Effectively Measure and Reduce Kernel Latencies for Real-time Constraints
Talkat Embedded Linux Conference + OpenIoT Summit North America 2017 · on Tuesday, 2017-02-21· at 11:30· Atrium Ballroom·
summary
Shows how to measure and cut Linux kernel response down to microseconds using PREEMPT_RT, CPU isolation, cache lockdown, mlock/hugepages and IRQ affinity, and by replacing spinlocks with sleeping/low-spin locks. Adds scheduler profiling hooks and a microsecond-scale visualization of per-core run-queue events (using cyclictest, hackbench, perf and in-house Prieta tasks). Reports scheduler durations usually <10 µs, with bursts up to max schedule duration 34µs under heavy/short-burst loads, and clustered wakeups 10–20 µs apart that amplify jitter.
AI-written summary · about
chapters (4)
- 0:00Start
- 7:26How To Measure the Latency
- 10:16Efficient Way To Perform Microscope Measurement
- 29:43Experiment Configuration
related talks
26mReal-time Linux: what is next?
36mJourney of improving RT performance on Single CPU equipped PREEMPT_RT Linux
32mCPU Power Saving Methods for Real-time Workloads
50mInside the RT Patch- 36mJourney of improving RT performance on Single CPU equipped PREEMPT_RT Linux
51mTuning RT Kernel to Improve Schedule Performance with Intel Platform
49mDebugging Methodologies for Realtime Issues in Linux Systems
39mEvaluation of PREEMPT_RT in Virtualized Environments