Skip to content

Separate network nodes versus network services on the controllers

What is compared

The network group of a Kolla-Ansible inventory decides where north-south traffic leaves the cloud: the Neutron L3, DHCP and metadata agents with ML2/OVS, the OVN gateway chassis with ML2/OVN, and, in the sample inventories, HAProxy and keepalived. This page compares two common layouts:

  • Network services on the controllers: the same hosts are in control and network. This is what the all-in-one and many small multinode deployments do (see three controllers and two computes).
  • Dedicated network nodes: two or more hosts only in network, the controllers only in control.

Neither is right in general. What changes is where routed and floating IP traffic flows, what fails together, what an upgrade or a reconfiguration restarts, and how many hosts there are to run.

What the network group carries

From the multinode sample inventory and the roles of Kolla-Ansible 2026.1:

Service Group in the multinode sample inventory ML2/OVS ML2/OVN
HAProxy and keepalived (the VIPs) loadbalancer:children = network yes yes
Neutron L3, DHCP, metadata agents neutron:children = network yes not used
Open vSwitch network, compute yes yes
OVN controller as gateway chassis ovn-controller-network = network (gets enable-chassis-as-gw) not used yes
OVN metadata agent compute, network not used yes
Neutron server, OVN databases and northd control — —
neutron_external_interface (br-ex) hosts in network; computes too with provider networks, DVR or neutron_ovn_distributed_fip

So "dedicated network nodes" in the sample inventory also moves the VIPs off the controllers. That is a choice of the sample, not a rule: the production architecture guide describes network nodes as hosting the Neutron agents "along with haproxy / keepalived", but nothing stops an inventory from setting [loadbalancer:children] to control.

What changes in globals.yml and the inventory

# network services on the controllers
[control]
controller-[1:3]

[network:children]
control

# dedicated network nodes
[control]
controller-[1:3]

[network]
network-[1:2]

globals.yml does not change between the two, except interfaces: neutron_external_interface must exist, with the same name or per-host values, on every host of network. With dedicated network nodes, the controllers no longer need an interface on the external or provider networks at all (unless computes or controllers still carry provider networks for another reason), and the network nodes need one on the tunnel network and the API network.

For the agents to fail over, ML2/OVS needs enable_neutron_agent_ha: true (L3 HA routers and several DHCP agents per network) and at least two hosts in network. With ML2/OVN, every host of network is a gateway chassis and OVN schedules router gateways across them by priority (and by availability zone when neutron_ovn_availability_zones is set, not covered here).

What each layout protects against, and what it does not

Event Network on the controllers Dedicated network nodes
A controller is lost The routers it hosted fail over (L3 HA or OVN gateway priority) and its control services are lost at the same time Only control services are affected; routing is unaffected
A network node is lost (no such host) Routers fail over to the other network nodes; the control plane is unaffected
reconfigure or upgrade of the control plane Restarts API and data plane containers on the same hosts, at the same time Control and network roles can be restarted at different moments, with --limit used with care
Heavy north-south or SNAT traffic Competes with the databases, RabbitMQ and APIs for CPU and NIC bandwidth Isolated on hosts sized for it
Kernel or OVS problem on a gateway host Can take a controller down with it Stays on a network node

Neither layout makes east-west traffic between instances leave the computes: with VXLAN or Geneve it goes compute to compute. And neither replaces distributed routing: with DVR in dvr mode on the computes (neutron_compute_dvr_mode: dvr, ML2/OVS) or neutron_ovn_distributed_fip (ML2/OVN), floating IP traffic of instances leaves from their own compute; only SNAT for instances without a floating IP still goes through the network group. With neutron_compute_dvr_mode: dvr_no_external, floating IP traffic stays centralised on the network group too.

What it costs to run

  • Hosts. Dedicated network nodes are at least two more hosts (one gives no failover), with NICs on the external networks.
  • Cabling and VLANs. External and provider networks reach only the network nodes (and the computes if distributed), not the controllers: a smaller exposure of the control plane to outside networks.
  • Upgrades. More hosts to upgrade, but control plane restarts and data plane restarts can be sequenced separately. The operating guide says that using --limit "is not recommended", because of known bugs (bug 2054348), so test such a sequence on a lab.
  • Moving later. Moving from one layout to the other on a running platform means adding the new hosts to network, moving routers and DHCP networks to them (ML2/OVS) or letting OVN reschedule gateways, then removing the old ones from the group: the procedure the upstream Adding and removing hosts guide gives for controllers, applied to network agents. With the sample inventory it also moves the VIPs, which is a separate, visible change.

What an operator gains and loses

Network on the controllers Dedicated network nodes
Fewer hosts and cables Routing failures and control plane failures are independent
One kind of host to size and patch Network nodes sized for packet processing, controllers for databases and APIs
A controller loss is also a router failover A controller maintenance does not move routers
Simpler for labs, small sites and edge More hosts, more interfaces, more to monitor

Validity and limits

  • Written from the Kolla-Ansible 2026.1 multinode sample inventory (ansible/inventory/multinode), its production architecture and Neutron guides, the ovn-controller role (tasks/setup-ovs.yml, gateway chassis on ovn-controller-network) and ansible/group_vars/all/neutron.yml (computes_need_external_bridge). Not built on a lab for this page. Status draft.
  • No measurement of throughput or failover time is given; those depend on hardware, the number of routers and the back-end.
  • Not covered: BGP-based designs (neutron-bgp-dragent, OVN BGP), hardware offload, and network nodes in separate availability zones (neutron_ovn_availability_zones).

Provenance: written by agent:ozarland-author, from the Kolla-Ansible 2026.1 documentation, sample inventory and code. 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.