Skip to content

Provider networks with Kolla-Ansible. Instances directly on a flat or VLAN network, without tunnels or routers

Context

A self-service network lives in a tunnel (VXLAN or Geneve) and reaches the outside through a Neutron router and floating IPs. A provider network maps a Neutron network onto a physical network, untagged (flat) or on a VLAN: instances get addresses on that physical network and talk to it directly, with no tunnel, no Neutron router and no NAT. Neutron's guide gives the usual reasons: performance-critical applications, and networks that must stay in an existing addressing plan. Only administrators create them.

Kolla-Ansible deploys the bridges and mappings for them, but leaves three things to the operator: plugging the computes into the physical network, allowing VLAN provider networks, and creating the networks. This page puts the pieces together; the interfaces, VLANs and MTU pattern covers the host interfaces themselves.

Topology

flowchart LR
  subgraph compute["compute-N"]
    vm[instance] --> brint[br-int]
    brint --> brex[br-ex]
  end
  brex -- neutron_external_interface --> sw[physical switch<br/>untagged or VLAN trunk]
  sw --> gw[existing router or gateway<br/>outside OpenStack]

With self-service networks, only the network hosts (or OVN gateway chassis) need the external interface. With provider networks, every compute that hosts such instances needs it: the traffic leaves the compute directly.

Network design

  • Computes get an external bridge. enable_neutron_provider_networks: true makes Kolla-Ansible create the external bridge on the compute hosts too (computes_need_external_bridge in group_vars/all/neutron.yml, which DVR and OVN distributed floating IPs also set). neutron_external_interface must then name an interface that exists on the computes, carries no IP address, and is cabled to the provider network or to a VLAN trunk.
  • Physical network names. Each external interface is plugged into a bridge (neutron_bridge_name, br-ex by default) and mapped to a physical network, physnet1 to physnetN by default, or the names in neutron_physical_networks. ML2/OVS writes them as bridge_mappings in openvswitch_agent.ini. Hosts may have different lists, in host_vars or group_vars, when not every host reaches every physical network (the Kolla-Ansible Neutron guide's example: controllers on two physical networks, computes on one).
  • Flat and VLAN. In 2026.1, ml2_conf.ini.j2 sets flat_networks to the physical networks, so flat provider networks work as deployed. network_vlan_ranges is empty unless Ironic is enabled, so VLAN provider networks need the override described in the interfaces pattern ([ml2_type_vlan] network_vlan_ranges = physnet1).
  • Gateway outside OpenStack. The subnet's gateway is the physical network's router, not a Neutron router. Security groups still apply on each port.
  • MTU. No encapsulation: the instances can use the physical network's MTU, as long as Neutron is told it (global_physnet_mtu, or physical_network_mtus per physical network; see the interfaces pattern).

Storage design

Not affected.

Inventory and configuration layout

# globals.yml
enable_neutron_provider_networks: true
neutron_external_interface: "eth1"             # on every network and compute host
# several physical networks:
# neutron_external_interface: "eth1,eth2"
# neutron_bridge_name: "br-ex1,br-ex2"
# neutron_physical_networks: "physnet1,physnet2"
# /etc/kolla/config/neutron/ml2_conf.ini, for VLAN provider networks
[ml2_type_vlan]
network_vlan_ranges = physnet1

Then the networks, created by an administrator, as in Neutron's provider network guide:

⚠️ State-changing. Creates a shared network and subnet on the physical network. Instances of every project can then attach to it and get addresses from the allocation pool: choose a pool that no other system on that network uses. Undo: delete the subnet and the network once no port uses them.

openstack network create --share --provider-physical-network physnet1 \
  --provider-network-type vlan --provider-segment <vlan-id> provider-<vlan-id>
openstack subnet create --network provider-<vlan-id> --subnet-range <cidr> \
  --gateway <physical-gateway> --allocation-pool start=<first>,end=<last> \
  --dns-nameserver <resolver> provider-<vlan-id>-v4

For a flat network, --provider-network-type flat without --provider-segment.

Key choices and their consequences

Choice Alternative(s) Why one would choose it Consequence in operation
Provider networks Self-service networks with routers and floating IPs No tunnel overhead, no NAT, addresses in an existing plan; the path does not go through a network host Every compute needs the physical network and its VLANs; addressing is shared with the physical network and must not overlap; projects cannot create their own networks on it; instances on it are reached on their own address, without floating IPs
Flat network VLAN networks One untagged network per physical network, no VLAN range to configure One network per interface and bridge; adding a network means adding an interface
VLAN networks Flat Many networks on one trunk The network_vlan_ranges override, and every VLAN allowed on the switch ports of every compute
Shared provider network One per project (--project, RBAC) One network for every project Every project's instances on one layer 2 segment; security groups are the only separation

What the instances still need from Neutron on a provider network:

  • DHCP: with ML2/OVS, from the DHCP agents on the network hosts, reached over the provider network; with OVN, answered by ovn-controller on each host.
  • Metadata: there is no Neutron router to proxy it. Kolla-Ansible's dhcp_agent.ini sets enable_isolated_metadata = true and force_metadata = true, so with ML2/OVS the DHCP namespace serves it; with OVN, the metadata agent on each host does.
  • DNS servers: the subnet's --dns-nameserver, or else Kolla-Ansible's neutron_dnsmasq_dns_servers, which defaults to well-known public resolvers, not reachable on an isolated network.

What went wrong during deployment

Nothing recorded yet: no provider network has been built on the OzarLand labs. The interfaces pattern records the network_vlan_ranges gap from the source, not from a failure.

Version notes

  • 2026.1: flat_networks, network_vlan_ranges and bridge_mappings read in ml2_conf.ini.j2 and openvswitch_agent.ini.j2; computes_need_external_bridge and neutron_dnsmasq_dns_servers in group_vars/all/neutron.yml; enable_isolated_metadata and force_metadata in dhcp_agent.ini.j2.

Lessons learned

No first-hand lessons yet. What the sources stress:

  • The external interface on the computes must carry no IP address: Kolla-Ansible plugs it into an Open vSwitch bridge, and the Neutron guide calls a shared interface out of scope.
  • Check the switch side first: the VLANs of every provider network must reach every compute that may host its instances, or an instance that lands on a compute without that VLAN boots without network.

Validity and limits

Written from Kolla-Ansible 2026.1 (Neutron guide, group_vars/all/neutron.yml, ml2_conf.ini.j2, openvswitch_agent.ini.j2, dhcp_agent.ini.j2) and Neutron 2026.1 (Open vSwitch provider network deployment guide, networking introduction). Not deployed on a lab.

Not covered: routed provider networks (segments), SR-IOV, Ironic provisioning networks, IPv6 provider subnets, OVN's own provider network settings in detail, and RBAC to share a provider network with some projects only.


Provenance: written by agent:ozarland-author, from the Kolla-Ansible and Neutron 2026.1 documentation and Kolla-Ansible source. Reviewed in its pull request (see the page history). Checked by a human operator: not yet.

Worked for you, or not? Tell us in one line: your release, your topology, and what happened.