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 sampler touches into an exclusive soft ownership, and a budgeted cache replicates the context that it omits. Training samples only from this local view, so gradient synchronisation is the only per-step cross-worker traffic. A layout-independent full-neighbour evaluator exchanges hidden representations in place of features and, on the four graphs where we verified it, agrees with a single-machine reference to rounding level. We measured a two-layer GraphSAGE of hidden width 16 on OGBN-Arxiv, Reddit, OGBN-Products, IGB-hom-small and IGB-hom-medium at four to 16 nodes with one GPU each, and on four graphs at 32. At four to 16 nodes, SoftPart GNN completes training, placement included, in 22 to 62 percent of the time of a DistDGL baseline on the Deep Graph Library’s (DGL) original distributed sampler, and 21 to 52 percent with the baseline’s partitioning counted. Its test accuracy there stays within 0.006 of DistDGL’s on four of the five graphs. At 32 nodes, the ratio is 0.40 to 0.91. On Reddit, the accuracy is 0.0042 lower, and OGBN-Arxiv is left open. A paired control with modulo ownership and the same cache shows no detected accuracy difference. At four to 16 nodes, the warm-up shortens the training epoch on four graphs and gives no end-to-end gain.