VXLAN Broadcast Domain Size Limitations

One of the attendees of my Building Next-Generation Data Center online course tried to figure out whether you can build larger broadcast domains with VXLAN than you could with VLANs. Here’s what he sent me:

I’m trying to understand differences or similarities between VLAN and VXLAN technologies in a view of (*cast) domain limitation.

There’s no difference between the two on the client-facing side. VXLAN is just an encapsulation technology and doesn’t change how bridging works at all (read also part 2 of that story).

The only difference between VLAN-based fabric and VXLAN-based fabric is the core transport. VLAN-based fabric uses STP/MLAG in the fabric core, TRILL/SPB/… based fabrics use their own routing protocols, and VXLAN uses IP routing. Edge flooding and learning behavior remains the same.

I covered the basics of TRILL and SPB (in case anyone is still interested) in Data Center Infrastructure for Networking Engineers webinar. Roger Lapuh did a deeper dive into SPB during his presentation @ Leaf-and-Spine Fabrics webinar. His presentation is accessible with free ipSpace.net subscription.

EVPN is a different story as it’s IP aware, but keep in mind that EVPN became SIP of Networking; every implementation supports a different subset of features.

Several EVPN-based fabrics support ARP proxy at the fabric edge, reducing the number of broadcasts caused by ARPs… assuming someone is not ARPing for a non-existent IP address, in which case you’d probably see those ARPs flooded. Test, test, test… and make sure you also test all possible crazy scenarios.

EVPN-based fabric could implement pure IP transport and turn off flooding altogether, turning what looks like a VLAN into stable routed IP network (admittedly doing routing on host routes). I don’t think any vendor is brave enough to do that.

Yes, we know and understand why we should keep VLAN size limited (let’s say 1K hosts/guests/) but what about VXLAN segment size?

The same limitations apply. Although EPVN-based fabrics (whether using VXLAN, MPLS or GRE) could reduce the amount of ARP traffic, there’s nothing stopping a single host from blasting the network with a gazillion RARPs per second (because why not) and impacting everyone else in the same segment.

Am I right that from business risk perspective I should keep VXLAN domain small as well because someone or something can impact all my 12.000 VM's in one VXLAN? Or is this technology resistant against broken frames/packets, flooding…?

A single flooding domain is a single failure domain. A VXLAN VNI (unless turned into a pure routed solution) is a single flooding domain regardless of what the fabric and microsegmentation vendors are telling you. QED.

Long story short: Bridging doesn’t scale. Keep your failure domains small.

3 comments:

  1. https://packetpushers.net/podcast/weekly-show-408-running-secure-ethernet-fabrics-with-extreme-networks
    God podcast from the “pushers” :-)
    I like when they talk about overly networking vs fabric.
    Replies
    1. Is Vxlan with EVPN better den SPB in Campus?
    2. VXLAN needs just IP transport in the underlay - a well-known and well-tested technology. SPB is newer, and might thus contain more bugs on the transport side. EVPN is newer on the edge side. Choose your poison ;)
Add comment
Sidebar