Advancing BlueStore with Real-World Insights
summary
A Python scraper tool captures a 73-hour RBD trace (4 OSDs, ~300M ops) and reveals concrete opportunities: coalescing disjoint writes would touch ~4.6M extra segments but only nets ~12% benefit so likely not worth added complexity; simple prefetching of sequential reads cuts read IO by ~63% for sequences and ~39% overall in the model; ALRU cache is within ~10% of an oracle but biased by PG placement; a replayer and zero pool exist and work but need better reporting, compression handling, and mixing of workloads.
AI-written summary · about
chapters (10)
- 0:00Introduction and motivation
- 1:05Limitations of synthetic workloads
- 3:01Scraper tool for real workloads
- 4:27Analyzing read-write alignment patterns
- 9:54Opportunities in RBD write-read correlation
- 11:03In-sequence read prefetching potential
- 16:56Optimizing onode cache misses
- 22:23Workload replay and testing tools
- 23:48Efficient testing improvements
- 24:54Q&A