Speed Up Creation of VM with Pass-Through GPU

Discover techniques to accelerate VM creation with GPU pass-through in this 20-minute conference talk by Liang Li from DIDI Chuxing. Explore the challenges slowing down GPU VM instance creation and learn about various optimization strategies. Delve into topics such as speeding up VFIO_PIN_PAGES_REMOTE, pre-zeroing free pages, pinning memory in bulk, enhancing VFIO DMA mapping efficiency, PCI reset optimization, KVM metadata optimization, and configuration optimization. Gain insights into the effects of these optimizations and their practical implications for improving VM performance with GPU pass-through.

Intro
AGENDA
BACK GROUND
BASE LINE DATA
WHAT SLOW DOWN CREATION OF A GPUVM INSTANCE
SPEED UP VFIO_PIN_PAGES_REMOTE
PRE ZERO OUT FREE PAGES
PIN MEMORY IN BULK
MAKEVFIO DMA MAP MORE EFFICIENT
PCI RESET OPTIMIZATION
KVM META DATA OPTIMIZATION
CONFIGURATION OPTIMIZATION
EFFECT OF OPTIMIZATION
CONCLUSION