michaelspedersen
commented
5 years ago Changed from "Spec" project as the plugin by itself it more relevant as a use-case. Workload infrastructure and use-case code has been added already, so this ticket will be closed.
I expect some of the work can be reused in specifications going forward, e.g. DANM and Multus
Goal: Add workload infrastructure and use-case example utilizing SRIOV Network Device Plugin
Deploying the plugin (workload infrastructure): For the example, only physical functions (PFs) will be utilized. To utilize virtual functions (VFs) on
n2.xlargeinstances provided by Packet, additional configuration of the Kubernetes worker node is needed:During initial testing the plugin was successfully tested with VFs, so this can also be done after going through the above configuration steps
Prior to deploying the plugin the host vSwitch (VPP) deployed alongside K8s must be stopped. The PFs used by the plugin will be the same two that was used by the vSwitch.
Deploying the use-case The initial use-case will feature a single container, with one vfio-pci backed PF allocated to it. The container will be running VPP, and due to requirements for both hugepages and access to the vfio-pci kernel module, the container is running in privileged mode. Due to this limitation, the SRIOV Network Device Plugin can not be fully utilized. Under normal operation, the plugin will provide only the requested interfaces to the container, but due to the privileged operating mode all interfaces will be visible (and usable) in the container. To work around this limitation the VPP configuration will be updated based on the information pushed to the container (through an environment variable), and as such only the expected interface will be utilized.
More information about the spec will be added alongside the scripts once added to the CNF Testbed repo.