Energy model accuracy
summary
Linux EAS uses static per-CPU tables mapping frequency to capacity and power, but real-device sweeps show up to 20% variation across workloads and OPPs; medium and big cores often differ far less than expected so placement decisions break. Practical fixes explored: runtime model learning with an energy meter (hardware often lacks low-power accuracy), per-thread runtime updates, and using PMU counters or IPC estimates to classify memory-bound vs compute-bound behavior. Watch if wanting practical tradeoffs and implementation ideas.
AI-written summary · about
chapters (8)
- 0:00Introduction to energy model
- 1:27Current energy model limitations
- 3:35Accuracy problems in predictions
- 7:28Factors affecting performance
- 10:26Solutions for better accuracy
- 17:55Task-specific adjustments
- 22:17Detecting model inaccuracies
- 23:40Conclusion and next steps
related talks
37mPower Measurement and Attribution for Processes and Hardware Devices in the Linux...
42mPower Measurement and Attribution in Kernel for Embedded System Processes
20mUpdating Energy Model from Thermal
48mAn Unbiased Look at the Energy Aware Scheduler (EAS)
27mMeasuring Carbon Footprint of Personal Computing
32mCPU Power Saving Methods for Real-time Workloads
01h 11mPerformance engineering - being friendly to your hardware
44mReal-Time Linux on Embedded Multicore Processors