The Two New I/O Controllers and BFQ
summary
Compares the new io.cost and io.latency cgroup controllers against BFQ on SSDs and finds the controllers often fail: a target group gets up to 47× more throughput under BFQ and total throughput is up to 3× higher with BFQ. Root cause is that io.cost uses a simple linear per-I/O cost model and a feedback V_rate throttle that misestimates service on drives with complex, time‑varying transfer functions, causing unnecessary throttling, poor latency control and slow recovery.
AI-written summary · about
chapters (4)
related talks
25mRecent developments in I/O control and testing
49mBFQ I/O scheduler: more throughput, control and efficiency
41mA Solution to High Latencies Caused by I/O
18mIO and cgroups, the Current and Future Work
32mI'm BFQ, I'm dynamite (or almost ...)
45mLinux Storage Stack Explained
50mOptimizing for NVMe Drives: The 10 Microsecond Challenge
24mOptimizing Storage Performance for 4–5 Million IOPs