Per OSD Recovery Bandwidth Control Based on dmClock
summary
Prototype implements dmClock to enforce per-OSD bandwidth limits for recovery, exposing three fixed levels (low/medium/high) plus a self-adaptive mode that adjusts per-OSD targets by degraded-object counts (example: OST0 330 vs OST1 200 → factors 1.22 and 0.8). Tests on a ~122‑PG cluster keep client throughput around 10,000 IOPS while showing recovery-priority mode finishes full recovery in about 6 minutes versus over 30 minutes with strict low-limit mode, with the safety/mixed mode trading intermediate latency and recovery time.
AI-written summary · about
chapters (5)
- 0:00Start
- 0:49Background balance client ops with recovery ops
- 4:28Design & Implementation High Lights
- 11:43OSD recovery bandwidth
- 16:00cluster bandwidth control
related talks
44mCeph QoS Refinements for Background Operations Using MClock
18mDayu: Fast and Low-interference Data Recovery in Very-large Storage Systems
28mRock around the clock (synchronization
38mEfficient Ceph Performance Troubleshooting in Production Using eBPF
24mAverting crisis at multi-petabyte scale
25mExplicitly Supporting Stretch Clusters in Ceph
14mUnderstanding RDMA Microarchitecture Resources for Performance Isolation
19mProtocol-Aware Recovery for Consensus-Based Storage