P3: Distributed Deep Graph Learning at Scale
Talkat 15th USENIX Symposium on Operating Systems Design and Implementation (OSDI '21) · 2021-07-14..07-16·
summary
P3 (pcube) is a distributed GNN runtime that separates graph structure from features, never ships full feature vectors across the network and instead pushes compact computation graphs and synchronizes intermediate activations. It uses column-wise feature partitioning, a hybrid model/data-parallel push‑pull pipeline, and optional structure caching. Prototype runs on DGL with 16 NVIDIA P100 GPUs and reports up to 7× speedup, feature traffic dropping from ~80% to ~20% and pipelining giving up to 50% improvement; benefits shrink when hidden dims grow.
AI-written summary · about
chapters (6)
- 0:00Introduction
- 0:36Graph Neural Networks
- 2:14Graph Processing Literature
- 7:25Hybrid Parallelism
- 11:12Results
- 13:43Summary
related talks
26mDeep Graph Library: Deep Graph learning at scale
14mLeapGNN: Accelerating Distributed GNN Training Leveraging Feature-Centric Model Migration
16mAccelerating Distributed Graph Learning by Using Collaborative In-Network...
17mBGL: GPU-Efficient GNN Training by Optimizing Graph Data I/O and Preprocessing
31mDGL Operator: Distributed Graph Neural Network Training with DGL and K8s
18mTC-GNN: Bridging Sparse GNN Computation and Dense Tensor Cores on GPUs
18mGPU-Disaggregated Serving for Deep Learning Recommendation Models at Scale
49mDistributed Deep Learning