Evaluation of PREEMPT_RT in Virtualized Environments
summary
Measurements use cyclictest on Docker, KVM and Jailhouse with a tiny injector driver called bloxys to emulate misbehaving drivers; containers hit host-like worst‑case latencies (~26–30 µs) once C‑states and privileges are fixed, KVM guests show higher and irregular worst cases (50 µs on one vCPU, ~200 µs on another), Jailhouse guests match native host latencies (~45 µs), and stressing shared L2 cache multiplies latencies (~6×), so CPU pinning, NO_HZ_FULL and IRQ affinity are required.
AI-written summary · about
chapters (6)
- 0:00Introduction and approach
- 3:08Containerization evaluation
- 16:01Hardware virtualization with KVM
- 24:06Jailhouse partitioning
- 28:43Hardware resource impact
- 33:58Conclusion and Q&A
related talks
36mJourney of improving RT performance on Single CPU equipped PREEMPT_RT Linux- 36mJourney of improving RT performance on Single CPU equipped PREEMPT_RT Linux
52mEffectively Measure and Reduce Kernel Latencies for Real-time Constraints
26mReal-time Linux: what is next?
54mIntroduction to Realtime Linux
53mPREEMPT_RT - How Not to Break It.
44mReal-Time Linux on Embedded Multicore Processors
24mPREEMPT_RT Q&A with tglx