Memory-Efficient GPU-Based Exact and Parallel Triangle Counting in Large Graphs

Abstract

Triangle counting is one of the basic research topics in many practical problems, such as clustering coefficients, transitivity ratio and traffic network complexity, etc., all of which can be converted into triangle counting issues. In this paper, a memory efficient GPU-based parallel triangle counting algorithm is proposed for large undirected graphs, which based on Polak's parallel forward triangle counting algorithm and improved the algorithm flow, edges sorting and triangle counting kernel function to increase the memory usage efficiency and run time performance. Algorithm comparison with the description of proposed algorithm steps and the implementation details is also introduced in the paper. Performance evaluations show that the proposed algorithm achieves 5 to 10 times speedup over the existing parallel forward algorithm, moreover, the GPU memory usage is reduced to a maximum of 50% compared to the existing algorithm.