GNNAdvisor: An Adaptive and Efficient Runtime System for GNN Acceleration on GPUs
Talkat 15th USENIX Symposium on Operating Systems Design and Implementation (OSDI '21) · 2021-07-14..07-16·
summary
GNNAdvisor implements an adaptive GPU runtime that uses 2D workload management and community-aware node remapping plus a lightweight analytical model for automatic parameter selection. It partitions neighbors and embedding dimensions to reduce warp divergence and atomics, customizes shared-memory layouts, and remaps node IDs to improve locality. Evaluation on Quadro and V100 GPUs reports on-average 3.02× speedup over DGL and up to 4.10× over NeuGraph, with ~47% fewer atomics and ~57% less DRAM traffic.
AI-written summary · about
chapters (14)
- 0:00Intro
- 0:27Graphs are everywhere...
- 1:02Graph Analystics: Goals and Methods
- 2:24GNN: Graph Neural Networks
- 2:50Existing GNN Acceleration Solutions
- 3:56Overview of GNNAdvisor
- 6:04Input Extraction (cont'd)
- 7:162D Workload Management (cont'd)
- 8:40Specialized Memory Optimization (cont'd)
- 9:00Design Optimization
- 10:16Evaluation (cont'd): Overall Performance
- 10:39Evaluation (cont'd): Optimization Analysis
- 11:11Evaluation (cont'd): Additional Studies
- 11:38Key Focus & Contributions
related talks
15mMGG: Accelerating Graph Neural Networks with Fine-Grained Intra-Kernel Communication...
18mTC-GNN: Bridging Sparse GNN Computation and Dense Tensor Cores on GPUs
17mBGL: GPU-Efficient GNN Training by Optimizing Graph Data I/O and Preprocessing
15mIdentifying and Analyzing Pitfalls in GNN Systems
01h 52mNeural Network Accelerator Design
19mNeuGraph: Parallel Deep Neural Network Computation on Large Graphs
16mGMI-DRL: Empowering Multi-GPU DRL with Adaptive-Grained Parallelism
19mOptimizations and abstractions for sparse machine learning