StripeFinder: Erasure Coding of Small Objects Over Key-Value Storage Devices
Talkat 12th USENIX Workshop on Hot Topics in Storage and File Systems (HotStorage '20) · 2020-07-13..07-14
summary
StripeFinder exposes that KVMD stores DxP metadata objects and quantifies the cost: for P=2 and 16-byte keys, KVMD gives byte amplification 2.8x (D=4) and 3.3x (D=8) for 256-byte objects, nullifying parity gains; StripeFinder cuts metadata bytes by ~D and yields 2.2x (D=4) and 1.9x (D=8). Object-count amplification drops from 4.5 to 1.8 when 10 finder entries are combined. Conclusion: parity helps for mid-sized objects but tiny objects remain best replicated; implementation complexity and key-size ratio limit applicability.
AI-written summary · about
chapters (12)
- 0:00Intro
- 0:21Context
- 1:32Erasure Coding Goals
- 2:35For Large Objects: Splitting
- 3:55For Small Objects: Multi-Packing?
- 4:23For Small Objects: Uni-Packing
- 6:30The Metadata Overhead of Uni-Packing
- 8:19Stripe Finder: Reducing Byte Amplification
- 8:31StripeFinder: Reducing Byte Amplification
- 11:38StripeFinder: Reducing Object Amplification
- 12:53Conclusions
- 14:44Future Directions
