Skip to content

Latest commit

 

History

History
57 lines (45 loc) · 3.45 KB

sriov.md

File metadata and controls

57 lines (45 loc) · 3.45 KB
Raw

SRIOV Network Device Plugin and SRIOV CNI plugin

Cluster configuration options

In order to install SRIOV Network Device Plugin set sriov_net_dp_enabled value to true in your group vars file. Setting it to false will disable SRIOV Network Device Plugin installation and cause other related options to be ignored.

sriov_net_dp_enabled: true

You can also change the Kubernetes namespace used for the plugin deployment - by default it's kube-system.

sriov_net_dp_namespace: kube-system

It's possible to build and store image locally or use one from public external registry. If you want to use image from the Docker Hub (recommended and faster option), please use false.

sriov_net_dp_build_image_locally: false

The example_net_attach_defs dictionary allows for enabling/disabling automatic creation of example net-attach-def objects. If you want to create the example net-attach-def resource for use with the SRIOV Network Device Plugin and SRIOV CNI, please set below variable to true.

example_net_attach_defs:
  sriov_net_dp: true

Worker node specific options

There's also a set of configuration options that are applied in per-node manner.

First set of variables enables SRIOV for selected network adapters, by setting iommu_enabled as true and passing names of the physical function interfaces. There's also an option to define how many virtual functions should be created for each physical function. In below example dataplane_interfaces configuration will create 6 VFs for 18:00.0 PF interface (PF0) and 4 VFs for 18:00.1 PF interface (PF1). VFs will be attached to default driver defined via 'default_vf_driver'. In our case to 'iavf' driver for PF0 and to 'vfio_pci' driver for PF1. If you need to assign different driver to specific VFs then 'sriov_vfs' section is used. It contains list of pairs vf_name and required driver. In our case VFs vf_00 and vf_05 are attached to 'vfio_pci' driver for PF0. PF1 does not require any specific driver, so 'sriov_vfs' section contains empty list. Name of the first VF is 'vf_00', the second VF has name 'vf_01' and so on. Name of the last VF is derived from 'sriov_numvfs - 1'. In our case 'vf_05'. So, for PF0 VFs devices names are vf_00-vf_05, 6 devices in total. This configuration will also add IOMMU kernel flags, and as a result will reboot the target worker node during deployment.

dataplane_interfaces:
  - bus_info: "18:00.0"
    sriov_numvfs: 6
    default_vf_driver: "iavf"
    sriov_vfs:
      vf_00: "vfio-pci"
      vf_05: "vfio-pci"
  - bus_info: "18:00.1"
    sriov_numvfs: 4
    default_vf_driver: "vfio-pci"
    sriov_vfs: []

dataplane_interfaces can be also configured automatically. All compatible NICs will be discovered and configured. Default VF driver is iavf, which can be changed by modifying dataplane_interface_default_vf_driver. Amount of VFs will be configured to maximum available on your NIC.

dataplane_interface_default_vf_driver: "iavf"
dataplane_interfaces: []

Next option defines whether the SRIOV CNI plugin will be installed on the target worker node. Setting it to true will cause the Ansible scripts to build and install SRIOV CNI plugin in the /opt/cni/bin directory on the target server.

sriov_cni_enabled: true`

Please refer to the SRIOV Network Device Plugin and SRIOV CNI documentation to get more details about sriov resources and usage examples.