How to split a cake? GPUs with OpenStack
summary
Shows a working OpenStack setup that slices heterogeneous GPUs into multiple instance types: an Intel iGPU and an NVIDIA T4 are configured with kernel modules and Nova to expose mdev profiles / mediated devices as flavors, producing multiple performance tiers rather than a single passthrough. The demo files map profiles to flavors, exposes driver and kernel limits (documentation notes vendor caps and monitoring gaps), and reports practical limits (driver/host constraints, up to ~40 lanes reported) — it works with caveats.
AI-written summary · about
chapters (9)
- 0:00Introduction to GPUs in OpenStack
- 2:13Hardware virtualization challenges
- 5:12Nova and virtual GPU profiles
- 8:10Kernel modules and performance tuning
- 11:59Nova configuration and flavors
- 14:21Creating virtual GPUs with Nova
- 17:48Custom GPU flavors and limitations
- 20:28Instance management and monitoring
- 22:44Future features and Q&A