A Parallel Data Mining Algorithm for PageRank Computation

We study the utility of graphics processing units (GPUs) for an acceleration of the data mining PageRank algorithm and a reduction of the memory size of the web graph. We first present a new web graph representation using a compressed format in order to reduce the memory allocation of the web graph. Then, this web graph is simply partitioned into small chunks to be processed on the GPUs' device. The basic steps of the algorithm are then split up into parallel operations allowing to exploit the computing power of GPUs in the CUDA language as best as possible. In the experiments, we have tested the algorithm using GPUs with a set of real web data, and compared the computation with a CPU-based one. The obtained results show that the proposed PageRank computation on GPUs outperforms the CPU version by a factor of 100, reducing at the same time the web graph memory storage by 93, 928%.