Optimizing Distributed GNN Training on Large Graphs

Professor Wu, Chuan   (Principal Investigator (PI))

36

2021-09-01

1093580

Optimizing Distributed GNN Training on Large Graphs

Distributed ML System, Flow Scheduling, GNN Training, Graph Sampling, Placement

Network,Others - Computing Science and Information Technology

Engineering (E)

17207621

General Research Fund (GRF)

2021

On-going

1 [Algorithms for Communication and Computation Scheduling in Distributed GNN Training]: Design efficient and near-optimal graph sampling/GNN gradient communication and graph store/sampler/trainer execution scheduling algorithms, given graph store/sampler/trainer placement. 2 [Algorithms for Graph Store, Sampler and Trainer Placement]: Design efficient and near-optimal placement strategies for graph stores, samplers and trainers, jointly achieving GNN training time minimization with communication/computation scheduling. 3 [Joint Graph Partition, Sampling and Caching Design]: Design joint, efficient approaches for graph partition, sampling and graph data caching on training machines, to further reduce inter-machine traffic and expedite GNN training convergence. 4 [Implementation and Evaluation]: Implement a distributed GNN training system using our algorithms and strategies, and evaluate it with real-world GNN training workloads in AI clouds.