A lock-free approach to parallelizing personalized PageRank computations on GPU