May 2026· IEEE International Conference on Data Engineering· pp. 1831-1844· 0 citations· 61 references
Computer Science
TL;DR
SmartNIC-based in-network prediction is a practical complement to partitioning and compression techniques for communicationefficient full-graph GNN training at scale, and results indicate SmartNIC-based in-network prediction is a practical complement to partitioning and compression techniques for communicationefficient full-graph GNN training at scale.
Abstract
Full-graph GNN training delivers high accuracy but scales poorly on multi-server clusters due to heavy, irregular internode embedding exchanges. We present SNI-GNN, a SmartNICassisted full-graph training system that reduces communication while preserving accuracy by predicting remote embeddings innetwork. 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 intermediateresult reuse. We provide error and convergence bounds showing that predictor bias remains controlled under bounded secondorder 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 communicationefficient full-graph GNN training at scale.
CoRe-GNN is proposed, which performs both propagations in parallel at each layer: a coarsened inter-cluster term capturing long-range structure, and a local intra-cluster term preserving per-node discriminability.
Antonin Joly, Nicolas Keriven, Aline Roumy· 0 citations
SGS-GNN improves F1-scores by 4% relative to full training and up to 30% on heterophilic graphs and outperforms state-of-the-art methods by 4–7% at similar sparsity levels while reducing peak memory usage by up to 3.9×.
Siddhartha Shankar Das, Naheed Anjum Arafat, Muftiqur Rahman et al.· Proceedings of the 32nd ACM...· 0 citations
Taurus is presented, a single-machine system for GNN inference on graphs that do not fit in RAM, supporting both full-graph inference and fanout-sampled inference, and outperforms the strongest layer-wise baseline, DGI.
Many GNN-based link prediction models use a shared message-passing encoder to amortize node encoding across edges. However, in practice, neighborhood expansion makes the edge-level mini-batches approach near full-graph computation, weakening this benefit. Existing scalable techniques enable efficient mini-batch training but make recomputation decisions implicitly via structural access, either recomputing all accessed nodes or only the target nodes in the batch. We propose Access-Decoupled Recomputation (ADR), which treats recomputation as an explicit optimization decision while preserving the original graph structure without neighborhood sampling. We realize ADR as AFLiP, an efficient and structure-preserving framework for GNN-based link prediction. Across diverse models and datasets, AFLiP matches or exceeds full-recomputation accuracy, while sampling and static reuse exhibit dataset-dependent degradation. Compared to full recomputation, AFLiP reaches the best checkpoint up to 3.1× faster and scales to web-scale graphs where full recomputation is infeasible, and achieves up to 16× higher throughput than sampling on Friendster.
Jeongmin Bae, Min-Soo Kim· Proceedings of the 32nd ACM...· 0 citations
Graph Neural Networks (GNNs) are popular deep learning models designed to process graph-structured data through recursive neighborhood aggregations in the message passing process. When applied to semi-supervised node classification, the message-passing enables GNNs to understand short-range spatial interactions, but also causes them to suffer from over-smoothing and over-squashing. These challenges hinder model expressiveness and prevent the use of deeper models to capture long-range node interactions (LRIs) within the graph. Popular solutions for LRIs detection are either too expensive to process large graphs due to high time complexity or fail to generalize across diverse graph structures. To address these limitations, we propose a mechanism called information flow control, which leverages a novel connectivity measure, called information flow score, to address over-smoothing and over-squashing with linear computational overhead, supported by theoretical evidence. Building on this mechanism, we introduce DeltaGNN, to the best of our knowledge among the first scalable (featuring linear computational and memory complexity overhead) and generalizable (capable of effectively handling graphs with diverse homophily, density, and topology) architectures for long-range and short-range interaction detection. We benchmark our model across 10 real-world datasets, including graphs with varying sizes, topologies, densities, and homophilic ratios, showing superior performance with limited computational complexity.
Kevin Mancini, Islem Rekik· IEEE Transactions on Pattern...· 2 citations
A deeply fused GraphSAGE-GRU cell that embeds independent inductive GraphSAGE(SAmple and aggreGatE) encoders directly into each GRU gate, enabling simultaneous spatio-temporal feature extraction at every time step while remaining topology-agnostic.
Xuran Chen· Poster Volume 0008 The 2026...· 0 citations