Baleen: ML Admission & Prefetching for Flash Caches
summary
Baleen uses coordinated ML admission and prefetching, training on an episodes model that groups accesses from admission to eviction and optimizes an end-to-end metric called Disk-head Time. For a fixed flash write-rate budget it cuts peak Disk-head Time by 12% versus prior policies; Baleen‑TCO picks the optimal flash write rate and reduces estimated TCO by 17% on seven Meta traces. Prefetching supplies most gains but requires a separate confidence model; code and traces are available for offline simulation and deployment.
AI-written summary · about
chapters (11)
- 0:00Baleen and flash caching
- 2:19Episodes model for ML training
- 3:03Caching in block storage
- 6:10Training on episodes vs accesses
- 9:01Efficient episode generation
- 12:28Episode-based admission policies
- 13:52Write rates vs disk costs
- 15:40Baleen vs baseline policies
- 17:54Baleen’s cost-saving benefits
- 18:37Effective prefetching insights
- 19:43Conclusion and future work
related talks
20mTectonic-Shift: A Composite Storage Fabric for Large-Scale ML Training
20mAccelerating Model Loading in LLM Inference by Programmable Page Cache
24mBLAS-on-flash: An Efficient Alternative for Large Scale ML Training and Inference?
15m3L-Cache: Low Overhead and Precise Learning-based Eviction Policy for Caches
19mOffloading-Efficient Sparse AI Systems
20mLearning Relaxed Belady for Content Distribution Network Caching
14mHyCache: Hybrid Caching for Accelerating DNN Input Preprocessing Pipelines
15mBidaw: Enhancing Key-Value Caching for Interactive LLM Serving via Bidirectional...