BGL: GPU-Efficient GNN Training by Optimizing Graph Data I/O and Preprocessing
summary
Distributed GNN training system BGL builds a dynamic cache engine and a two-level CPU+GPU cache that together cut feature-retrieval traffic and reduce feature-retrieval time by about 80%, increasing cache hit ratio (proximity-aware BFS ordering raises hits from ~8% to ~14% in their figure). A new multi-level graph partitioning cuts cross-partition communication by 25–44%, profiling-based resource isolation balances pipeline stages, and end-to-end BGL yields ~1.9x mean speedup (up to 2.7x) over prior systems.
AI-written summary · about
chapters (8)
- 0:00Introduction to GNN
- 1:00Training bottlenecks
- 2:13Identifying core challenges
- 4:10Proposed BGL architecture
- 5:31Technical design details
- 10:00Optimization strategies
- 12:51Evaluation and results
- 15:37Conclusion
related talks
20mLegion: Automatically Pushing the Envelope of Multi-GPU System for Billion-Scale...
15mMGG: Accelerating Graph Neural Networks with Fine-Grained Intra-Kernel Communication...
14mGLIST: Towards In-Storage Graph Learning
14mLeapGNN: Accelerating Distributed GNN Training Leveraging Feature-Centric Model Migration
18mTC-GNN: Bridging Sparse GNN Computation and Dense Tensor Cores on GPUs
13mGNNAdvisor: An Adaptive and Efficient Runtime System for GNN Acceleration on GPUs
16mAccelerating Distributed Graph Learning by Using Collaborative In-Network...
19mOptimizations and abstractions for sparse machine learning