CPUfreq/sched and VM guest workload problems

on YouTube ↗15m

" summary

Workloads in VMs show large CPUfreq/sched regressions versus host; a virtual MMIO CPU frequency control device plus DT property changes are implemented so guest reads trap to host and writes clamp VCPU frequency, and an optional BPF handling path avoids costly userland hops. Synthetic traces show host Fmin→Fmax 180 ms, VM without fixes never migrates, VM with fixes 183 ms; Fmin→Fmax per-core 47 ms host vs 34 ms without fix vs 46 ms with fix; big-CPU ramp 133 ms host vs 188 ms without vs 135 ms with. Benchmarks (PCMark, Roblox, ML tests) regain most performance, and an additive Uclamp change fixes slow ramps when multiple threads migrate onto one CPU.

AI-written summary · about

" chapters (10)

1. 0:00Introduction to VM workloads
2. 1:05Fixing VM frequency awareness
3. 2:09Synthetic workload performance
4. 3:51Uclamp aggregation issues
5. 5:36Guest-host communication methods
6. 6:17BPF solution for MMIO
7. 7:19Benchmark results overview
8. 8:23Real-world workload benchmarks
9. 9:03Discussion on additive Uclamp
10. 10:50Q&A session begins