Measuring the Impacts of the Preempt-RT Patch
summary
Shows measured trade-offs of the Preempt-RT patch: it cuts worst-case latencies toward a 200 µs target but increases kernel overhead and context switches (observed ~700k cs/s under stress). Uses perf, pidstat, vmstat, mpstat plus stress-ng and cyclictest to compare raw compute, network and I/O. Threaded IRQs appear as tasks in htop, FIFO-heavy workloads jump from ~8s to ~70s, while iperf bandwidth stays similar. Recommends tuning CPU idle states, disabling CPU frequency scaling and SMT and checking firmware SMM and shared SIMD units.
AI-written summary · about
chapters (25)
- 0:00Introduction
- 0:46Simulation Platform
- 1:27Test Bench
- 2:17Embedded Product
- 4:08RealTime Operating System
- 5:38Linux
- 8:07RT Path
- 10:02Perf
- 11:06Tati
- 11:38NonRT tasks
- 14:35PIDstat
- 15:39VMstat
- 18:05Context switches
- 18:23PIDstats
- 18:42IPerf
- 19:23StressEnergy
- 20:50CPU
- 21:41Recap
- 22:07Building a realtime operating system
- 22:50CPU idle state
- 23:52Devfs
- 24:38Know your system
- 27:49CPU idle
- 29:36More information
- 30:14Questions
related talks
42mWhen the OS gets in the way
09mInterrupts: The Hidden World of Linux Performance
29mImprove Xeon IRQ throughput with posted interrupt
40mPerformance Testing Done Right
37mAdvanced Performance Profiling Using Perf Tools
49mResponding in a Timely Manner
01h 06mPerformance engineering: being friendly to your hardware
55mPerformance tuning of Red Hat Enterprise Linux