hesela.devBaleen
Baleen is a research flash-cache design that coordinates learned admission and prefetching to reduce HDD backend work under a flash-write budget.
Training groups accesses into residency episodes using an assumed eviction age, then learns admission and prefetching decisions. This offline construction is not future knowledge available to a running cache. [1]
The evaluation uses seven Meta traces and a chronological split. The first day trains models, later days test them. Simulated results and testbed validation are distinct evidence. [1]
The author artifact documents a RejectX baseline, model training and simulation. [2]
Relations
Citations
- Wong et al. Baleen: ML Admission & Prefetching for Flash Caches. FAST 2024, pp. 347-371; sections 3.1, 3.4 and 5.
- Baleen-FAST24 author artifact, pinned README at 4e3a9205b1f1bd72bffd66dd52bd5b06509089e9.
Sources & caveats
- Primary-source synthesis; the author simulator, ML training and testbed measurements were not reproduced.