FreshGNN: Reducing Memory Access via Stable Historical Embeddings for Graph Neural Network Training
Summary: FreshGNN caches historical node embeddings to avoid repeated raw-feature loads in GNN mini-batch training, lowering GPU–CPU traffic. A gradient+staleness policy caches only stable embeddings, yielding 59% fewer memory accesses and 3.4–20.5x speedups on large graphs with <1% accuracy loss. (summarized by gpt-5-mini on Feb 09 2026)
Incoming Non-self Citations Over Time
Authors
- 1. Kezhao Huang (Tsinghua University)
- 2. Haitian Jiang (New York University)
- 3. Minjie Wang (Amazon)
- 4. Guangxuan Xiao (Massachusetts Institute of Technology)
- 5. David Wipf (Amazon)
- 6. Xiang Song (Amazon)
- 7. Quan Gan (Amazon)
- 8. Zengfeng Huang (Fudan University)
- 9. Jidong Zhai (Tsinghua University)
- 10. Zheng Zhang (Amazon)
BibTeX Citation
@article{huang_vldb24,
title = {{FreshGNN: Reducing Memory Access via Stable Historical Embeddings for Graph Neural Network Training}},
author = {Huang, Kezhao and Jiang, Haitian and Wang, Minjie and Xiao, Guangxuan and Wipf, David and Song, Xiang and Gan, Quan and Huang, Zengfeng and Zhai, Jidong and Zhang, Zheng},
journal = {PVLDB},
series = {{VLDB} '24},
volume = {17},
number = {6},
pages = {1473--1486},
doi = {10.14778/3648160.3648184},
url = {https://doi.org/10.14778/3648160.3648184},
year = {2024}
}Incoming Citations (Sorted by Pagerank)
Showing 5 of 5 citing papers.
| Rank | Citing Paper | Year | Venue | Pagerank |
|---|---|---|---|---|
| 5,079 | NeutronOrch: Rethinking Sample-based GNN Training under CPU-GPU Heterogeneous Environments | 2024 | VLDB | 6.3717342e-05 |
| 6,803 | OUTRE: An OUT-of-core De-REdundancy GNN Training Framework for Massive Graphs within A Single Machine | 2024 | VLDB | 5.7684339e-05 |
| 10,309 | A Comprehensive Benchmark on Spectral GNNs: The Impact on Efficiency, Memory, and Effectiveness | 2026 | SIGMOD | 5.093636e-05 |
| 10,890 | Heta: Distributed Training of Heterogeneous Graph Neural Networks | 2025 | VLDB | 5.093636e-05 |
| 10,975 | Faster Convergence in Mini-batch Graph Neural Networks Training with Pseudo Full Neighborhood Compensation | 2025 | VLDB | 5.093636e-05 |
Previous
Page 1 / 1
Next
Outgoing Citations (Sorted by Pagerank)
Showing 1 of 1 cited papers.
Citations counted here include only citations to other VLDB/SIGMOD/CIDR/PODS papers in this database.
| Rank | Cited Paper | Year | Venue | Pagerank |
|---|---|---|---|---|
| 1,132 | SANCUS: Staleness-Aware Communication-Avoiding Full-Graph Decentralized Training in Large-Scale Graph Neural Networks | 2022 | VLDB | 0.00012041292 |
Previous
Page 1 / 1
Next