Running Ceph on Flashcache

on YouTube ↗18m

" summary

OVH runs Ceph on flashcache across ~150 clusters on roughly 11,500 servers with ~44 PB storage, ~2 PB NVMe and over 70,000 flashcache instances. Real problems and fixes: cache pollution from deep-scrub/backups is mitigated with PID blacklisting and a sequential skip threshold in flashcache; client latency drops after fixing an xattr / metadata block-size mismatch and documenting the option; an XFS deadlock under low memory is identified and fixed. Measured result: ~50% more IO and ~40% write hit on flash.

AI-written summary · about

" chapters (7)

1. 0:00Introduction to talk
2. 1:32Why choose Flashcache
3. 2:42Cache pollution problem
4. 6:58Matching block sizes
5. 9:34Deadlock on low memory
6. 12:07Flashcache performance gains
7. 15:36Future of Flashcache