SoftPart GNN: Distributed Graph Neural Network Training Without Offline Graph Partitioning
Distributed graph neural network training usually starts with an offline partitioner such as METIS, which fixes the layout before any workload is seen. SoftPart GNN is a distributed GraphSAGE runtime on PyTorch DistributedDataParallel (DDP) without one. A short online warm-up turns the vertices that the training sample...