SNI-GNN: SmartNIC-Assisted Full-Graph GNN Training with In-Network Embedding Prediction

2026-08-06

AI Summary

This work addresses the high communication overhead and limited scalability of full-graph GNN training on multi-server clusters, primarily caused by frequent and irregular cross-node embedding exchanges. To overcome this, the study introduces SmartNICs into full-graph GNN training for the first time, proposing an in-network remote embedding prediction mechanism based on a lightweight linear trend predictor. This approach is synergistically combined with boundary node importance sampling and an asynchronous DPU-GPU pipeline to achieve efficient co-design of communication and computation. The method provides convergence guarantees under controlled error bounds, reduces communication volume by 21–45%, and achieves end-to-end speedups of 1.3–3.6× over BNS-GCN and up to 1.29× over SANCUS, with negligible accuracy loss (≤0.01). It further enables efficient scaling to graphs with tens of millions of edges across 16 GPUs.

Abstract

Full-graph GNN training delivers high accuracy but scales poorly on multi-server clusters due to heavy, irregular inter-node embedding exchanges. We present SNI-GNN, a SmartNIC-assisted full-graph training system that reduces communication while preserving accuracy by predicting remote embeddings in-network. SNI-GNN deploys a lightweight linear-trend predictor on SmartNICs to refine cached historical embeddings, coupled with an importance-based boundary-node sampling policy and an asynchronous DPU--GPU data pipeline with intermediate-result reuse. We provide error and convergence bounds showing that predictor bias remains controlled under bounded second-order dynamics and yields standard non-convex convergence with inexact gradients. Implemented on NVIDIA BlueField-3, SNI-GNN integrates with state-of-the-art full-graph systems, cuts communication by 21--45\%, achieves 1.3--3.6$\times$ end-to-end speedups over BNS-GCN and up to 1.29$\times$ over baseline SANCUS, with accuracy loss $\leq 0.01$, and scales efficiently to 16 GPUs on graphs with up to tens of millions of edges. These results indicate SmartNIC-based in-network prediction is a practical complement to partitioning and compression techniques for communication-efficient full-graph GNN training at scale.

Problem

• Graph Neural Networks
• Full-graph Training
• Communication Overhead
• Scalability
• Distributed Training

Innovation

• SmartNIC
• in-network prediction
• full-graph GNN training
• embedding communication reduction
• asynchronous DPU-GPU pipeline