Leaf & Spine Architectures

[matgar]

Why would the ip of a VTEP exist in 2 different leafs?

Both leafs are effectively acting as a single logical unit - think of them as two ToRs in the same rack.  To get to the same resources (servers, etc.), you could go to either leaf - why treat them as two discrete entities when they both do the same job?

My understanding/way of thinking is that the VTEP is part of the "invisible" network infrastructure where simplicity is wanted and that its within the VXLANS that you might want to do configure anycast or server load-balancing etc.

Also ECMP in a spine/leaf setup is as far as I know supposed to happen in the leaf.
Ie LEAF3 has 2 paths to LEAF2 via either SPINE1 or SPINE2.

Doesn't ECMP happen anywhere there is more than one route with equal costs to the same destination?  So if the spine needed to reach a host, and there were two equal-cost paths to reach it via LEAF1 or LEAF2, is that not ECMP?  I'm not being snarky here I'm seriously asking - I've been known for being dumb before and if I'm misunderstanding the definition of ECMP I'd like to alleviate that.

EDIT - Oh, and welcome to the forum.

True ECMP happens wherever there is more than one equal cost route. It was less a comment of what ECMP is, but rather where its expected to happen in your design.
My meaning was that for predictability (troubleshooting/understanding) purposes if nothing else, it would be better if each spine only had one route to the destination.
All I've read of spine/leaf architecture has been with the design that the multiple routes are in the leafs and not in the spines. (well with exception to your post that is.)
New ways to do things isn't necessarily bad but in this case you seem to have ended up with unexpected/unwanted behavior.

Edit: And thanks for the welcome.

[wintermute000]

Sorry for the spam...

Winter - if you do find out if there was anything in particular that sealed the deal for Cisco, I'd love to know.  I probably suspect it was the typical pants dropping on price though (I wonder how long they can keep doing that, or how many times they can pull it off before customers realize it's only a one-time price, and when they go to renew, prepare to get YOUR pants dropped). 

I know it may happen once in a while due to some feature, but I don't remember ever losing for technical reasons.

Related, found this hilarious comments thread, some serious insider (allegedly) b1tching. If TLDR go near the bottom when they start talking about CAP - Cisco's Customer Assurance Program - and the alleged massive boondoggle of the symantec ACI deployment, leading from/into general slagging off of the Insieme BU and the lack of market readiness of ACI.

http://www.bradreese.com/blog/10-4-2015.htm#COMMENT



Having seen two live ACI deployments (small ones by US standards too) and having heard their horror stories, I'm not really that surprised and do lean towards the b1tching being towards the 'truth' side of the equation

[burnyd]

OK OK I get it. When you are talking MLAG, the arista switches are NOT in a traditional stack, its some kind of Nexus vPC type feature correct? (i.e. the switches are still separate entities and separate control planes but can present a shared etherchannel to a downstream host)


Also, what is HER?

Head end replication.

MLAG is like vpc but it actually works.

[wintermute000]

out of curiosity, what about vPC do you find not working? (aside from bugs... QA is down the toilet these days at the big C.... as well as the idiosyncracies of routing over a vPC or not)

I know Juniper has a similar feature as well, can't recall what its called.

In my mind MLAG = multi chassis etherchannel = a stack or VSS = single control plane but I guess its just terminology

[burnyd]

out of curiosity, what about vPC do you find not working? (aside from bugs... QA is down the toilet these days at the big C.... as well as the idiosyncracies of routing over a vPC or not)

How much time do you have haha.

Bugs are really where the problem lies. 

[NetworkGroover]

I think simplicity is key there.. though it would take time to really dig into the details to adequately answer the question.

From a 10,000 foot view, the simple fact you have to have a second dedicated peer-link (is this still the case?) kinda highlights a complexity problem... and some of the other limitations that I have yet to run into with MLAG - it just works.  I'm sure it sounds bias... but don't take it from me.. ask other folks who've worked with it - I have yet to hear complaints.

[that1guy15]

Oh man did I miss out on this thread while away for Christmas!