Skip to main content

Multus and SR-IOV

Using Multus

Multus CNI is a CNI plugin that enables attaching multiple network interfaces to pods. Multus does not replace CNI plugins, instead it acts as a CNI plugin multiplexer. Multus is useful in certain use cases, especially when pods are network intensive and require extra network interfaces that support dataplane acceleration techniques such as SR-IOV.

Multus can not be deployed standalone. It always requires at least one conventional CNI plugin that fulfills the Kubernetes cluster network requirements. That CNI plugin becomes the default for Multus, and will be used to provide the primary interface for all pods.

To enable Multus, add multus as the first list entry in the cni config key, followed by the name of the plugin you want to use alongside Multus (or none if you will provide your own default plugin). Note that multus must always be in the first position of the list. For example, to use Multus with canal as the default plugin you could specify:

# /etc/rancher/rke2/config.yaml
- multus
- canal

This can also be specified with command-line arguments, i.e. --cni=multus,canal or --cni=multus --cni=canal.

For more information about Multus, refer to the multus-cni documentation.

Using Multus with Cilium

To use Cilium with Multus the exclusive config needs to be disabled. You can do this by using the following HelmChartConfig:

# /var/lib/rancher/rke2/server/manifests/rke2-cilium-config.yaml
kind: HelmChartConfig
name: rke2-cilium
namespace: kube-system
valuesContent: |-
exclusive: false

Using Multus with the containernetworking plugins

Any CNI plugin can be used as secondary CNI plugin for Multus to provide additional network interfaces attached to a pod. However, it is most common to use the CNI plugins maintained by the containernetworking team (bridge, host-device, macvlan, etc) as secondary CNI plugins for Multus. These containernetworking plugins are automatically deployed when installing Multus. For more information about these plugins, refer to the containernetworking plugins documentation.

To use any of these plugins, a proper NetworkAttachmentDefinition object will need to be created to define the configuration of the secondary network. The definition is then referenced by pod annotations, which Multus will use to provide extra interfaces to that pod. An example using the macvlan cni plugin with Multus is available in the multus-cni repo.

Multus IPAM plugin options

host-local IPAM plugin allocates ip addresses out of a set of address ranges. It stores the state locally on the host filesystem, therefore ensuring uniqueness of IP addresses on a single host. Therefore, we don't recommend it for multi-node clusters. This IPAM plugin does not require any extra deployment. For more information:

Using Multus with SR-IOV

SR-IOV experimental support was added in v1.21.2+rke2r1, and is fully supported starting with the April 2023 releases: v1.26.4+rke2r1, v1.25.9+rke2r1, and v1.24.13+rke2r1

Using the SR-IOV CNI with Multus can help with data-plane acceleration use cases, providing an extra interface in the pod that can achieve very high throughput. SR-IOV will not work in all environments, and there are several requirements that must be fulfilled to consider the node as SR-IOV capable:

  • Physical NIC must support SR-IOV (e.g. by checking /sys/class/net/$NIC/device/sriov_totalvfs)
  • The host operating system must activate IOMMU virtualization
  • The host operating system includes drivers capable of doing sriov (e.g. i40e, vfio-pci, etc)

The SR-IOV CNI plugin cannot be used as the default CNI plugin for Multus; it must be deployed alongside both Multus and a traditional CNI plugin. The SR-IOV CNI helm chart can be found in the rancher-charts Helm repo. For more information see Rancher Helm Charts documentation.

After installing the SR-IOV CNI chart, the SR-IOV operator will be deployed. Then, the user must specify what nodes in the cluster are SR-IOV capable by labeling them with

kubectl label node $NODE-NAME

Once labeled, the sriov-network-config Daemonset will deploy a pod to the node to collect information about the network interfaces. That information is available through the sriovnetworknodestates Custom Resource Definition. A couple of minutes after the deployment, there will be one sriovnetworknodestates resource per node, with the name of the node as the resource name.

NOTE: the SR-IOV CNI chart from rancher-charts now includes the node-feature-discovery chart as an automatic dependency. This chart deploys a small daemonset that automatically labels each node based on the capabilities detected on that node. This works for both hardware and software features. In particular, node-feature-discovery can automatically add the label when it detects a compatible node. For more information, see the NFD documentation.

However, the latest versions of the sriov-network-operator also include a whitelist of supported hardware so sriov will actually be available only with the NICs on that list. If you want to use the SR-IOV CNI with a NIC that is not on the list, you will need to update the supported-nic-ids configMap yourself.

For more information about how to use the SR-IOV operator, please refer to sriov-network-operator