Filter-based join algorithms on uniprocessor and distributed-memory multiprocessor database machines

Abstract

The hybrid-hash algorithm and its parallel variant have been recently found to outperform all other algorithms in joining disk-based large relations on uniprocessor and ring-interconnected distributed-memory multiprocessor database machines. This paper presents several extensions to the centralized and distributed hybrid-hash algorithms. These extensions are based on the usage of one or more bit-vectors as filters. A comparative performance study to the presented algorithms is carried out. In a uniprocessor environment, this study shows that one of the proposed filter-based algorithms outperforms all of the other ones, including the hybrid-hash algorithm. In a distributed environment, the filter-based algorithms are found to suffer from a serious problem, namely, overloading the interconnection network with the transmission of large size bit-vectors. Different compression schemes are proposed to reduce the size of a transmitted bit-vector. The augmentation of the distributed version of best-performing centralized algorithm with one of the proposed compression schemes have been found to outperform all of the other algorithms and substantially improves the performance of the join operation.