Identifying and Analyzing Pitfalls in GNN Systems
summary
Analysis finds that omitting training accuracy hides system bugs and yields misleading optimizations: framework runtime overhead dominates for small graphs (node count < 4M), and DGL’s PyTorch integration inflates memory use by about 12 GB memory overhead. A new prototype, GRAPHPY, establishes baseline runtimes and memory and outperforms prior systems; a prior kernel-fusion system (DGN) runs 1.48x slower and gives only 6.4% memory saving. Recommendations are to report accuracy, use mid-size datasets, measure framework overhead, and compare to GRAPHPY.
AI-written summary · about
chapters (13)
- 0:00GNN system pitfalls
- 1:45Graph storage formats
- 3:53Sparse kernels in GNN
- 4:13Framework runtime overhead
- 6:26Training time impact
- 7:31Memory overhead in DGL
- 9:20Evaluation pitfalls in GNN
- 10:25Design pitfalls affecting accuracy
- 11:51Missing transpose kernels impact
- 12:53Backward operations errors
- 13:16GraphPi performance and accuracy
- 13:40Kernel fusion benefits
- 14:21Summary and open questions
related talks
13mGNNAdvisor: An Adaptive and Efficient Runtime System for GNN Acceleration on GPUs
18mTC-GNN: Bridging Sparse GNN Computation and Dense Tensor Cores on GPUs
15mMGG: Accelerating Graph Neural Networks with Fine-Grained Intra-Kernel Communication...
17mBGL: GPU-Efficient GNN Training by Optimizing Graph Data I/O and Preprocessing
29mGraph Neural Networks + LLM on
19mNeuGraph: Parallel Deep Neural Network Computation on Large Graphs
16mPreview: Graph NNs
31mGraph-Based Machine Learning with Python