Support for DPDK

Support for DPDK

In order to get as close to wire-line speed as possible for virtual machines, we want to install and utilize OVS with DPDK on some compute nodes.

Problem description

DPDK-backed OpenVSwitch and vhostuser features are fully merged in OpenStack Mitaka. With this set of features operator could achieve a boost to the networking performance, and unlike SR-IOV, operator can still control traffic via OpenFlow rules.

Proposed changes

Enabling DPDK requires:

  • Discovering compatible hardware by hardcoded compatible driver and model list, that extracted from DPDK website and drivers source code and hardcoded into Fuel settings.
  • Providing operator with information about compatible NICs.
  • Proper network configuration with dedicated Private network for VLAN segmentation.
  • Configuring CPU isolation for OVS and DPDK PMD.
  • Configuring memory amount for DPDK for every NUMA node.
  • Making configurations on compute side, including nova, neutron, interface binding to the DPDK and OpenVSwitch.

Web UI

On Nodes tab, in Interfaces configuration screen for every interface or bond should be checkbutton to enable DPDK on network interface. It will be shown only if interface is DPDK-capable:


For bond, it will be shown only if all bonded interfaces are DPDK-capable:


Only Private network with VLAN segmentation could be placed on DPDK enabled interface or bond.

If no networks is placed on DPDK interface, that means that no changes should be done in deployment.

These validations should be done both on UI and in API.


The nailgun-agent should collect and send information information about PCI-ID of NIC. This information should be passed to nailgun in the next format

  "meta": {
    "interfaces": [
         "pci_id": String,
         "numa_node": Number

The nailgun will compare PCI-ID against hardcoded into openstack.yaml list of supported hardware. List of supported hardware should also —Āontains dpdk driver name for every PCI-ID, that used in network scheme

  igb_uio: [8086:1010, 8086:1012, ... ]
  <dpdk_driver>: <LIST OF PCI-IDS>

Data model

Next DPDK-related information is stored in interface_properties field of NodeNICInterface:

  • PCI-ID of interface.
  • NUMA node of PCI device.
  • Whether DPDK is enabled by user or not.

Availability should be calculated when nailgun receive information from agent.

Data model for interface_properties should look like this

      "interface_properties" : {
         "dpdk" : {
            "enabled": Boolean or null,
            "available": Boolean
         "pci_id": String,
         "numa_node": Number

For bond interface, only dpdk dict is needed.

When operator configures interface as DPDK to use it for Private network, astute.yaml will be extended as following

  • Node-level parameter enabled will enable DPDK in OpenVSwitch on node. Other parameters should be defined and provided by HugePages and NUMA/CPU pinning features

    • ovs_core_mask: OpenVSwitch cpu core mask in hexa format
    • ovs_pmd_core_mask: OpenVSwitch core mask in hexa format for PMD threads
    • ovs_socket_mem: List of amounts of memory to allocate per NUMA node
      enabled: True
      ovs_core_mask: 0x4
      ovs_pmd_core_mask: 0x6
      ovs_socket_mem: 128,128,128,128
  • Network transformations should include vendor specific attrubute datapath_type: netdev to br-prv bridge:

      - action: add-br
        name: br-prv
        provider: ovs
          datapath_type: netdev
  • Interface should be added using add-port action with provider dpdkovs directly into OVS br-prv bridge:

      - action: add-port
        name: enp1s0f0
        bridge: br-prv
        provider: dpdkovs
  • Bond should be added using add-bond action with provider dpdkovs directly into OVS br-prv bridge:

      - action: add-bond
        bridge: br-prv
        provider: dpdkovs
          mode: balance-rr
        - enp1s0f0
        - enp1s0f1
        name: bond0
  • New vendor specific attribute dpdk_driver should be added from hardcoded list of supported hardware (described above):

           bus_info: '0000:01:00.0'
           driver: ixgbe
           dpdk_driver: igb_uio


Only payload for interfaces and node agent API handlers will be changed as described in Nailgun-agent and Data model sections.



RPC Protocol

Only payload changes.

Fuel Client

User will able to modify DPDK enablement on interface and network assignment it with fuel node --network --download/--upload because of REST API payload changes. No changes required in client.



Fuel Library

Fuel library will consume data from astute.yaml.

  • OpenVSwitch will be configured to use dpdk.
  • vhostuser_socket_dir will be configured in plugin.ini on compute node to enable vhostuser in neutron.
  • l23network will configure interface as dpdk, connect it to ovs bridge and store it in config.


To achieve the same networking performance SR-IOV could be used. Comparing to it, DPDK allows to use experimental Security Groups engine.

Upgrade impact


Security impact


Notifications impact


End user impact

User interface impact is described in Web UI section.

Performance impact

Performance penalties are not expected.

Deployment impact

This feature will require to use VLAN segmentation and dedicated DPDK capable network interface for Private network.

Developer impact


Infrastructure impact

This feature will be tested on virtual environment.

Documentation impact




Primary assignee:
Mandatory design review:

Work Items

  • Collecting information about interfaces
  • Enable DPDK configuration in fuel-library
  • Support of configuring DPDK via fuel API
  • Support of DPDK on UI
  • Manual testing
  • Create a system test for DPDK


This feature depends on HugePages and NUMA/CPU pinning features.

Testing, QA

  • Test API/CLI cases for the configuring DPDK
  • Test WEB UI cases for the configuring DPDK
  • Test that DPDK is discovered and configured properly
  • Performance testing

Acceptance criteria

User should be able to deploy compute nodes with network interface in DPDK mode, and boot a VM with vhostuser and HugePages enabled.