Parallel Association Rules Mining on GPU: CUDA

General-purpose computation on graphics processing unit (GPGPU) is playing an important role in super-computing. In this paper, we proposed a parallel association mining solution based on GPU using CUDA-- CudaApriori. The support-counting step for candidate frequent item-sets is off-loaded from CPU to GPU. Firstly, candidate frequent item sets and transactions are partitioned in the pattern of thread block and grid of thread blocks of GPU. Secondly, the task of support counting is performed in parallel by massive threads with the simple matching computation, suitable to stream access model of GPU. In our experimental work, we simulated transactions on both our CudaApriori and the standard Apriori. The result shows that CudaApriori produces. a 10-fold performance enhancement of frequent k-itemsets (k > 2) mining phase to Apriori and outperforms it by up to 80% at whole. In spite of the data, transmission of CudaApriori between GPU and CPU is the extra cost to Apriori, its performance reducing can be neglected with the high 2GB/s speed road.