MTU Considerations

Problem

The cluster has multiple network layers, each adding encapsulation overhead. VXLAN adds 50 bytes, GRE adds 28 bytes, OVN adds its own headers. If the MTU is set to the standard 1500 on all interfaces, encapsulated packets exceed the physical link MTU, causing fragmentation, dropped packets, or silent TCP retransmissions that cripple performance.

Where MTU Matters

Interface	MTU	Why
eno1	1500	Physical NIC, public internet — standard
eno1.4000	1400	VLAN for inter-node traffic — headroom for VXLAN if needed
br-ex	1500	OpenStack external bridge — VMs connect here
OVN overlay	1400	Neutron global_physnet_mtu: 1400, path_mtu: 1400
VM interfaces	~1400	Inherited from Neutron subnet MTU
Netbird WG	1280	WireGuard default — conservative for cross-network paths

Configuration

From argus/infrastructure/yaook/neutron.yaml:

neutronConfig:
  DEFAULT:
    global_physnet_mtu: 1400
    advertise_mtu: True
    enable_tcp_mss_clamping: True
  neutronML2Config:
    ml2:
      path_mtu: 1400
      physical_network_mtus: enp3s0:1400

From hephaestus/nixosModules/vlanConfiguration.nix — the VLAN sub-interface:

interfaces."eno1.4000" = {
  mtu = 1400;
};

advertise_mtu: True tells Neutron to set the MTU option in DHCP responses, so VMs automatically configure their interface MTU. enable_tcp_mss_clamping: True adjusts TCP MSS to fit within the reduced MTU, preventing black-hole fragmentation where SYN packets succeed but data transfers fail.

The default OpenStack network (non-external) uses MTU 1342 — slightly lower to leave room for additional VXLAN encapsulation within the overlay.