PMWG: CPU Idle governor and irq prediction
summary
Implementation of an idle governor that predicts the next wake source and subtracts wake-up overhead shows that exit latencies are highly variable up to 2–3 ms (and shift by ~60% with OPP), so naive residency measurements overestimate sleep. A prototype on HiKey (Hikey 8960) compares a shallow-WFI policy and a prediction-based governor, finds about 8% energy saving on multimedia but firmware can mask savings, and measures prediction overhead around 14 μs. Proposed fixes: per-cluster dynamic exit latencies, refcounted idle domains, or firmware returning actual sleep time.
AI-written summary · about
chapters (4)
- 0:00Start
- 1:44Idle states characteristics
- 22:25Consumption vs idle states
- 47:42What is going on? (Jankbench test1)
related talks
30mHow to get realistic c-state latency and residency?
24mLimit the idle CPU search depth and use CPU filter during task wake up
21mLinux per cpu idle injection
32mCPU Power Saving Methods for Real-time Workloads
25mMultiple system-wide low power-states
49mDebugging Methodologies for Realtime Issues in Linux Systems
52mEffectively Measure and Reduce Kernel Latencies for Real-time Constraints
47mEvolution of Suspend-to-Idle Support in The Linux Kernel