Recent developments in I/O control and testing
summary
Compares BFQ against Facebook’s I/O cost controller and I/O latency controller and shows the new controllers routinely misallocate bandwidth: interferer groups get near zero share, targets or total throughput collapse (observed up to 47× lower per-group bandwidth and up to 3× lower aggregate throughput on SSDs). Root cause is controllers relying on noisy, time‑varying I/O cost and device‑saturation estimates (v_rate, busy level, lagging groups), causing the controller to throttle the device unnecessarily; tracing patches reproduce and explain the failures. Testing work is integrated into LKFT.
AI-written summary · about
chapters (8)
- 0:00Introduction and overview
- 1:04State of I/O controllers in Linux
- 3:18New I/O latency and cost controllers
- 8:15Root cause of controller failures
- 13:38Failure analysis of I/O cost controller
- 17:58Real-world analogy for controller issues
- 20:52Why BFQ outperforms new controllers
- 21:35Transition to I/O testing presentation
related talks
27mThe Two New I/O Controllers and BFQ
18mIO and cgroups, the Current and Future Work
41mA Solution to High Latencies Caused by I/O
49mBFQ I/O scheduler: more throughput, control and efficiency
42mcgroup2 Resource Control strategies w/ resctl-demo
32mI'm BFQ, I'm dynamite (or almost ...)
38mUh-oh, It's I/O Ordering!
24mOptimizing Storage Performance for 4–5 Million IOPs