How to get realistic c-state latency and residency?
summary
Uses the hikey960 and a simple LT workload plus an arm energy probe to compare CPU-idle, cluster-off and WFI power without specialized gear. Measures exit latency by firing thousands of simultaneous absolute-timer wakeups (big-cluster worst-case ~3 ms, little ~1.6 ms), measures entry by increasing sleep lengths to find the breakpoint (CPU ~0.5 ms, cluster ~1 ms), and derives residency by plotting power vs sleep period to find a ~5 ms break-even and strong dependence on OPP. Scripts and automation are provided.
AI-written summary · about
chapters (10)
- 0:00Introduction to c-state latency
- 1:54Why measure c-state latency
- 4:42Tools for measuring latency
- 7:35Measuring exit latency
- 16:53Entry latency measurement
- 21:10Residency time overview
- 23:24Calculating residency time
- 26:18Power consumption tests
- 28:29OPP impact on residency
- 29:16Conclusion and next steps
related talks
56mPMWG: CPU Idle governor and irq prediction
32mCPU Power Saving Methods for Real-time Workloads
52mEffectively Measure and Reduce Kernel Latencies for Real-time Constraints
42mWhen the OS gets in the way
48mCyclictest latency, find out why with tracing
49mDebugging Methodologies for Realtime Issues in Linux Systems
40mLatency hints for CFS task
28mLatency hints for CFS task