Speed Up Creation of a VM With Pass Through GPU
summary
Claims specific kernel and QEMU patches that cut GPU pass-through VM creation from ~38s to 3.3s for 1 GPU 48GB and from ~120s to 4.4s for 4 GPUs 192GB (>95% reduction). Uses perf/flamegraphs to find hotspots and implements pre-zero free page, get_user_pages_contig/get_user_page_long_term, cached IOA mappings and IOA splitting, parallel PCI resets, and reverse-map skipping by counting. Patches land as RFCs and local benchmarks show large, reproducible speedups with modest compatibility caveats.
AI-written summary · about
chapters (14)
- 0:00Intro
- 1:02AGENDA
- 1:37BACK GROUND
- 2:40BASE LINE DATA
- 3:50WHAT SLOW DOWN CREATION OF A GPUVM INSTANCE
- 5:22SPEED UP VFIO_PIN_PAGES_REMOTE
- 6:45PRE ZERO OUT FREE PAGES
- 7:50PIN MEMORY IN BULK
- 8:55MAKEVFIO DMA MAP MORE EFFICIENT
- 10:59PCI RESET OPTIMIZATION
- 12:19KVM META DATA OPTIMIZATION
- 14:33CONFIGURATION OPTIMIZATION
- 15:29EFFECT OF OPTIMIZATION
- 17:28CONCLUSION
related talks
20mPassthrough/Headless GPU Gets Ahead
33mGPU Para-Virtualization on Xen
26mAnalysis of AMD HW-assisted vIOMMU Implementation and Performance
40mAchieving the Ultimate Performance with KVM
21mNVMe Emulation Performance Optimization
44mImproving kexec boot time
30mBeyond the DC Walls: Building a Nationwide GPU Fabric With KubeVirt and Wide-Area…
25mExploring Intel Sub-Page Protection in Context of Live Migration