Q: Why does OVN use STT and Geneve instead of VLANs or VXLAN (or GRE)?

A: OVN implements a fairly sophisticated packet processing pipeline in “logical datapaths” that can implement switching or routing functionality. A logical datapath has an ingress pipeline and an egress pipeline, and each of these pipelines can include logic based on packet fields as well as packet metadata such as the logical ingress and egress ports (the latter only in the egress pipeline).

The processing for a logical datapath can be split across hypervisors. In particular, when a logical ingress pipeline executes an “output” action, OVN passes the packet to the egress pipeline on the hypervisor (or, in the case of output to a logical multicast group, hypervisors) on which the logical egress port is located. If this hypervisor is not the same as the ingress hypervisor, then the packet has to be transmitted across a physical network.

This situation is where tunneling comes in. To send the packet to another hypervisor, OVN encapsulates it with a tunnel protocol and sends the encapsulated packet across the physical network. When the remote hypervisor receives the tunnel packet, it decapsulates it and passes it through the logical egress pipeline. To do so, it also needs the metadata, that is, the logical ingress and egress ports.

Thus, to implement OVN logical packet processing, at least the following metadata must pass across the physical network:

  • Logical datapath ID, a 24-bit identifier. In Geneve, OVN uses the VNI to hold the logical datapath ID; in STT, OVN uses 24 bits of STT’s 64-bit context ID.
  • Logical ingress port, a 15-bit identifier. In Geneve, OVN uses an option to hold the logical ingress port; in STT, 15 bits of the context ID.
  • Logical egress port, a 16-bit identifier. In Geneve, OVN uses an option to hold the logical egress port; in STT, 16 bits of the context ID.

See ovn-architecture(7), under “Tunnel Encapsulations”, for details.

Together, these metadata require 24 + 15 + 16 = 55 bits. GRE provides 32 bits, VXLAN provides 24, and VLAN only provides 12. Most notably, if logical egress pipelines do not match on the logical ingress port, thereby restricting the class of ACLs available to users, then this eliminates 15 bits, bringing the requirement down to 40 bits. At this point, one can choose to limit the size of the OVN logical network in various ways, e.g.:

  • 16 bits of logical datapaths + 16 bits of logical egress ports. This combination fits within a 32-bit GRE tunnel key.
  • 12 bits of logical datapaths + 12 bits of logical egress ports. This combination fits within a 24-bit VXLAN VNI.
  • It’s difficult to identify an acceptable compromise for a VLAN-based deployment.

These compromises wouldn’t suit every site, since some deployments may need to allocate more bits to the datapath or egress port identifiers.

As a side note, OVN does support VXLAN for use with ASIC-based top of rack switches, using ovn-controller-vtep(8) and the OVSDB VTEP schema described in vtep(5), but this limits the features available from OVN to the subset available from the VTEP schema.