Accelerate and Scale GNNs with Deep Graph Library and GPUs
, Senior Applied Scientist, AWS
Graphs play important roles in many applications, including drug discovery, recommender systems, fraud detection, and cybersecurity. Graph neural networks (GNNs) are the current state-of-the-art method for computing graph embeddings in these applications. Amazon and NVIDIA built a partnership on developing Enterprise Deep Graph Library (DGL) for large-scale GNNs. We'll discuss the recent improvements of DGL on NVIDIA GPUs in the DGL 0.9 release cycle resulting from the partnership. We've enabled seamless interaction of RAPIDS cuGraph and DGL to allow model developers to use graph analytics in GNN models easily. We'll show the benefits of having an integrated graph analytics and graph neural network environment for users wanting to calculate graph embeddings at scales. We also take advantage of unified virtual addressing to accelerate large-scale GNN training in a single machine and add mixed-precision training to accelerate the GNN computation on GPUs. In this release cycle, we further demonstrate the scalability of the entire pipeline of DGL’s distributed training, including data loading, distributed training and distributed inference, on graphs with billions of nodes and tens or even hundreds of billions of edges on a cluster of GPU machines. Finally, we'll talk about the DGL container built by the NVIDIA team to enable easy use of DGL on NVIDIA GPUs.