Traditional data center architectures centralize the routing. All inter-vlan traffic must transit the core, even when hosts share a network access point. It would be more efficient to perform inter-vlan routing at the network edge. One could route at the edge, but there is a requirement to retain a single L2 domain across the core. Are there any data center architectures that deliver distributed L3 forwarding in a single L2 domain?
3 Answers
A lot of the content in this question is very Cisco-ish, let's dial that back and dispel some mythical assumptions.
1. I need to do Layer 2 at the core
Not unless that's all your hardware is capable of; you could quite freely have point-to-point (links/portchannels) between everything in your core and bring up OSPF/ISIS adjacencies, this is arguably superior to a large L2 domain at the core since you're now precluded from inflicting L2 loop prevention upon yourself; topology changes are now going to be handled by your IGP. If you need to multipath, ECMP takes the place of things like GLBP.
2. Traditional DC architecture centralizes routing
Depends what you mean by "Tradition" - IPv4 is "traditionally" classful, modern IPv4 is not. In the same vein, a modern datacenter architecture will make the vast majority of non-access links between equipment pass traffic over L3, because from a security and performance standpoint, you want to mimimize the size of your broadcast domains and avoid reliance on L2 loop prevention which tends to impact traffic flow for significantly longer periods than an L3 protocol when a topology change happens (recent development such as SPB and TRILL not withstanding)
3. Traffic between broadcast/multipoint domains must happen at the Core
Again, this is down to the hardware you have; if your distribution layer is capable of L3 forwarding on par with your core and also happens to be the termination point for the particular domains you wish to route traffic between, there's no logical reason to send that traffic up to the core only to have it sent back down.
So, with all this in mind, why do large, flat L2 networks remain so pervasive? Because it's "simple" and doesn't require too much thought. Setting up multi-area OSPF and ensuring optimum paths throughout the network by configuring link costs appropriately makes for an excellent architecture, but it requires you to think. spinning up a few VLANs, using something like GLBP and letting spanning tree just "do its thing" does not. The real WTF however is that a decent L2 setup should have a well-thought out STP configuration just as OSPF should.
More than likely you'd have to go with one of the newer overlay networks or using Plexxi's switches. I'm sure you could find some way of doing this through traditional means, but I'm not sure how stable it would be.
L3 at TOR is for sure not uncommon. Most larger DCs (Google's, Facebook's) run L3 to the TOR, but they also have phenomenally written applications that don't require VMotion/L2 to recover from failure.
It sounds though that your requirements are more SMB and similar pain points for most virtualized DCs. If it's a greenfield deployment I would strongly suggest looking at Plexxi's offering. If not then an overlay might be the right thing for you.
HP's IRF has some resemblance to what you're talking about, although possibly not in the way you're suggesting. It presents a single system image, but the switching and routing is distributed amongst its nodes. So if routing is enabled and the switch knows from ARP & MAC tables that the destination is on the same switch, it doesn't traverse the IRF stack's backbone. But that's more like having a distributed core than having a two- or three-layer architecture.