Self-adaptive Graph Traversal on GPUs

Abstract

GPU's massive computing power offers unprecedented opportunities to enable large graph analysis. Existing studies proposed various preprocessing approaches that convert the input graphs into dedicated structures for GPU-based optimizations. However, these dedicated approaches incur significant preprocessing costs as well as weak programmability to build general graph applications. In this paper, we introduce SAGE, a self-adaptive graph traversal on GPUs, which is free from preprocessing and operates on ubiquitous graph representations directly. We propose Tiled Partitioning and Resident Tile Stealing to fully exploit the computing power of GPUs in a runtime and self-adaptive manner. We also propose Sampling-based Reordering to further optimize the memory efficiency of SAGE through a lightweight and effective node reordering technique on the fly. Extensive experiments demonstrate that SAGE can achieve superior graph traversal performance over existing approaches under different architectural scenarios, i.e., single-GPU, out-of-core, and multi-GPU.