[2025 PyTorch Conference Poster] Spectral Prefetcher: A Graph Neural Network-Based Storage Prefetcher

This repository contains the implementation of Spectral Prefetcher, a spectral graph neural network (inspired by SGDP) designed for rapid and precise I/O access pattern prediction in cloud storage systems. Link to paper.

🏗️ Architecture

Problem - Interleaved Access Pattern

Despite the widespread adoption of SSDs in cloud environments, I/O access (disk reads and writes) remains a significant performance bottleneck. To mitigate access latency, various prefetching techniques have been developed. However, as servers increasingly host multi-user workloads, the efficacy of traditional prefetchers diminishes.

Solution - Graph Neural Network to see a Global Picture

In this paper, we introduce "Spectral Prefetcher", a spectral graph neural network for rapid and precise predictions. This approach:

• ✅ 21.8% improvement in hit rate over baseline
• ✅ 79.9% reduction in training and inference latency over state-of-the-art GNN prefetcher
• ✅ 33.3% decrease in memory usage

📊 Results

Figure 2: Hit Rate Comparison

Our experimental results demonstrate significant improvements in hit rate across various workloads:

• Microsoft Research Cambridge traces
• Alibaba cloud traces
• Seagate enterprise workloads

The Spectral Prefetcher (SGDP) consistently outperforms traditional prefetchers (Linux kernel prefetcher, LEAP) and other ML-based approaches (LSTM, spatial GNN).

Figure 3: Performance Metrics

Comprehensive evaluation showing:

• Training time: 79.9% faster than baseline GNN
• Memory footprint: 33.3% reduction in model size

Dataset: Microsoft Research Cambridge Traces

The traces were collected by Microsoft Research Cambridge from a data center of Microsoft Windows servers over a one-week period in February 2007 and can be downloaded from the SNIA IOTTA repository:

Microsoft. MSR Cambridge Traces. http://iotta.snia.org/traces/388, 2022.

📖 Citation

If you use this code in your research, please cite:

@article{yang2024spectral,
  title={Leveraging AI for Faster Storage Access: a Graph-Neural-Network-Based Prefetcher},
  author={Yang, Zeyuan and Kurniawan, Daniar H.},
  journal={University of Chicago},
  year={2024}
}

Check out the original SGDP paper by Yiyuan Yang et al.

👥 Authors

• Zeyuan Yang - University of Chicago - faradawny@gmail.com
• Daniar H. Kurniawan - University of Chicago - daniar@uchicago.edu

📝 License

This project is part of academic research at the University of Chicago.

---

For questions or issues, please contact faradawny@gmail.com