Manual install & Configuration

This document discusses what is required for manual installation or integration into a production OpenStack deployment tool of conventional architectures that include the following types of nodes:

  • Controller - Runs OpenStack control plane services such as REST APIs and databases.

  • Network - Runs the layer-2, layer-3 (routing), DHCP, and metadata agents for the Networking service. Some agents optional. Usually provides connectivity between provider (public) and project (private) networks via NAT and floating IP addresses.

    Note

    Some tools deploy these services on controller nodes.

  • Compute - Runs the hypervisor and layer-2 agent for the Networking service.

Packaging

Open vSwitch (OVS) includes OVN beginning with version 2.5 and considers it experimental. From version 2.13 OVN has been released as separate project. The Networking service integration for OVN is now one of the in-tree Neutron drivers so should be delivered with neutron package, but older versions of this integration were delivered with independent package, typically networking-ovn.

Building OVS from source automatically installs OVN for releases older than 2.13. For newer releases it is required to build OVS and OVN separately. For deployment tools using distribution packages, the openvswitch-ovn package for RHEL/CentOS and compatible distributions automatically installs openvswitch as a dependency. Ubuntu/Debian includes ovn-central, ovn-host, ovn-docker, and ovn-common packages that pull in the appropriate Open vSwitch dependencies as needed.

A python-networking-ovn RPM may be obtained for Fedora or CentOS from the RDO project. Since Ussuri release OVN driver is shipped with neutron package. A package based on the older branch of networking-ovn can be found at https://trunk.rdoproject.org/.

Fedora and CentOS RPM builds of OVS and OVN from the master branch of ovs can be found in this COPR repository: https://copr.fedorainfracloud.org/coprs/leifmadsen/ovs-master/.

Controller nodes

Each controller node runs the OVS service (including dependent services such as ovsdb-server) and the ovn-northd service. However, only a single instance of the ovsdb-server and ovn-northd services can operate in a deployment. However, deployment tools can implement active/passive high-availability using a management tool that monitors service health and automatically starts these services on another node after failure of the primary node. See the Frequently Asked Questions for more information.

  1. Install the openvswitch-ovn and networking-ovn packages.

  2. Start the OVS service. The central OVS service starts the ovsdb-server service that manages OVN databases.

    Using the systemd unit:

    # systemctl start openvswitch
    

    Using the ovs-ctl script:

    # /usr/share/openvswitch/scripts/ovs-ctl start  --system-id="random"
    
  3. Configure the ovsdb-server component. By default, the ovsdb-server service only permits local access to databases via Unix socket. However, OVN services on compute nodes require access to these databases.

    • Permit remote database access.

      # ovn-nbctl set-connection ptcp:6641:0.0.0.0 -- \
                  set connection . inactivity_probe=60000
      # ovn-sbctl set-connection ptcp:6642:0.0.0.0 -- \
                  set connection . inactivity_probe=60000
      # if using the VTEP functionality:
      #   ovs-appctl -t ovsdb-server ovsdb-server/add-remote ptcp:6640:0.0.0.0
      

      Replace 0.0.0.0 with the IP address of the management network interface on the controller node to avoid listening on all interfaces.

      Note

      Permit remote access to TCP ports: 6640 (OVS) to VTEPS (if you use vteps), 6642 (SBDB) to hosts running neutron-server, gateway nodes that run ovn-controller, and compute node services like ovn-controller and ovn-metadata-agent. 6641 (NBDB) to hosts running neutron-server.

  4. Start the ovn-northd service.

    Using the systemd unit:

    # systemctl start ovn-northd
    

    Using the ovn-ctl script:

    # /usr/share/openvswitch/scripts/ovn-ctl start_northd
    

    Options for start_northd:

    # /usr/share/openvswitch/scripts/ovn-ctl start_northd --help
    # ...
    # DB_NB_SOCK="/usr/local/etc/openvswitch/nb_db.sock"
    # DB_NB_PID="/usr/local/etc/openvswitch/ovnnb_db.pid"
    # DB_SB_SOCK="usr/local/etc/openvswitch/sb_db.sock"
    # DB_SB_PID="/usr/local/etc/openvswitch/ovnsb_db.pid"
    # ...
    
  5. Configure the Networking server component. The Networking service implements OVN as an ML2 driver. Edit the /etc/neutron/neutron.conf file:

    • Enable the ML2 core plug-in.

      [DEFAULT]
      ...
      core_plugin = ml2
      
    • Enable the OVN layer-3 service.

      [DEFAULT]
      ...
      service_plugins = ovn-router
      
  6. Configure the ML2 plug-in. Edit the /etc/neutron/plugins/ml2/ml2_conf.ini file:

    • Configure the OVN mechanism driver, network type drivers, self-service (tenant) network types, and enable the port security extension.

      [ml2]
      ...
      mechanism_drivers = ovn
      type_drivers = local,flat,vlan,geneve
      tenant_network_types = geneve
      extension_drivers = port_security
      overlay_ip_version = 4
      

      Note

      To enable VLAN self-service networks, make sure that OVN version 2.11 (or higher) is used, then add vlan to the tenant_network_types option. The first network type in the list becomes the default self-service network type.

      To use IPv6 for all overlay (tunnel) network endpoints, set the overlay_ip_version option to 6.

    • Configure the Geneve ID range and maximum header size. The IP version overhead (20 bytes for IPv4 (default) or 40 bytes for IPv6) is added to the maximum header size based on the ML2 overlay_ip_version option.

      [ml2_type_geneve]
      ...
      vni_ranges = 1:65536
      max_header_size = 38
      

      Note

      The Networking service uses the vni_ranges option to allocate network segments. However, OVN ignores the actual values. Thus, the ID range only determines the quantity of Geneve networks in the environment. For example, a range of 5001:6000 defines a maximum of 1000 Geneve networks. On the other hand, these values are still relevant in Neutron context so 1:1000 and 5001:6000 are not simply interchangeable.

      Warning

      The default for max_header_size, 30, is too low for OVN. OVN requires at least 38.

    • Optionally, enable support for VLAN provider and self-service networks on one or more physical networks. If you specify only the physical network, only administrative (privileged) users can manage VLAN networks. Additionally specifying a VLAN ID range for a physical network enables regular (non-privileged) users to manage VLAN networks. The Networking service allocates the VLAN ID for each self-service network using the VLAN ID range for the physical network.

      [ml2_type_vlan]
      ...
      network_vlan_ranges = PHYSICAL_NETWORK:MIN_VLAN_ID:MAX_VLAN_ID
      

      Replace PHYSICAL_NETWORK with the physical network name and optionally define the minimum and maximum VLAN IDs. Use a comma to separate each physical network.

      For example, to enable support for administrative VLAN networks on the physnet1 network and self-service VLAN networks on the physnet2 network using VLAN IDs 1001 to 2000:

      network_vlan_ranges = physnet1,physnet2:1001:2000
      
    • Enable security groups.

      [securitygroup]
      ...
      enable_security_group = true
      

      Note

      The firewall_driver option under [securitygroup] is ignored since the OVN ML2 driver itself handles security groups.

    • Configure OVS database access and L3 scheduler

      [ovn]
      ...
      ovn_nb_connection = tcp:IP_ADDRESS:6641
      ovn_sb_connection = tcp:IP_ADDRESS:6642
      ovn_l3_scheduler = OVN_L3_SCHEDULER
      

      Note

      Replace IP_ADDRESS with the IP address of the controller node that runs the ovsdb-server service. Replace OVN_L3_SCHEDULER with leastloaded if you want the scheduler to select a compute node with the least number of gateway ports or chance if you want the scheduler to randomly select a compute node from the available list of compute nodes.

    • Set ovn-cms-options with enable-chassis-as-gw in Open_vSwitch table’s external_ids column. Then if this chassis has proper bridge mappings, it will be selected for scheduling gateway routers.

      # ovs-vsctl set open . external-ids:ovn-cms-options=enable-chassis-as-gw
      
  7. Start the neutron-server service.

Network nodes

Deployments using OVN native layer-3 and DHCP services do not require conventional network nodes because connectivity to external networks (including VTEP gateways) and routing occurs on compute nodes.

Compute nodes

Each compute node runs the OVS and ovn-controller services. The ovn-controller service replaces the conventional OVS layer-2 agent.

  1. Install the openvswitch-ovn and networking-ovn packages.

  2. Start the OVS service.

    Using the systemd unit:

    # systemctl start openvswitch
    

    Using the ovs-ctl script:

    # /usr/share/openvswitch/scripts/ovs-ctl start --system-id="random"
    
  3. Configure the OVS service.

    • Use OVS databases on the controller node.

      # ovs-vsctl set open . external-ids:ovn-remote=tcp:IP_ADDRESS:6642
      

      Replace IP_ADDRESS with the IP address of the controller node that runs the ovsdb-server service.

    • Enable one or more overlay network protocols. At a minimum, OVN requires enabling the geneve protocol. Deployments using VTEP gateways should also enable the vxlan protocol.

      # ovs-vsctl set open . external-ids:ovn-encap-type=geneve,vxlan
      

      Note

      Deployments without VTEP gateways can safely enable both protocols.

    • Configure the overlay network local endpoint IP address.

      # ovs-vsctl set open . external-ids:ovn-encap-ip=IP_ADDRESS
      

      Replace IP_ADDRESS with the IP address of the overlay network interface on the compute node.

  4. Start the ovn-controller service.

    Using the systemd unit:

    # systemctl start ovn-controller
    

    Using the ovn-ctl script:

    # /usr/share/openvswitch/scripts/ovn-ctl start_controller
    

Verify operation

  1. Each compute node should contain an ovn-controller instance.

    # ovn-sbctl show
      <output>