an approximate analysis and verified by simulations. Note that CDM's approach can improve performance in database systems that use a RAID (redundant array of inexpensive disks).< <ETX xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">&gt;</ETX>