Low-Latency Real-Time (RT) Kernel: The Cornerstone of Arm's 5G ORAN Strategy
summary
Arm measures RT latency on server cores with cyclictest and finds worst-case spikes over 100 µs on 64 cores, while the target is under 10 µs. Tracing with trace-cmd pins the problem to getrandom/CRNG contention (CRNG mix_state) triggered by kernel stack offset randomization exhausting a shared entropy pool; applying RT tuning (isolating IRQs/CPUs, performance governor, disabling ticker) plus a CRNG patch reduces max latency to about 8–9 µs even at 64 cores.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 2:11Real-time latency in 5G
- 7:30Latency requirements and OSI model
- 11:09Benchmark setup and methodology
- 12:54Initial latency data
- 16:11Bug reports and fixes
- 19:21Kernel tuning for real-time
- 24:08Latency results with tuning
- 25:35Kernel stack randomization impact
- 26:40Root cause analysis with Trace-CMD
- 37:11Debugging high latency with function tracing
- 39:43Entropy exhaustion in RNG
- 41:10Final results and patches
- 41:33Q&A
related talks
50mLow-Latency Real-Time (RT) Kernel: The Cornerstone of Arm's 5G ORAN Strategy
26mTuning Real-Time Linux for ARM SoCs
33mTowards PREEMPT_RT for the Full Task Isolation
47mWhat Every Driver Developer Should Know about RT
43mRTLA: Real-time Linux Analysis Toolset
35mPreempt-RT Raspberry Pi Linux
40mTips, Tricks, and Gotchas for Linux Real-Time Tuning
36mPreempt-RT Latency Benchmarking of the Cortex-A53 Processor