high performance computing on graphics processing units: hgpu.org

In this paper, we propose a novel method, GSM (Gunrock Subgraph Matching), to compute graph matching (subgraph isomorphism) on GPUs. In contrast to previous approaches, GSM is BFS-based: possible matches are explored simultaneously in a breadth-first strategy and thus can be mapped onto GPUs in a massively parallel fashion. Our implementation on the Gunrock graph analytics framework follows a filtering-and-verification strategy. While previous work requires one-/two-step joining, we use one-step verification to decide the candidates in current frontier of nodes. Our implementation has a speedup up to 4x over previous GPU state-of-the-art implementation.