BFQ I/O scheduler: more throughput, control and efficiency
summary
BFQ claims up to five‑times higher throughput in hard server scenarios and reports a ~150% throughput uplift and ~80% faster app start in recent changes. It adds service‑time based injection to fill device service holes without breaking bandwidth guarantees, replaces bandwidth probing, and recovers throughput on deceptive workloads. A Chromebook demo shows Chromebook demo of immediate Chrome start during a heavy download. Also trims overhead (amortized ~0.6 µs per scheduler event) by disabling useless merges on SSDs.
AI-written summary · about
chapters (19)
- 0:00Introduction and speaker
- 2:21Previous BFQ achievements
- 3:29Live Chromebook demo
- 8:13New improvements overview
- 10:00Throughput enhancement details
- 15:49Loss of control problem
- 18:13Injection mechanism explained
- 21:28Throughput results after changes
- 22:37Responsiveness timeout fix
- 24:45Reducing execution time
- 27:01Hardware firmware impact
- 30:49Drive latency vs throughput
- 32:46Vendor firmware insights
- 34:31BFQ algorithm and CPU usage
- 35:47Parallelism challenges
- 39:34Multi-queue CPU efforts
- 44:01Global control vs locality
- 45:48Fairness and interference
- 49:02Q&A session
related talks
41mA Solution to High Latencies Caused by I/O
32mI'm BFQ, I'm dynamite (or almost ...)
27mThe Two New I/O Controllers and BFQ
25mRecent developments in I/O control and testing
38mFixing Real-Time Scheduler Throttling in the Linux Kernel
16mBPF Based Telemetry, Metrics, and Accounting on Android
17mSBB: Eliminating Centralized Bottlenecks in Userspace Network Runtime
25mLinux needs a Scheduler QOS API -- and it isn't nice(2)