2

In all spine leaf architecture explainers I've seen I've come across the concept of the border leaf which is used as the connection point of the entire clos network to an upstream router, be it either a WAN or core router. What confuses me is why they would only ever use one leaf (or one leaf pair, in case of l2 paired leafs using MLAGs) for the uplink. If connectivity permits, shouldn't it be possible to connect multiple (paired) leafs to the same upstream router, each having their own unique IP and redistributing the same subnets? ECMP and the like should then take care of optimising traffic flow, right?

What would be the down sides of having multiple/all leafs being "border leafs" to the same upstream router?

Example spine leaf architecture with border leaf: enter image description here

3
  • Since you refer to a clos network, you seem to be talking about a multi-tier architecture, perhaps even the traditional 3-tier hierarchy. Please be more specific about your scenario. I think a diagram would be a great help.
    – Zac67
    Apr 27, 2023 at 9:34
  • I'm not referring to a traditional 3-tier architecture but a spine leaf architecture: packetcoders.io/the-history-of-spine-and-leaf-architectures
    – alex
    Apr 27, 2023 at 10:23
  • 1
    It's unusual that the border connections would require the same bandwidth as the core. If you make every leaf a border, you're just pushing all the contention to the upstream device.
    – Ron Trunk
    Apr 27, 2023 at 13:18

1 Answer 1

3

Typically, connections to the WAN/Internet are done by routers. Connecting every leaf in your network to a set of routers would require routers with LOTS of ports, and the price per port on routers is usually a lot higher than on switches.

Also, it would introduce a lot more complexity, since every router would have BGP connections with every leaf. This greatly increases the number of BGP sessions and routes in your routing tables.

If your spine/leaf setup consists of tens or hundreds of leafs (which is not that uncommon), this really blows up fast, not just in ports and costs, but also in things like cabling.

Usually, you pick a few leafs on specific characteristics (location, traffic volume, etc) and connect those to the WAN/Internet.

2
  • Scale makes a lot of sense, thank you. Two follow-up questions: is there anything preventing one from using two or three independent (non l2-paired) leafs to connect to the same router? Would I have to take care of route prioritisation myself or could I make use of features like ECMP?
    – alex
    Apr 28, 2023 at 8:25
  • I don't think there's one generic answer to those questions since it all depends on your exact setup, but if it's just dynamic routing (like BGP), picking any leafs would work, just like ECMP could work. We connect four leafs on two MLAG pairs with ECMP enabled on an Arista spine/leaf setup with Juniper core routers.
    – Teun Vink
    Apr 28, 2023 at 8:47

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.