To InfiniBand and Beyond!
summary
KubeVirt VMs receive real InfiniBand VFs via the SR-IOV Network Operator plus the IB SR-IOV CNI, using VFIO on Mellanox ConnectX‑7 cards at up to 400 Gbps in the demo. Two nodes run one VM each, VFs pass through into guests, kernels load RDMA and IPoIB modules, and InfiniBand pings and stats validate connectivity; it is a working proof‑of‑concept, upstream code exists but not yet productized. Future work uses DRA for per‑VF attributes and co‑scheduling.
AI-written summary · about
chapters (13)
- 0:00InfiniBand and KubeVirt intro
- 1:03What is InfiniBand
- 3:56RDMA kernel bypass
- 6:24InfiniBand SR-IOV in Kubernetes
- 9:15SR-IOV operator config
- 11:25KubeVirt overview
- 13:12KubeVirt VM network bindings
- 15:38Live demo: InfiniBand with KubeVirt
- 21:02Current scope and hardware challenges
- 22:29Cloud-native acceleration with DRA
- 26:25KubeVirt DRA adoption
- 27:48Summary
- 29:34Q&A
related talks
37mHigh Performance Networking with Kubevirt
27mKubeVirt vDPA Workflow: From Host to Pod to Domain
27mHigh Performance KubeVirt in Action
36mIt’s All for the Users. More Durable, Secure, and Pluggable. KubeVirt v0.53
36mBringing vDPA to Life in Kubernetes
34mRDMA Enabled Kubernetes for High Performance Computing
20mKubeVirt Benchmar...
32mKubeVirt + Istio: Virtual Machines on Kubernetes
