KubeVirt Deep Dive: Virtualized GPU Workloads on KubeVirt
Talkat KubeCon + CloudNativeCon North America 2019 · on Wednesday, 2019-11-20· at 10:55· Room 1AB - San Diego Convention Center, Upper Level·
summary
KubeVirt runs full VMs as pods using virt-controller, virt-launcher and virt-handler, mapping VM CPU/memory to pod requests and PVCs for VM disks. NVIDIA supplies an open source KubeVirt GPU device plugin that discovers GPUs and vGPUs, registers via gRPC, and on allocation returns a PCI address set as a pod env var which virt-launcher injects into libvirt/domXML for passthrough. The flow yields VMs with GPUs (including Windows gaming) running on Kubernetes; PCI/vNUMA scheduling work is ongoing.
AI-written summary · about
chapters (11)
- 0:00Introduction to KubeVirt
- 1:31Use cases and philosophy
- 4:23Managing VMs like pods
- 5:37Under the hood architecture
- 9:21Storage and network integration
- 12:53GPU support motivation
- 14:06GPU device plugin architecture
- 16:53GPU device discovery flow
- 18:59GPU allocation and VM launch
- 21:25KubeVirt feature utilization
- 24:38Questions and answers
related talks
31mKubeVirt: What If Kubernetes Orchestrated Your VMs?
43mKubeVirt - Kubernetes, Virtualization and Your Future Data Center
32mKubeVirt - making running virtual machines in a k8s cluster
25mKubeVirt: Converge IT infrastructure into one single Kubernetes platform
30mKubeVirt: bringing virtual machines in a Kubernetes world
37mHigh Performance Networking with Kubevirt
36mIt’s All for the Users. More Durable, Secure, and Pluggable. KubeVirt v0.53
31mKubeVirt: What if? What if We Could Create Virtual Machines in Our K8s Cluster?