Neutron OVS DVR - Distributed Virtual Router

Neutron Distributed Virtual Router implements the L3 Routers across the Compute Nodes, so that tenants intra VM communication will occur without hiting the Network Node. (East-West Routing)

Also Neutron Distributed Virtual Router implements the Floating IP namespace on every Compute Node where the VMs are located. In this case the VMs with FloatingIPs can forward the traffic to the External Network without reaching the Network Node. (North-South Routing)

Neutron Distributed Virtual Router provides the legacy SNAT behavior for the default SNAT for all private VMs. SNAT service is not distributed, it is centralized and the service node will host the service.

Problem description

Today Neutron L3 Routers are deployed in specific Nodes (Network Nodes) where all the Compute traffic will flow through.

Problem 1: Intra VM traffic flows through the Network Node

In this case even VMs traffic that belong to the same tenant on a different subnet has to hit the Network Node to get routed between the subnets. This would affect Performance.

Problem 2: VMs with FloatingIP also receive and send packets through the Network Node Routers

Today FloatingIP (DNAT) translation done at the Network Node and also the external network gateway port is available only at the Network Node. So any traffic that is intended for the External Network from the VM will have to go through the Network Node.

In this case the Network Node becomes a single point of failure and also the traffic load will be heavy in the Network Node. This would affect the performance and scalability.

This would also help the Neutron networks to be on par with the Nova parity.

Proposed change

The proposal is to distribute the L3 Routers across the Compute Nodes when required by the VMs.

In this case there will be Enhanced L3 Agents running on each and every compute node ( This is not a new agent, this is an updated version of the existing L3 Agent). Based on the configuration in the L3 Agent.ini file, the enhanced L3 Agent will behave in legacy(centralized router) mode or as a distributed router mode.

Also the FloatingIP will have a new namespace created on the specific compute Node where the VM is located Each Compute Node will have one new namespace for FIP, per external network that will be shared among the tenants. An External Gateway port will be created on the Compute Node for the External Traffic to Flow through.

Default SNAT functionality will still be centralized and will be running on a Service Node.

The Metadata agent will be distributed as well and will be hosted on all compute nodes and the Metadata Proxy will be hosted on all the Distributed Routers.

The existing DHCP server will still run on the Service Node. There are future plans to distributed the DHCP. ( This will be addressed in a different blueprint)

This implementation is specific to ML2 with OVS driver.


An alternative is to use a Kernel Module. But we did not pursue this since there was a dependency for the Kernel Module to be part of the upstream linux distribution before we push this patch.

Data model impact

There are couple of minor data model changes that will be addressed by this blueprint.

  1. Router object Data Model.

|     Field      |    Type      | Null | Key | Default |
| tenant_id      | string(256)  | Yes  |     | NULL    |
| id             | string(36)   | NO   | PRI |         |
| name           | string(256)  | YES  |     | NULL    |
| status         | string(16)   | YES  |     | NULL    |
| admin_state_up | boolean      | YES  |     | NULL    |
| gw_port_id     | string(36)   | YES  | MUL | NULL    |
| enable_snat    | boolean      | NO   |     |         |
| distributed    | boolean      | YES  |     | NULL    |

Add “distributed” flag to the router object data model. This will enable the agent to take necessary action based on the router model.

  1. SNAT Agent to host mapping data model.

A new table for the service node enhanced L3 agent to track the SNAT service on each node.

| Field            | Type         | Null | Key | Default |
| id               | string(36)   | NO   | PRI |         |
| router_id        | string(36)   | YES  | MUL | NULL    |
| host_id          | string(255)  | YES  |     | NULL    |
| l3_agent_id      | string(36)   | YES  | MUL | NULL    |
  1. ML2 DVR interface binding data model.







port_id host router_id vif_type vif_details vnic_type profile cap_port_filter driver segment status

string(36) string(255) string(36) string(64) string(4095) string(64) string(36) boolean string(64) string(36) string(16)




A new table that is used to hold port bindings for DVR router interfaces only. This is similar to the portbindings table, but this table will hold bindings only for dvr router interfaces.

The original portbindings table will also hold one-binding row for a dvr router interface, but that won’t hold binding information. That binding row is held there, only to ensure transparency of dvr presence to the tenants themselves.

Some of the significant fields in the above are: port_id - This refers to the port id of the DVR Router interface for which this binding is applied to. The port-id will refer to id field of the port table. host - This holds the host on which the DVR interface is bound. router_id - This field indicates for which router interface, this binding belongs. status - This field represents the status of the dvr interface port on the host, which is represented by this binding.

The status field value of the single-binding row for dvr router interface in the original portbindings table will now be an ORed result of the above status field of all such bindings available in the above table for dvr router interfaces.

  1. ML2 DVR Unique MAC Address Table:







host mac_address

string(255) string(32)




A new table that is used to hold Unique DVR Base mac assigned to OVS L2 agent that is running in DVR Mode.

For any given host where an OVS L2 Agent is running, only one MAC Address from the DVR Base Mac pool is allocated to that OVS L2 Agent. This allocation rpc cycle, completes during init() of the OVS L2 Agent.

In order to make OVS L2 Agent run in DVR Mode, enable_distributed_routing flag must be set to True in the [agent] section of ml2 ini file (ml2_conf.ini).

Similarly, the DVR Base Mac Address which represents start of the pool, need to be defined in neutron.conf

REST API impact

router-create Create a router for a given tenant.

router-create --name another_router --distributed=true

Admin can only set this attribute. The tenants need not be aware about this attribute in the router table. So it is not visible to the tenant.


POST /v2.0/routers
Accept: application/json




router-show Show information of a given router.


GET /v2.0/routers/a9254bdb-2613-4a13-ac4c-adc581fba50d
Accept: application/json



router-update Create a router for a given tenant.

Admin can only update a centralized router to a distributed router.

Note: Admin can only update a centralized router to a distributed router and not the other way around. For the first release we are targeting only from centralized to distributed.

Admin only context:

neutron router-update router1 --distributed=True

Admin only CLI commands:

l3-agent-list-hosting-snat   List L3 agents hosting a snat service.

This command will list the agent with the router-id and SNAT IP.

l3-agent-snat-add            Associate a snat namespace to an L3 agent.

This command will allow an admin to associate a SNAT namespace to an agent. This command will take the router ID as an argument.

l3-agent-snat-remove         Remove snat association from an L3 agent.

This command will allow an admin to remove or disassociate a SNAT service from the agent.

Security impact

Need to make sure the existing FWaaS and the Security Group Rules are not affected by the DVR.

Notifications impact


Other end user impact

Yes this change will have some impact on the python-neutronclient

The Admin level API proposed above will have to be implemented in the CLI.

Also there is an impact with Horizon to address the admin level API mentioned above.

Performance Impact

  • Improves Performance.

Inter VM traffic between the tenant’s subnet need not reach the router in the Network node to get routed and will be routed locally from the Compute Node. This would increase the performance substantially.

Also the Floating IP traffic for a VM from a Compute Node will directly hit the external network from the compute node, instead of going through the router on the network node.

Other deployer impact

Global Configuration to enable Distributed Virtual Router.


# To enable distributed routing this flag need to be enabled.
# It can be either True or False.
# If False it will work in a legacy mode.
# If True it will work in a DVR mode.

#router_distributed = True

# ovs_neutron_plugin.ini

# This flag need to be enabled for the L2 Agent to address
# DVR rules

#enable_distributed_routing = True

# l3_agent.ini
# This flag is required by the L3 Agent as well to run the L3
# agent in a Distributed Mode.
#distributed_agent = True

This will be disabled by default.

NOTE: This is for backward compatibility. For migration the admin might have to run the db-migration script and also re-start the agents with the right configuration to take effect.

If Cloud admin wanted to enable the feature this can be configured.

It currently uses the existing OVS binary in Linux Distribution. So there should not be any new binaries.

Developer impact

Multinode Devstack setup may be required to develop and test.

Services Impact - Some of the services such as the VPN and FW should be refactored to accomodate the distributed virtual routers. The respective services team will be working with the DVR team to refactor the services.



Primary assignee:

  • <swaminathan-vasudevan>

Other contributors:

  • <rajeev-grover>

  • <mbirru>

  • <michael-smith6>

  • <vivekanandan-narasimhan>

Work Items

  1. L3 Plugin Extension for DVR

  2. ML2 Plugin/OVS Agent for DVR

  3. L3 Enhanced Agent for DVR

  4. L3 Agent Scheduler for DVR

  5. L3 Driver/iplib for DVR


OVS (2.01 and above), L2-Pop.


Yes. Since we are implementing the Distributed Nature of routers, there need to be multinode setup for testing this feature so that the rules and actual namespace creation for the routers can be validated.

Single node infrastructure to test the feature may still be possible, but we need to validate.

Continuous integration testing to test the dvr at the gate will be considered.

Documentation Impact

Yes. There will be documentation impact and so documentation has to be modified to address the new deployment scenario.