Parallel Join Algorithms based on Parallel B+-trees

Within the last several years, a number of parallel algorithms for the join operation have been proposed. However, almost all of the algorithms did not take advantages of the underlying parallel storage structures or data declustering methods of the operand relations. This paper introduces the concept of parallel storage structure or declustering aware parallel join algorithm. A parallel B+-tree index is presented and two classes of parallel join algorithms, which take advantages of the underlying parallel B+-tree index, are proposed and analyzed in the paper. First, the parallel B+-tree index is described. A parallel B+-tree index of a relation consists of a set of improved B+-tree indices (IB-tree for short) distributed among the processing nodes in a parallel computing system. Each IB-tree is built on a partition of the relation partitioned by one of the three strategies of range-partition, hash-partition and round-robin-partition. Second, two classes of parallel join algorithms based on the parallel B+-tree index are presented and analyzed. One is based on the range-partition strategy. The other is based on the hash-partition strategy. It is shown that the parallel execution times of the algorithms are linearly proportional to max{N/P, M/P}, where N and M are the numbers of tuples of the operand relations and P is the number of processing nodes. Finally, the two classes of parallel join algorithms are compared with the well known parallel join algorithms, parallel hybrid-hash join, parallel Grace-hash join and parallel merge-sort join, in practice. Theoretical and experimental results show that the algorithms based on parallel B+-tree are very efficient in case of at least one operand relation having a parallel B+-tree index on the join attributes.