GNNAdvisor: An Efficient Runtime System for GNN Acceleration on GPUs
share
As the emerging trend of the graph-based deep learning, Graph Neural Networks (GNNs) recently attract a significant amount of research attention from various domains. However, existing GNN implementations fail to catch up with the evolving GNN architectures, ever-increasing graph size, and node embedding dimensionality, thus, suffering from an unsatisfied performance. To break this hurdle, we propose GNNAdvisor, an efficient runtime system to systematically accelerate GNN applications on GPUs. First, GNNAdvisor spots the graph structure information (e.g., graph community) as a new driving force to facilitate GNN acceleration. Besides, GNNAdvisor implements a novel yet highly-efficient group-based workload management tailored for GNN computation to improve the thread-level performance on GPUs. GNNAdvisor further capitalizes on the GPU memory hierarchy for acceleration by gracefully coordinating the execution of GNNs according to the characteristics of the GPU memory structure. Moreover, GNNAdvisor incorporates a Modeling Estimating strategy to offer sufficient flexibility for automatic performance tunning across various GNN architectures and input datasets. Extensive experiments show that GNNAdvisor provides an average of 3.02x, 4.36x, and 52.16x speedup over the state-of-the-art GNN execution frameworks, Deep Graph Library (DGL), NeuGraph, and GunRock, respectively.
READ FULL TEXT