uclamp in CFS: Fairness, latency, and energy efficiency
summary
Current upstream uclamp uses max aggregation, which lets many tasks share a low frequency and hides true compute demand, causes PELT saturation and frequency spikes, and relies on coarse buckets. Proposal applies clamps at the source by adding a per-task util_avg_uclamp that is propagated to the runqueue and uses sum-style aggregation (or a new pelt-derived signal) so load balancing can spill tasks and energy models remain accurate. Prototype claims ~50% less uclamp code and shows preliminary functional improvements.
AI-written summary · about
chapters (5)
- 0:00Introduction to uclamp
- 1:47Current uclamp aggregation issues
- 7:31Problems with max aggregation
- 11:22Solutions and upstream RFC
- 14:50Summary and next steps
related talks
40mLatency hints for CFS task
48mAn Unbiased Look at the Energy Aware Scheduler (EAS)
28mLatency hints for CFS task
13mHarvesting Spare CPU Resources in Container Systems
50mSCHED_DEADLINE: It's Alive!
32mCPU Power Saving Methods for Real-time Workloads
21mLinux per cpu idle injection
32mImprove CPU Utilization by Designing a New Schedule Algorithm