Power Measurement and Attribution for Processes and Hardware Devices in the Linux...
summary
A kernel-side framework measures per-process and per-device energy by calibrating device power curves (baseline plus operating-range sweeps) and running multivariate regression over process counters pulled from proc process accounting to produce kJ estimates; attribution multiplies a process’s active fraction (example: 1.5s active of 10s => 15%) into the model. Validation uses hardware probes and is validated with multimeter ground truth; GPU reads use Nvidia SMI. Work surfaces accuracy versus instrumentation overhead, RAPL/rle timing limits, privacy concerns, and a planned GNOME-Usage panel and developer API.
AI-written summary · about
chapters (16)
- 0:00Power measurement in Linux
- 1:03Energy efficiency in computing
- 2:49Measuring app energy use
- 4:12Limitations of power measurement
- 7:22Accurate energy measurement frameworks
- 9:08Power models for hardware
- 10:37Why existing tools like PowerTop fall short
- 13:09Regression-based energy system
- 16:42Process energy attribution
- 19:15Energy modeling trade-offs
- 21:47Scaling energy measurement
- 23:57Privacy in energy tracking
- 24:40Energy to carbon emissions
- 26:49Energy measurement utility demo
- 28:17Key takeaways and future
- 29:21Q&A on energy measurement
related talks
42mPower Measurement and Attribution in Kernel for Embedded System Processes
27mMeasuring Carbon Footprint of Personal Computing
39mAbout The Need to Power Instrument The Linux Kernel
24mEnergy model accuracy
22mA generic energy model description
45mLinux perf tool metrics
51mLinux Systems Performance - Architecture
25mThe Power-Supply Subsystem