PGLBox: An Industrial-Scale GPU‑Accelerated Graph Learning Framework

The presentation begins with an overview of graph learning, describing graphs as a universal language for complex systems such as social networks, biological molecules, knowledge graphs, and recommendation systems, and outlines the evolution from spectral‑based methods to spatial‑based and message‑passing architectures.

It then discusses the challenges of scaling graph neural networks, highlighting bottlenecks in graph storage, neighbor sampling, and CPU‑GPU data transfer, and introduces GPU acceleration strategies including full‑graph placement in GPU memory, Unified Virtual Addressing (UVA) for CPU‑GPU memory sharing, and multi‑GPU NVLink communication.

The PaddlePaddle Graph Learning (PGL) framework is introduced as a Paddle‑based library that provides graph engines, programming interfaces, and pre‑built models. Building on PGL, the PGLBox engine implements a fully GPU‑powered training pipeline that supports hierarchical storage, pass‑level pipelining, and efficient feature retrieval, enabling training on graphs with billions of nodes and edges.

Industrial use cases are described, showing how graph representation learning powers recommendation, item‑based and user‑based collaborative filtering, and large‑scale heterogeneous graphs. The Q&A section addresses GPU hash tables, caching strategies, pipeline consistency, and multi‑node scaling.

Overall, the article demonstrates that with modern GPU hardware (e.g., 8×A100 with 640 GB total memory) and the PGLBox system, large‑scale graph neural network training can achieve up to 28× speed‑up over CPU‑based solutions, making end‑to‑end graph learning feasible for production workloads.