Ceph Performance Tuning: From Bluestore to RBD

on YouTube ↗38m

" summary

Uses CBT Upstream and fio/hsbench for repeatable tests on a 60‑OSD NVMe cluster and shows concrete wins: reads hit 4.5 million IOPS while burning ~8 cores per OSD; raising PG count 4096→16000 raises reads from ~3.1M to ~4.5M. RBD snap‑trim fixes cut OSD CPU from ~250% down to ~1 core; RocksDB TTL compaction drops insertion latency from ~60,000 ms to ~10,000 ms and disk use from ~90–95% to ~20–25%.

AI-written summary · about

" chapters (14)

1. 0:00Introduction and setup
2. 1:06Is Ceph slow?
3. 2:55Benchmarking tools overview
4. 5:00When Ceph performs best
5. 6:45SSD firmware impact on performance
6. 8:12PG count and performance scaling
7. 11:44MDS directory distribution
8. 13:09IO500 benchmark insights
9. 15:57RBD snap trimming CPU usage
10. 19:30OSD threading efficiency
11. 21:19PG log rocksdb interaction
12. 24:31RocksDB tombstone compaction issue
13. 30:50Summary and key takeaways
14. 31:53Q&A session