Asynchronous parallel algorithm for mining association rules on a shared-memory multi-processors

Abstract

Mining association rules from large databases is an important problem in data mining. There is a need to develop parallel algorithm for this problem because it is a very costly computation process. However, all proposed parallel algorithms for mining association rules follow the conventional level-wise approach. On a shared-memory multi-processors, they will impose a synchronization in every iteration which degrades greatly their performance. The deficiency comes from the contention on the shared I/O channel when all processors are accessing their database partitions in the shared storage synchronously. An asynchronous algorithm APM has been proposed for mining association rules on shared-memory multiprocessors. All participating processors in APM generate candidates and count their supports independently without synchronization. Furthermore, it can finish the computation with less I/O than required in the level-wise approach. The algorithm has been implemented on a Sun Enterprise 4000 multi-processors with 12 nodes. The experiments show that APM has super performance than other proposed synchronous algorithms.