RL-Watchdog: A Fast and Predictable SSD Liveness Watchdog on Storage Systems
summary
A kernel-level watchdog implements a LWW liveness probe, a reinforcement-learning timeout predictor (RLTP) and a fast-failure-notify path (FFN) inside Linux 6.0.0. RLTP learns three kernel metrics, uses ~384 bytes per SSD, stabilizes within 120 seconds and predicts LWLC latency with up to 99.8% accuracy. Evaluated with power-fault injection on single SSDs and RAID, it cuts application data loss up to 96.7% (82.4% on buffered writes) and shortens detection time up to 97.9% without raising IO tail latency.
AI-written summary · about
chapters (10)
- 0:00RL-Watchdog introduction
- 1:20SSD failure detection importance
- 3:05Timeout-based detection problems
- 7:21RL-Watchdog core schemes
- 11:41Evaluation setup and failure injection
- 13:09Impact on data loss and detection time
- 14:55Real-world performance impact
- 15:39RLTP prediction accuracy
- 16:23False positives impact on I/O
- 16:46Conclusion
related talks
19mEvaluating File System Reliability on Solid State Drives
13mRemap-SSD: Safely and Efficiently Exploiting SSD Address Remapping to Eliminate...
12mFusionRAID: Achieving Consistent Low Latency for Commodity SSD Arrays
45mBig Data and SSDs
20mNVMe SSD Failures in the Field: the Fail-Stop and the Fail-Slow
18mOPTR: Order-Preserving Translation and Recovery Design for SSDs with a Standard Block Device Interface
53mOvercome Blast Radius Anxiety With Ultrafast Rebuilds and Performance for High-capacity SSDs
15mRemoving Double-Logging with Passive Data Persistence in LSM-tree based Relational...