BatchGNN: Efficient CPU-Based Distributed GNN Training on Very Large Graphs
This work addresses the challenge of scalable GNN training for large-scale graph data, though it is incremental as it builds on existing distributed systems with optimizations.
The paper tackles the problem of efficiently training Graph Neural Networks (GNNs) on very large graphs by introducing BatchGNN, a distributed CPU-based system that reduces communication overhead through techniques like macrobatching and caching, achieving an average 3x speedup over DistDGL on OGBN graphs and scaling to terabyte-sized graphs.
We present BatchGNN, a distributed CPU system that showcases techniques that can be used to efficiently train GNNs on terabyte-sized graphs. It reduces communication overhead with macrobatching in which multiple minibatches' subgraph sampling and feature fetching are batched into one communication relay to reduce redundant feature fetches when input features are static. BatchGNN provides integrated graph partitioning and native GNN layer implementations to improve runtime, and it can cache aggregated input features to further reduce sampling overhead. BatchGNN achieves an average $3\times$ speedup over DistDGL on three GNN models trained on OGBN graphs, outperforms the runtimes reported by distributed GPU systems $P^3$ and DistDGLv2, and scales to a terabyte-sized graph.