Last week I wrote a post about my thoughts after an informal talk that Radia Perlman gave at Brocade. I had a bunch of thoughts during and immediately after she spoke, but the more interesting questions came several hours afterwards. Bear with me as I share these questions and (in honor of unconfirmed or semi-hypothetical things) speculate a bit.
First, for just a hint of context, recall that Dr. Perlman invented Spanning Tree Protocol (STP) in 1985 and more recently has helped to establish the TRILL standard. With that background, is TRILL just the “Son of STP” to be followed in 20 years by the “Son of TRILL”? Um, no.
You see, Radia was never a fan of STP. She was working at DEC in Layer 3 protocols (as a part of DECnet) and felt strongly that “Ethernets” should not be connected together by extending Layer 2, but instead devices in one L2 network should speak to devices in another L2 network by routing between these networks using Layer 3. And that might have worked fantastically well… if only all the applications running on Layer 2 networks had been written using Layer 3 (like good little OSI citizens).
But alas, back in those days there were many apps written right on top of Layer 2, and so routing would not meet the need. Nowadays, though, everything is written with TCP/IP, right? So, that leads to the first question I’d like to ask Radia.
Q: If years ago you thought we should use L3 routing to connect our tiny shared-media networks, why are you supporting TRILL now when the number of apps sitting on Layer 2 is negligible? In other words, why not route everything?
(Well, a good journalist would attempt to contact the person in question, but that would spoil my speculative fun. And in a way this is more authentic, because this way you’ll know what was in my head and I can’t disown it after Radia chimes in with some truly wise and obvious response. And actually you can see her 2004 answer here, but the game is slightly different today. And so here goes my first order speculation.)
Let’s guess that her answer would go something like this: “Sure, most, if not all, apps are now written on TCP/IP, but many apps still demand large L2 networks for mobility or require their peers in the same subnet, which rules out routing. These subnets need to be huge and Spanning Tree doesn’t cut it. Hence TRILL.”
Okay. Gotcha. Makes sense. At least, I can see what problem is being targeted. But others are targeting that a different way with these new encapsulation ideas that allow an architect to spread a logical (or, if you prefer, virtual) Layer 2 network anywhere. Who needs jumbo physical L2 in such world?
I could speculate about Radia’s answer to this, but you might think me a wimp if I continue to duck ownership my ideas. That said, I’ll try to be a tad more macho and acknowledge this is my own answer:
First, let’s pause for a moment and notice that those encapsulation schemes are being offered up by server virtualization folks. Sure, those encapsulation ideas are interesting for the up-and-coming new crowd of hyperscale multi-tenant virtualized data centers where every server has a built-in virtual switch, and an orchestration scheme to support them all. “But wait,” I hear you say, “that’s all fine for hypervisor to hypervisor communication; but to be useful, virtualized applications need to communicate with things outside the virtualized world.” Right you are. We would need additional gadgets to bridge from the real world to this tunnel-ridden universe. (And when poking around orchestrated virtual networks, you’re getting into my neck of the woods: Software Defined Networking.)
Stop! Enough of these mythical and phantasmagorical creatures! Most of us mortals spend our time in ordinary data centers, solving tough but honest problems. We don’t need your stinking multi-tenancy and tunneled universes! Please do a simple and frank appraisal of these new networking hobgoblins options!
All right, all right. Since you asked so nicely, I resolve that in the New Year I’ll work on a straightforward and only slightly controversial and sensationalized post or two on Ethernet Fabrics versus OpenFlow! It should be easy to recognize. Just set up a search for my working title: “Clash of the New Network Titans!”
Or you could just subscribe to EthernetFabric.com’s feed.
