Reproduce “iSpLib: A Library for Accelerating Graph Neural Networks using Auto-tuned Sparse Operations” 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
Md Saidul Hoque Anik, Pranav Badhe, Rohit Gampa, Ariful Azad
Indiana University Bloomington researchers developed iSpLib, a specialized library that accelerates Graph Neural Network training and inference on CPUs by optimizing core sparse matrix operations. The library utilizes an auto-tuning mechanism and cache-enabled backpropagation to achieve up to 50x speedup over PyTorch baselines for GNN models like GCN, GraphSAGE, and GIN, without compromising model accuracy.
Key insights
- Sparse matrix operations (SpMM, SDDMM, FusedMM) are the primary bottleneck in GNN computations, particularly on CPU architectures.
- Auto-tuning kernel parameters to specific hardware and graph dataset characteristics significantly improves performance for sparse operations.
- Caching intermediate results during the forward pass can effectively accelerate the backward pass in GNN training by reducing redundant computations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.