PruneInfer: Exact Full-Neighborhood GNN Inference of Large Datasets on a Single GPU via Full Topology Pruning

Abstract

Graph neural network (GNN) inference in deployment often requires deterministic and exact predictions, which in turn require each inference run to aggregate complete dependency information from the full graph topology. However, over large graphs, full-graph forward propagation is usually infeasible on GPU due to limited memory, and slow on CPU due to slow computation. Existing systems address this limitation through exact full-neighborhood inference with different execution schemes. In these schemes, DGL-style systems repeatedly spill and reload intermediate embeddings between CPU and GPU, while GDL-style systems reduce this spill traffic by partitioning the graph but repeatedly transfer overlapping raw features and recompute overlapping intermediate embeddings across expanded subgraphs. This paper presents PruneInfer, an efficient single-GPU system for exact full-neighborhood GNN inference. PruneInfer fully prunes redundant topology across partitioned subgraphs and explicitly reuses previously computed intermediate embeddings to avoid redundant data movement and computation. To further reduce and hide communication overhead, PruneInfer incorporates F2-Cache, a frequency-guided two-tier cache, and a cross-layer asynchronous pipeline (CLAP). Experiments on four large real-world graphs and three representative GNN models show that PruneInfer achieves up to 4.49 × speedup over DGL and up to 7.05 × speedup over GDL-GNN across the evaluated datasets. PruneInfer is the only system on our experimental configuration that maintains accuracy comparable to full-graph inference across all evaluated models.