Performance Evaluation of Grid Based Multi-Attibute Record Declustering Methods

I/O subsystem is widely accepted as one of the principal bottlenecks for high performance parallel databases systems. The emergence of parallel I/O architectures has made the problem of data declustering, i.e. fragmenting a le of records and allocating the pieces to diierent disks, one of prime importance. This is evident from the growing activity in this area. In this study we focus only on multi-attribute declustering methods which are based on some type of grid-based partitioning of the data space. While a number of such declustering methods exist, we believe a good performance evaluation of their relative merits is lacking. Almost all performance analyses so far have been theoretical, where exact conditions on number of disks, sizes of attribute domains, and query shapes and sizes have been derived, for which a certain declustering method is optimal. Also, most conditions exist for partial match queries. We believe that in practice putting restrictions on the size of attribute domains is debatable and on the shape and size of queries is unacceptable. Thus, to answer the question how do various declustering schemes perform under a wide range of query and database scenarios (both relative to each other and to the optimal)?, we have carried out a detailed performance evaluation. Parameters that are varied include shape and size of queries, database size and number of attributes, and the number of disks. The theoretical contribution of this paper is in showing that there exists no declustering methods that is strictly optimal for range queries if the number of disks is more than 5. Our ndings are: (i) for large queries all methods perform almost the same and are close to optimal, (ii) there can be a substantial diierence for small queries, (iii) performance of the methods is quite sensitive to query shape, and (iv) the relative diierence between the methods' performance as well as their deviation from optimality decreases with the size and number of attributes in a query. Thus, we conclude that information about common queries on a relation ought to be used in deciding the declustering for it, and that this is especially crucial for small queries. Also, since there is no clear winner, parallel database systems must support a number of declustering methods.