Leveraging AI for Faster Storage Access: A Graph-Neural-Network-Based Prefetcher
summary
A graph neural network–based prefetcher converts raw addresses into deltas then builds a weighted graph with temporal edges plus low‑weight fully connected edges so the model sees a global view of interleaved multi‑tenant IO. It trains a GNN (not an LSTM) on Microsoft Alibaba and Tencent traces and reports 33% runtime savings, 33% smaller model size and 79% less training time. The pipeline outputs next‑address predictions used to prefetch into CPU cache and reduce storage stalls.
AI-written summary · about
chapters (9)
- 0:00AI-driven storage access
- 1:49Prefetching speeds storage
- 2:33Traditional prefetching limits
- 4:05Multi-tenant workload challenges
- 6:22Access patterns to graphs
- 8:35GNNs for global patterns
- 9:23Graphs into ML models
- 10:06GNN performance gains
- 11:27Key takeaways
related talks
17mBGL: GPU-Efficient GNN Training by Optimizing Graph Data I/O and Preprocessing
16mPreview: Graph NNs
19mOffloading-Efficient Sparse AI Systems
16mCognitive SSD: A Deep Learning Engine for In-Storage Data Retrieval
11mLitePred: Transferable and Scalable Latency Prediction for Hardware-Aware Neural...
14mGLIST: Towards In-Storage Graph Learning
09mBuild a UI that Learns - Intelligent Prefetching
14mHyCache: Hybrid Caching for Accelerating DNN Input Preprocessing Pipelines