Parallel pattern mining on Graphics Processing Units

Abstract

Frequent pattern mining is a field with many practical applications, where large computational power and speed are needed. Many state-of-the-art frequent pattern mining applications are an inefficient solutions for both shared memory and multiprocessor systems due to problems with parallelism and memory. One of possible solutions to the problem is the use of Graphics Processing Unit (GPU) in the system along with modification of classical pattern mining algorithms in such a way, that the sequential part of algorithm is run on host and the parallel part on GPU. Such solution allows for considerable speed-up (of up to two orders of magnitude), but for more complicated problems and FPM algorithms it can be hard to achieve. So far there were presented 3 modifications of the most basic Apriori algorithm for solving GPGPU (general-purpose computation on graphics hardware) problems. Each of proposed parallel implementations (PBI, TBI, GPA) is suited only for frequent itemset mining, furthermore even the best of them (TBI) obtains results only 2-5 times faster then CPU-based versions of FP-growth algorithm. On ICCC'12 there was presented parallel version of more complex GSP algorithm adjusted for GPGPU problems, capable of obtaining results 50-100 times faster then CPU-based version, with the same accuracy. Not only is the GPGPU implementation the fastest, but also it is suited for solving more complex frequent pattern mining problems. This paper presents extension of proposed modifications to more complex algorithms from Apriori family. The modifications are evaluated both theoretically and with use of experimental setup consisting of nVidia Tesla card and CUDA parallel computing platform. Solution proposed in the paper uses more complex, faster frequent pattern mining algorithms then GSP, making it well suited for solving real-time GPGPU problems for very large data sets.