Inking Out Inefficiencies in Ceph Erasure Coding
summary
The Tentacle release ships a new Erasure Coding implementation called fast EC that boosts small‑read IOPS roughly threefold versus legacy EC and reduces space waste so very small objects match replica efficiency; overwrite path adds parity‑delta writes cutting an 8+2 read/modify/write from 10 reads+8 writes to about 3 reads+3 writes. Direct reads (client fetches from OSDs) will land in Umbrella to reach replica‑par read latency, defaulting to a 16k default chunk size. On‑disk format changes require an OSD/mon upgrade and a pool conversion flag but clients need not all be upgraded.
AI-written summary · about
chapters (20)
- 0:00Introduction to Fast EC
- 0:44What is Erasure Coding
- 3:18Performance of Replica vs EC
- 5:05Improving Read Performance
- 7:57Small Object Efficiency
- 9:45Cost Savings vs Replica
- 10:25Direct Reads Future
- 11:57Overwrite Improvements
- 13:55Parity Delta Writes
- 16:04Write Performance Gains
- 17:29Scaling with K Value
- 18:52Chunk Size Optimization
- 20:38Storage Efficiency Impact
- 23:17Recommendations for Chunk Size
- 24:00Backfill Recovery Performance
- 25:02Future Plans Overview
- 26:25Pool Migration Benefits
- 27:47Simplifying EC Pools
- 28:50Stretch Clusters Support
- 29:33Q&A Session