![[2602.01872] Grappa: Gradient-Only Communication for Scalable Graph Neural Network Training](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.01872] Grappa: Gradient-Only Communication for Scalable Graph Neural Network Training
Summary
Grappa introduces a gradient-only communication framework for scalable training of Graph Neural Networks (GNNs), improving speed and accuracy while minimizing network load.
Why It Matters
As GNNs become more prevalent in various applications, optimizing their training processes is crucial. Grappa addresses the challenges of distributed training, offering a solution that enhances efficiency and accuracy, making it valuable for researchers and practitioners in machine learning and distributed computing.
Key Takeaways
- Grappa reduces communication overhead by using gradient-only updates during GNN training.
- The framework achieves up to 13x faster training speeds compared to existing systems.
- It maintains accuracy for deeper models through innovative gradient aggregation techniques.
- Grappa is model-agnostic and compatible with common deep-learning frameworks.
- The approach is effective on commodity hardware, making it accessible for broader use.
Computer Science > Distributed, Parallel, and Cluster Computing arXiv:2602.01872 (cs) [Submitted on 2 Feb 2026 (v1), last revised 16 Feb 2026 (this version, v2)] Title:Grappa: Gradient-Only Communication for Scalable Graph Neural Network Training Authors:Chongyang Xu, Christoph Siebenbrunner, Laurent Bindschaedler View a PDF of the paper titled Grappa: Gradient-Only Communication for Scalable Graph Neural Network Training, by Chongyang Xu and 2 other authors View PDF Abstract:Cross-partition edges dominate the cost of distributed GNN training: fetching remote features and activations per iteration overwhelms the network as graphs deepen and partition counts grow. Grappa is a distributed GNN training framework that enforces gradient-only communication: during each iteration, partitions train in isolation and exchange only gradients for the global update. To recover accuracy lost to isolation, Grappa (i) periodically repartitions to expose new neighborhoods and (ii) applies a lightweight coverage-corrected gradient aggregation inspired by importance sampling. We present an asymptotically unbiased estimator for gradient correction, which we use to develop a minimum-distance batch-level variant that is compatible with common deep-learning packages. We also introduce a shrinkage version that improves stability in practice. Empirical results on real and synthetic graphs show that Grappa trains GNNs 4x faster on average (up to 13x) than state-of-the-art systems, achieves better a...
