Lower MTU for migration from a VXLAN OVS deployment

If your pre-migration OVS deployment uses the VXLAN tunneling protocol, you might need to reduce the network maximum transmission unit (MTU) by 8 bytes before migrating to OVN, which uses the Geneve tunneling protocol.

Before you begin

About this task

Important
Consider performing this procedure in a dedicated maintenance window period before the migration.

VXLAN packets reserve 50 bytes of data for header content. This includes 42 bytes of standard outer headers plus an 8-byte VXLAN header. If the physical network uses the standard ethernet MTU of 1500 bytes, you can set the MTU on your VXLAN networks to 1450 and traffic can pass without fragmentation.

Geneve packets reserve 58 bytes of data for header content. This includes the 42 bytes of standard outer headers plus a 16-byte Geneve header. Thus, if the physical network has an MTU less than 1508, you must reduce the MTU on your pre-migration OpenStack VXLAN networks by 8 bytes to avoid the need for fragmentation.

Note
If your physical network can transmit at least 58 bytes more than your OpenStack VXLAN network MTU without fragmentation, skip this procedure and proceed to Migrate from OVS to OVN. For example, you can skip this procedure if your physical network is configured for 9000-byte jumbo frames and your openstack network MTU is 8942 or less.

The RHOSP OVN migration tool automatically lowers the MTU by 8 bytes on VXLAN and GRE overcloud networks. In the following procedure, you use the tool to:

  • Increase the frequency of DHCP renewals by reducing the DHCP T1 timer to 30 seconds.
  • Reduce the MTU size on existing VXLAN networks by 8 bytes.

If your deployment does not use DHCP to configure all VM instances, you must manually reduce MTU on the excluded instances.

Procedure

  1. Run ovn_migration.sh reduce-dhcp-t1.

    This lowers the T1 parameter of the internal neutron DHCP servers that configure the dhcp_renewal_time in /var/lib/config-data/puppet-generated/neutron/etc/neutron/dhcp_agent.ini in all the nodes where DHCP agent is running.

    $ ovn_migration.sh reduce-dhcp-t1 | sudo tee -a /var/log/ovn_migration_output.txt
  2. Verify that the T1 parameter has propagated to existing VMs. The process might take up to four hours:
    1. Log in to one of the Compute nodes.
    2. Run tcpdump over one of the VM taps attached to a project network.
    3. If T1 propagation is successful, expect to see requests occur approximately every 30 seconds: 
      [heat-admin@overcloud-novacompute-0 ~]$ sudo tcpdump -i tap52e872c2-e6 port 67 or port 68 -n

tcpdump: verbose output suppressed, use -v or -vv for full protocol decode

listening on tap52e872c2-e6, link-type EN10MB (Ethernet), capture size 262144 bytes

13:17:28.954675 IP 192.168.99.5.bootpc > 192.168.99.3.bootps: BOOTP/DHCP, Request from fa:16:3e:6b:41:3d, length 300

13:17:28.961321 IP 192.168.99.3.bootps > 192.168.99.5.bootpc: BOOTP/DHCP, Reply, length 355

13:17:56.241156 IP 192.168.99.5.bootpc > 192.168.99.3.bootps: BOOTP/DHCP, Request from fa:16:3e:6b:41:3d, length 300

13:17:56.249899 IP 192.168.99.3.bootps > 192.168.99.5.bootpc: BOOTP/DHCP, Reply, length 355
      Note
      This verification is not possible with cirros VMs. The cirros udhcpc implementation does not respond to DHCP option 58 (T1). Try this verification on a port that belongs to a full Linux VM. Red Hat recommends that you check all the different operating systems represented in your workloads, such as variants of Windows and Linux distributions.
  3. If any VM instances were not updated to reflect the change to the T1 parameter of DHCP, reboot them.
  4. Lower the MTU of the pre-migration VXLAN networks: 
    $ ovn_migration.sh reduce-mtu | sudo tee -a /var/log/ovn_migration_output.txt

    This step reduces the MTU network by network and tags the completed network with adapted_mtu. The tool acts only on VXLAN networks. This step will not change any values if your deployment has only VLAN project networks.

  5. If you have any instances with static IP assignment on VXLAN project networks, manually reduce the instance MTU by 8 bytes.

    For example, if the VXLAN-based MTU was 1450, change it to 1442.

    Note
    Perform this step only if you have manually provided static IP assignments and MTU settings on VXLAN project networks. By default, DHCP provides the IP assignment and MTU settings.

What to do next

Proceed to Migrate from OVS to OVN.