Unified-memory-based hybrid processing for partition-oriented subgraph matching on GPU

Subgraph isomorphism is a well known NP-hard problem that is to find all the matched subgraphs of a query graph in a large target graph. The state-of-the-art GPU-based solution is the vertex-oriented joining strategy, which is proposed by GSI. It effectively solves the problem of parallel write conflicts by taking vertices as processing units. However, this strategy may result in load-imbalance and redundant memory transactions. Moreover, due to the limitation of GPU memory, extending the existing frameworks to large-scale graph is difficult. To solve the first problem, we design a new storage structure Level-CSR and a new partition-oriented joining strategy. To avoid the influence of vertices with large degrees, we divide the dense vertices in traditional CSR into several GPU-friendly tasks and store them in Level-CSR. Then, an efficient execution strategy is designed based on the partitioned tasks. The partition strategy can reduce the load imbalance caused by the irregularity of real-world graphs, and further reduce the redundant global memory access caused by the redundant neighbor set accessing. To solve the second problem, we design a unified-memory-based hybrid processing strategy to support out-of-GPU subgraph matching. By leveraging unified-memory-based data movement and active vertex-based access mode switching strategy, our hybrid strategy can leverage the large storage of host memory and achieve efficient data movement between GPU and CPU. Besides, to further improve the performance, we propose a well-directed filtering strategy by exploiting a property of real-world graphs. The experiments show that compared with the state-of-the-art GPU based solutions, our approach can effectively reduce the number of unrelated candidates, minimize memory transactions, and achieve load balance between processors. And our unified-memory-based hybrid processing strategy can outperform the single access mode-based strategy under different parameter settings.