Reproduce “FastGL: A GPU-Efficient Framework for Accelerating Sampling-Based GNN Training at Large Scale” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Zeyu Zhu, Peisong Wang, Qinghao Hu, Gang Li, Xiaoyao Liang, Jian Cheng

The FastGL framework accelerates sampling-based GNN training on large-scale graphs by optimizing subgraph sampling, memory I/O operations, and GNN computation. It achieves an average speedup of 11.8x over PyG and demonstrates robust scalability across various configurations without compromising training accuracy.

Key insights

  • Exploiting node overlap across consecutive mini-batches can significantly reduce memory I/O traffic by reusing data already on the GPU.
  • Optimizing data placement and access patterns within the GPU's memory hierarchy (shared vs. global memory) improves GNN computation efficiency.
  • Fusing distinct operations into a single kernel during GPU-based subgraph sampling eliminates synchronization overheads, accelerating the sampling phase.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.