Parallel partition-based join algorithms on a hypercube database computer

Abstract

The problem for improving performances of database management systems has attracted great interests during the two decades. This thesis is, especially, concerned with the problem for improving the performance of the join operation, which is one of the most frequently used and one of the most complex operations in the realtional database management systems. Two partition-based join algorithms, the sort-partitioned join and the select partitioned join algorithms, are proposed for reducing the number of secondary storage accesses in single processor computers. We further investigate for the effecient partitioning techniques and propose two improved versions of the sort-partitioned join and an improved version of the select-partitioned join. Analytic comparisons show that the two proposed partition-based join algorithms always outperforms the sort-merge join and performs better tnan the hash-partitioned join in most cases. The analyses for the proposed join algorithms are validated in practical situations by simulation experiments. The simulation experiments show that the improved version of the select-partitioned join performs best in most cases. We propose a general execution model for the paralle execution of the partition-based join algorithms on a hypercube database computer and, from the execution model, several parallel partition-based join algorithms are designed, analyzed, and compared. The hypercube database computer consists of a set of subcubes where each subcube shares a secondary storage device. The execution model consists of three phases; global partitioning phase, local partitioning phaes, and local joining phase. The partitioning schemes studied for the single processor computers are extended and applied for each execution phase of the execution model. Several promising parallel partition-based join algorithms are derived from the execution model and their preformances are analyzed in terms of computation cost, communication cost, and I/O cost. Fr...