4.7.1 Design Considerations
The ESXi hosts’ VTEP network interfaces must be configured for at least 1,600 MTU through the network switches. This is to allow for the extra header bits that are applied to the packet size by VXLAN.
The concept of VXLAN replication to address scalability using Unicast Tunnel EndPoints (UTEP) and/or Multicast Tunnel EndPoints (MTEP) is documented in the design guide. However, consider the following aspects for a design implementation:
• Layer 2 deployments (L2 topology has no boundary for selecting UTEP and MTEP, because all VTEPs are on the same subnet)
o Small deployment – Unicast mode is a recommended configuration.
o Large deployment – Hybrid mode is more suited because the UTEP function cannot identify a VTEP boundary (VTEP are on same subnet) to provide efficient BUM replication per logical switch and thus scales very well.
• Layer 3 deployments
o In most cases, unicast deployment works because Layer 3 topology provides a VTEP IP addressing boundary, and therefore, UTEP efficiently replicates frame pre LS. No need for PIM because traffic is all unicast.
o For very large deployments, hybrid mode is recommended providing MTEP-based BUM replication as well eliminating configuration of L3 multicast (PIM).
• Hybrid mode deployments:
o VMware recommends configuring the external physical switch with IGMP querier along with IGMP snooping (this is an industry standard best practice for most switches, including Cisco, Arista, Dell, and Brocade).
o If you accidentally forget to configure IGMP querier in the physical switch, as long as IGMP snooping is defined, the hypervisor will send a join to the configured multicast address.
o Hybrid mode is preferred when large L2 multicast traffic from VMs requires replication.