Closing the B+-tree vs. LSM-tree Write Amplification Gap on Modern Storage Hardware...
summary
Shows three concrete changes — deterministic page shadowing, localized modification logging, and sparse redo logging — applied to a B+-tree implementation and evaluated on a ScaleFlux 3.2 TB computational storage drive with built‑in transparent compression. Uses a 150 GB dataset, 1 GB cache, record sizes 16/32/128 B and page sizes 8/16 KB. The modified B+-tree reduces write amplification by over 10×, matches or beats RocksDB’s (LSM-tree) write amplification, and improves random write throughput while keeping scan performance comparable.
AI-written summary · about
chapters (10)
- 0:00Intro
- 1:03Storage level Transparent Compression
- 3:07Reduce B+-tree Write Amplification
- 3:37Deterministic Page Shadowing
- 4:29Localized Page Modification Logging
- 6:06Spares Redo Logging
- 7:50Log-Flush-Per-Minute
- 10:23Log-Flush-Per-Commit
- 11:17Speed Performance Evaluation
- 12:25Conclusion
related talks
01h 16mDeep Dive into On-Disk Data Structures
39mAn Introduction to B-trees and LSM-trees for DEVs and DevOps
45mThe Database: A Sophisticated and Special Data Structure
34mWhat We Talk About When We Talk About Disk IO
01h 16mOn-Disk Data Structures: B-Trees, LSM-Trees and Fractal Trees
24mPerformance tuning in BlueStore & RocksDB
18mPreview: KV-Stores and Databases
21mMatrixKV: Reducing Write Stalls and Write Amplification in LSM-tree Based KV Stores