Set up SR-IOV networking

This document describes how to set up single-root input/output virtualization (SR-IOV) networking for Google Distributed Cloud. SR-IOV provides I/O virtualization to make a network interface card (NIC), available as network devices in the Linux kernel. This lets you manage and assign network connections to your pods. Performance is improved as packets move directly between the NIC and the pod.

Use this feature if you require fast networking to your pod workloads. SR-IOV for Google Distributed Cloud lets you configure the virtual functions (VFs) on the supported devices of your cluster nodes. You can also specify the particular kernel module to bind to the VFs.

This feature is available for clusters that run workloads, such as hybrid, standalone, and user clusters. The SR-IOV networking feature requires the cluster to have at least two nodes.

The setup process consists of the following high-level steps:

  1. Configure the cluster to enable SR-IOV networking.
  2. Configure the SR-IOV operator, a SriovOperatorConfig custom resource.
  3. Set up SR-IOV policies and configure your VFs.
  4. Create a NetworkAttachmentDefinition custom resource that references your VFs.

Requirements

The SR-IOV networking feature requires the official drivers for the network adapters to be present on the cluster nodes. Install the drivers before using the SR-IOV operator. Also, to use the vfio-pci module for your VFs, ensure that the module is available on the nodes where it's to be used.

Enable SR-IOV networking for a cluster

To enable SR-IOV networking for Google Distributed Cloud, add the multipleNetworkInterfaces field and the sriovOperator field to the clusterNetwork section of the Cluster object and set both fields to true.

apiVersion: baremetal.cluster.gke.io/v1
kind: Cluster
metadata:
  name: cluster1
spec:
  clusterNetwork:
    multipleNetworkInterfaces: true
    sriovOperator: true
...

The sriovOperator field is mutable, and can be changed after cluster creation.

Configure the SR-IOV operator

The SriovOperatorConfig custom resource provides global configuration for the SR-IOV networking feature. This bundled custom resource has the name default and is in the gke-operators namespace. The SriovOperatorConfig custom resource is honored for this name and namespace only.

You can edit this object with the following command:

kubectl -n gke-operators edit sriovoperatorconfigs.sriovnetwork.k8s.cni.cncf.io default

Here's an example of a SriovOperatorConfig custom resource configuration:

apiVersion: sriovnetwork.k8s.cni.cncf.io/v1
kind: SriovOperatorConfig
metadata:
  name: default
  namespace: gke-operators
spec:
  configDaemonNodeSelector:
    nodePool: "withSriov"
  disableDrain: false
  logLevel: 0

The configDaemonNodeSelector section lets you limit what nodes the SR-IOV operator can handle. In the preceding example, the operator is limited to only nodes that have a nodePool: withSriov label. If configDaemonNodeSelector field isn't specified, the following default labels are applied:

beta.kubernetes.io/os: linux
node-role.kubernetes.io/worker: ""

The disableDrain field specifies whether to perform a Kubernetes node drain operation before the node has to be rebooted or before a specific VF configuration is changed.

Create SR-IOV policies

To configure specific VFs in your cluster, you have to create a SriovNetworkNodePolicy custom resource in the gke-operators namespace.

Here's an example manifest for a SriovNetworkNodePolicy custom resource:

apiVersion: sriovnetwork.k8s.cni.cncf.io/v1
kind: SriovNetworkNodePolicy
metadata:
  name: policy-1
  namespace: gke-operators
spec:
  deviceType: "netdevice"
  mtu: 1600
  nodeSelector:
    baremetal.cluster.gke.io/node-pool: node-pool-1
  nicSelector:
    pfNames:
    - enp65s0f0
    deviceID: "1015"
    rootDevices:
    - 0000:01:00.0
    vendor: "15b3"
  numVfs: 4
  priority: 80
  resourceName: "mlnx"

The nodeSelector section lets you further limit the nodes on which the VFs have to be created. This limitation is on top of the selectors from the SriovOperatorConfig described in the previous section.

The deviceType field specifies the kernel module to use for the VFs. Available options for deviceType are:

  • netdevice for VF-specific standard kernel module
  • vfio-pci for the VFIO-PCI driver

The resourceName defines what name the VFs are represented as in the Kubernetes Node.

After the configuration process is done, your selected cluster nodes contain the defined resource as presented in the following example (notice the gke.io/mlnx):

apiVersion: v1
kind: Node
metadata:
  name: worker-01
spec:

status:
  allocatable:
    cpu: 47410m
    ephemeral-storage: "210725550141"
    gke.io/mlnx: "4"
    hugepages-1Gi: "0"
    hugepages-2Mi: "0"
    memory: 59884492Ki
    pods: "250"
  capacity:
    cpu: "48"
    ephemeral-storage: 228651856Ki
    gke.io/mlnx: "4"
    hugepages-1Gi: "0"
    hugepages-2Mi: "0"
    memory: 65516492Ki
    pods: "250"

The operator will always add the gke.io/ prefix to every resource you define with SriovNetworkNodePolicy.

Specify a NIC selector

For the SriovNetworkNodePolicy to function properly, specify at least one selector in the nicSelector section. This field contains multiple options on how to identify specific physical functions (PFs) in your cluster nodes. Most of the information required by this field is discovered for you and saved in the SriovNetworkNodeState custom resource. There will be an object per each node that this operator can handle.

Use the following command to view all the available nodes:

kubectl -n gke-operators get sriovnetworknodestates.sriovnetwork.k8s.cni.cncf.io -o yaml

Here's an example of a node:

apiVersion: sriovnetwork.k8s.cni.cncf.io/v1
kind: SriovNetworkNodeState
metadata:
  name: worker-01
  namespace: gke-operators
spec:
  dpConfigVersion: "6368949"
status:
  interfaces:
  - deviceID: "1015"
    driver: mlx5_core
    eSwitchMode: legacy
    linkSpeed: 10000 Mb/s
    linkType: ETH
    mac: 1c:34:da:5c:2b:9c
    mtu: 1500
    name: enp1s0f0
    pciAddress: "0000:01:00.0"
    totalvfs: 4
    vendor: 15b3
  - deviceID: "1015"
    driver: mlx5_core
    linkSpeed: 10000 Mb/s
    linkType: ETH
    mac: 1c:34:da:5c:2b:9d
    mtu: 1500
    name: enp1s0f1
    pciAddress: "0000:01:00.1"
    totalvfs: 2
    vendor: 15b3
  syncStatus: Succeeded

Set Physical Function partitioning

Pay special attention to the pfNames field of the nicSelector section. In addition to defining the exact PF to use, it lets you specify the exact VFs to use for the specified PF and the resource defined in the policy.

Here's an example:

apiVersion: sriovnetwork.k8s.cni.cncf.io/v1
kind: SriovNetworkNodePolicy
metadata:
  name: policy-1
  namespace: gke-operators
spec:
  deviceType: "netdevice"
  mtu: 1600
  nodeSelector:
    baremetal.cluster.gke.io/node-pool: node-pool-1
  nicSelector:
    pfNames:
    - enp65s0f0#3-6
    deviceID: "1015"
    rootDevices:
    - 0000:01:00.0
    vendor: "15b3"
  numVfs: 7
  priority: 80
  resourceName: "mlnx"

In the preceding example, the gke.io/mlnx resource uses VFs numbered 3-6 only and shows just four available VFs. Since the VFs are always created from the zero index, your requested number of VFs, numVfs, has to be at least as high as the range-closing value (counting from zero). This numbering logic is why numVfs is set to 7 in the preceding example. If you set a range from 3 to 4 (enp65s0f0#3-4), your numVfs must be at least 5.

When the partitioning isn't specified, the numVfs defines the VFs range that is being used, which always starts from zero. For example, if you set numVfs=3 without specifying partitioning, VFs 0-2 are used.

Understand policy priority

You can specify multiple SriovNetworkNodePolicy objects to handle various vendors or different VF configurations. Managing multiple objects and vendors might become troublesome when multiple policies reference the same PF. To handle such situations, the priority field resolves the conflicts on a per-node basis.

Here is the prioritization logic for overlapping PF policies:

  1. A higher priority policy overwrites one with lower priority only when PF partitioning is overlapping.

  2. Same priority policies are merged:

    1. Policies are sorted by name and processed in that order
    2. Policies with overlapping PF partitioning are overwritten
    3. Policies with non-overlapping PF partitioning are merged and all present

A high priority policy is one with lower numerical value in the priority field. For example, the priority is higher for a policy with priority: 10, than for a policy with priority: 20.

The following sections provide policy examples for different partitioning configurations.

Partitioned PF

Deploying the following two SriovNetworkNodePolicy manifests results in two available resources: gke.io/dev-kernel and gke.io/dev-vfio. Each resource has two VFs that are non-overlapping.

kind: SriovNetworkNodePolicy
metadata:
  name: policy-1
spec:
  deviceType: "netdevice"
  nodeSelector:
    baremetal.cluster.gke.io/node-pool: node-pool-1
  nicSelector:
    pfNames:
    - enp65s0f0#0-1
  numVfs: 2
  priority: 70
  resourceName: "dev-kernel"
kind: SriovNetworkNodePolicy
metadata:
  name: policy-2
spec:
  deviceType: "vfio-pci"
  nodeSelector:
    baremetal.cluster.gke.io/node-pool: node-pool-1
  nicSelector:
    pfNames:
    - enp65s0f0#2-3
  numVfs: 4
  priority: 70
  resourceName: "dev-vfio"

Overlapping PF partitioning

Deploying the following two SriovNetworkNodePolicy manifests results in only the gke.io/dev-vfio resource being available. The policy-1 VF range is 0-2, which overlaps with policy-2. Due to naming, policy-2 is processed after policy-1. Therefore, only the resource specified in policy-2, gke.io/dev-vfio, is available.

kind: SriovNetworkNodePolicy
metadata:
  name: policy-1
spec:
  deviceType: "netdevice"
  nodeSelector:
    baremetal.cluster.gke.io/node-pool: node-pool-1
  nicSelector:
    pfNames:
    - enp65s0f0
  numVfs: 3
  priority: 70
  resourceName: "dev-kernel"
kind: SriovNetworkNodePolicy
metadata:
  name: policy-2
spec:
  deviceType: "vfio-pci"
  nodeSelector:
    baremetal.cluster.gke.io/node-pool: node-pool-1
  nicSelector:
    pfNames:
    - enp65s0f0#2-3
  numVfs: 4
  priority: 70
  resourceName: "dev-vfio"

Non-overlapping PF partitioning with different priorities

Deploying the following two SriovNetworkNodePolicy manifests results in two available resources: gke.io/dev-kernel and gke.io/dev-vfio. Each resource has two VFs that are non-overlapping. Even though policy-1 has higher priority than policy-2, since the PF partitioning is non-overlapping, we merge the two policies.

kind: SriovNetworkNodePolicy
metadata:
  name: policy-1
spec:
  deviceType: "netdevice"
  nodeSelector:
    baremetal.cluster.gke.io/node-pool: node-pool-1
  nicSelector:
    pfNames:
    - enp65s0f0
  numVfs: 2
  priority: 10
  resourceName: "dev-kernel"
kind: SriovNetworkNodePolicy
metadata:
  name: policy-2
spec:
  deviceType: "vfio-pci"
  nodeSelector:
    baremetal.cluster.gke.io/node-pool: node-pool-1
  nicSelector:
    pfNames:
    - enp65s0f0#2-3
  numVfs: 4
  priority: 70
  resourceName: "dev-vfio"

Check SR-IOV policy setup status

When you apply the SR-IOV policies, you can track and view the final configuration of the nodes in the SriovNetworkNodeState custom resource for the specific node. In the status section, the syncStatus field represents the current stage for the configuration daemon. The Succeeded state indicates that configuration is finished. The spec section of the SriovNetworkNodeState custom resource defines the final state of VFs configuration for that Node, based on the number of policies and their priorities. All the created VFs will be listed in the status section for the specified PFs.

Here is an example SriovNetworkNodeState custom resource:

apiVersion: sriovnetwork.k8s.cni.cncf.io/v1
kind: SriovNetworkNodeState
metadata:
  name: worker-02
  namespace: gke-operators
spec:
  dpConfigVersion: "9022068"
  interfaces:
  - linkType: eth
    name: enp1s0f0
    numVfs: 2
    pciAddress: "0000:01:00.0"
    vfGroups:
    - deviceType: netdevice
      policyName: policy-1
      resourceName: mlnx
      vfRange: 0-1
status:
  interfaces:
  - Vfs:
    - deviceID: "1016"
      driver: mlx5_core
      mac: 96:8b:39:d8:89:d2
      mtu: 1500
      name: enp1s0f0np0v0
      pciAddress: "0000:01:00.2"
      vendor: 15b3
      vfID: 0
    - deviceID: "1016"
      driver: mlx5_core
      mac: 82:8e:65:fe:9b:cb
      mtu: 1500
      name: enp1s0f0np0v1
      pciAddress: "0000:01:00.3"
      vendor: 15b3
      vfID: 1
    deviceID: "1015"
    driver: mlx5_core
    eSwitchMode: legacy
    linkSpeed: 10000 Mb/s
    linkType: ETH
    mac: 1c:34:da:5c:2b:9c
    mtu: 1500
    name: enp1s0f0
    numVfs: 2
    pciAddress: "0000:01:00.0"
    totalvfs: 2
    vendor: 15b3
  - deviceID: "1015"
    driver: mlx5_core
    linkSpeed: 10000 Mb/s
    linkType: ETH
    mac: 1c:34:da:5c:2b:9d
    mtu: 1500
    name: enp1s0f1
    pciAddress: "0000:01:00.1"
    totalvfs: 2
    vendor: 15b3
  syncStatus: Succeeded

Create a NetworkAttachmentDefinition custom resource

After you successfully configure the VFs on the cluster, and they are visible in the Kubernetes Node as a resource, you need to create a NetworkAttachmentDefinition that references the resource. Make the reference with a k8s.v1.cni.cncf.io/resourceName annotation. Here is an example NetworkAttachmentDefinition manifest that references the gke.io/mlnx resource:

apiVersion: "k8s.cni.cncf.io/v1"
kind: NetworkAttachmentDefinition
metadata:
  name: gke-sriov-1
  annotations:
    k8s.v1.cni.cncf.io/resourceName: gke.io/mlnx
spec:
  config: '{
      "cniVersion": "0.3.0",
      "name": "mynetwork",
      "type": "sriov",
      "ipam": {
        "type": "whereabouts",
        "range": "21.0.108.0/21",
        "range_start": "21.0.111.16",
        "range_end": "21.0.111.18"
      }
    }'

The NetworkAttachmentDefinition must have the sriov as the CNI type. Reference any deployed NetworkAttachmentDefinition custom resources in your pods with a k8s.v1.cni.cncf.io/networks annotation.

Here's an example of how to reference the preceding NetworkAttachmentDefinition custom resource in a pod:

apiVersion: v1
kind: Pod
metadata:
  name: samplepod
  annotations:
    k8s.v1.cni.cncf.io/networks: gke-sriov-1
spec:
  containers:
  ...

When referencing a NetworkAttachmentDefinition custom resource in workloads, you don't have to worry about the Pods' resources definitions, or placement in specific Nodes, which is done automatically for you.

The following example shows a NetworkAttachmentDefinition custom resource with a VLAN configuration. In this sample, every VF belongs to the 100 VLAN:

apiVersion: "k8s.cni.cncf.io/v1"
kind: NetworkAttachmentDefinition
metadata:
  name: gke-sriov-vlan-100
  annotations:
    k8s.v1.cni.cncf.io/resourceName: gke.io/mlnx
spec:
  config: '{
      "cniVersion": "0.3.0",
      "name": "mynetwork",
      "type": "sriov",
      "vlan": 100,
      "ipam": {
        "type": "whereabouts",
        "range": "21.0.100.0/21"
      }
    }'

Additional information

The following sections contain information to help you configure SR-IOV networking.

Node reboots

When the SR-IOV operator configures the nodes, the nodes may need to be rebooted. Rebooting nodes might be needed during VF or kernel configuration. The kernel configuration involves enabling support of the SR-IOV functionality in the operating system.

Supported Network Adapters

The following table lists the supported network adapters for version 1.30.x clusters:

Name Vendor ID Device ID VF device ID
Intel i40e XXV710 8086 158a 154c
Intel i40e 25G SFP28 8086 158b 154c
Intel i40e 10G X710 SFP 8086 1572 154c
Intel i40e XXV710 N3000 8086 0d58 154c
Intel i40e 40G XL710 QSFP 8086 1583 154c
Intel ice Columbiaville E810-CQDA2 2CQDA2 8086 1592 1889
Intel ice Columbiaville E810-XXVDA4 8086 1593 1889
Intel ice Columbiaville E810-XXVDA2 8086 159b 1889
Nvidia mlx5 ConnectX-4 15b3 1013 1014
Nvidia mlx5 ConnectX-4LX 15b3 1015 1016
Nvidia mlx5 ConnectX-5 15b3 1017 1018
Nvidia mlx5 ConnectX-5 Ex 15b3 1019 101a
Nvidia mlx5 ConnectX-6 15b3 101b 101c
Nvidia mlx5 ConnectX-6_Dx 15b3 101d 101e
Nvidia mlx5 MT42822 BlueField-2 integrated ConnectX-6 Dx 15b3 a2d6 101e
Broadcom bnxt BCM57414 2x25G 14e4 16d7 16dc
Broadcom bnxt BCM75508 2x100G 14e4 1750 1806