learnedftl · research

LearnedFTL

LearnedFTL is a research page-level flash translation layer that supplements a mapping cache with bitmap-qualified address prediction.

Related Observatory article: Read on hesela.com

After a cached-mapping miss, a per-logical-page bitmap selects either a model-derived address or the translation-page fallback. A set bit marks an exact prediction for that mapping, not permission to read approximate data. The paper clears the bit on an overwrite. [1]

Virtual physical page numbers make parallel placement easier to model. Group-based allocation brings related mappings together for training during garbage collection. [1]

In the pinned author artifact, model_predict compares its prediction with the emulator's mapping table, while the bitmap-clearing block in ssd_write is commented out. In ssd_read, model_hit_num increases before model eligibility and prediction success are checked; a failed prediction can still reach the translation-page fallback. That counter therefore does not, by itself, count successful avoided translation reads. This is a static observation of the pinned source, not a demonstrated data-loss event or a reproduced performance result. [2]

LeaFTL is a different design. Its bounded-error segments use out-of-band metadata to correct mispredictions. Similar names do not make the algorithms or evaluation artifacts interchangeable. [4]

The HPCA evaluation uses FEMU and reimplemented comparison algorithms. Retention-rewrite training is proposed but not evaluated there: the emulator lacks that rewrite mechanism, so the evaluated model training uses garbage collection. [1]

Relations

related
SSD mapping cache, LeaFTL, DFTL, Global Translation Directory (DFTL), FTL translation page, Flash translation layer (FTL), Garbage collection (flash)

Advantages

  • A qualified prediction can avoid a mapping-page read even when the mapping cache misses. [1]

Limitations

  • A historical accuracy bit must remain consistent with the current mapping; prediction coverage is not a crash-recovery guarantee. [1, 2]
  • The artifact's setup requires a virtual machine and substantial host memory. [3]

Citations

  1. Wang et al. LearnedFTL: A Learning-Based Page-Level FTL for Reducing Double Reads in Flash-Based SSDs. HPCA 2024, sections III-IV. (accessed 2026-11-22)
  2. LearnedFTL author artifact, commit 4702166cdc8daf32008a65a79f72ad991c2a991d: model_predict, ssd_read and ssd_write. (accessed 2026-11-22)
  3. LearnedFTL author artifact: build, virtual-machine setup and FTL selection at the inspected commit. (accessed 2026-11-22)
  4. Sun et al. LeaFTL: A Learning-Based Flash Translation Layer for Solid-State Drives. ASPLOS 2023, section 3. (accessed 2026-11-22)

Further reading

  1. Hesela: SSD mapping caches, learned lookup validation and evaluation boundaries. (accessed 2026-11-22)
Sources & caveats
  • Primary-paper synthesis and a static review of the pinned author implementation. No emulator run, hardware test, benchmark replication or independent review is claimed.
  • The counter interpretation follows source control flow; it does not establish the numerical impact on published graphs. Paper design and inspected artifact are distinguished.

JSON: /data/terms/learnedftl.json

Published by Hesela team · Updated

Report a correction · Editorial standards · Reuse guidance

Cite this page
Hesela team. LearnedFTL. Hesela. 2026-11-22. https://hesela.dev/terms/learnedftl/