Inventor of OpenFlow SDN Admits Most SDN Today Is Hype

darthcamaro writes "Every networking vendor today is talking about Software Defined Networking (SDN). The basic idea is that the control of the underlying networking hardware is abstracted by software. Martin Casado helped to come up with the whole topic with his 2005 Stanford thesis (PDF). Eight years later after selling his startup Nicira to VMware for $1.2 Billion, Casado sees the term SDN meaning everything and nothing to all people. From the article: '"I actually don't know what SDN means anymore, to be honest," Casado said. Casado noted that the term SDN was coined in 2009 and at the time it did mean something fairly specific. "Now it is just being used as a general term for networking, like all networking is SDN," Casado said. "SDN is now just an umbrella term for, cool stuff in networking."'"

Comments

Riding the hype train

[GeneralTurgidson]

I need to build a business around some new buzzword and sell it to VMware. Cloud and everything related to it has really stagnated development of other areas of IT in my opinion. Companies try and figure out WTF SDN is or how to integrate their networking stack with AWS instead of focusing on what's really happening in the IT world.

"Software defined computing"

[Anonymous Coward]

It's the way of the future.

So we aren't going to be able to replace...

[John Hasler]

...all the fiber optic cables with software? We aren't going to move everything to the cloud, including the cloud?

Re:So we aren't going to be able to replace...

[fuzzyfuzzyfungus]

Certainly not all of them; but I'm pretty sure that the box they are all plugged in to is, pretty much, using a software layer to abstract the ugly details of dumping traffic between them over a really, really, fat internal bus of some flavor.

And, in many cases, a single fiber is(thanks to software) being sliced up into a bunch of little VLANs to create a logical topology that (while it is ultimately constrained by the physical one) is substantially different than the physical topology, especially once you count aggregated port groups, redundant links, and so on.

'SDN' doesn't mean jack in part because everything except your 20 year old 10Mb hub is already doing some amount of software trickery(even dumb switches keep track of which MAC(s) are on which port, and anything with 'managed' in the title can do quite a bit more), with varying levels of ASIC vs. general-purpose-CPU and varying levels of correlation between the logical topology and the physical topology.

There just isn't a nice bright line(at least in terms of real-world use cases, obviously a VM chattering to itself over a loopback interface is 'software' and a passive ethernet tap is 'hardware') between what is 'software defined' and what isn't. They all obviously depend on hardware to execute the software; but the amount of additional logical complexity slopes up surprisingly smoothly.

Re:So we aren't going to be able to replace...

[istartedi]

We aren't going to move everything to the cloud, including the cloud?

Sure we can. It's all based on something I call the "metacircular evaluator". My consultancy and I can install MEs in all your software systems so that you can move the software into the software, and define your business in terms of your business. "My god!" the tech business reporter exclaimed, "this is the most revolutionary thing I've ever heard, tell me more about how these MEs work".

Well, you just have to re-write everything in Lis--. And then, before I could even finish, the room was empty.

Re:So we aren't going to be able to replace...

[Lennie]

SDN in practise just means, networking things (private networks, VPNs, loadbalancers, etc.) have an API so they can be automated.

So when you need to scale out, because your website has more visitors during the day then you don't just get new VMs but those VMs also get connected to the right networks or extra load balancers gets added as well.

The software in software defined networking, is the application specific software. That application can be that website as mentioned above or something completely different.

For example Google uses their self-developed software to reserve bandwidth for their different applications and data-replication jobs and handle link failover on the WAN-links between their datacenters.

Because they used OpenFlow their were able to save money on their WAN-links because they get better utilization than traditional methods. They have normal Google servers that 'directly' configure the forwarding tables.

pffffFFF FPGA HA HA HA!

[VortexCortex]

I'm glad he's laughing all the way to the bank. Gives me room for my new buzz-word compliant technology: Hardware Optimized for Software Systems (HOSS)

Shhhh, it's just ASIC in sheep's clothing.

Re:Now that he's cashed in...

[girlinatrainingbra]

SDN is the hot new buzzword, just like "cloud" computing has been for the last few years. Buzzwords fly by. I agree with you that he can afford to be honest, but he's not just being honest, he's pointing out the "fuzziness" of what the term "SDN" is being applied to.
.
I believe he had a fixed and set definition which he must have specified with some detail in his thesis (isn't that what you're supposed to do in a thesis? be specific?), but nowadays everyone and anyone is calling any configurability of the top or higher levels of networking as "Software defined networking".

Re:dunno

[postbigbang]

You mean "infecting" servers with router VM appliances that smoke government blocks by creating backdoor VPNs, proxies, shadow VLANs and stuff?

Never happens.

Virtual circuit network

[Animats]

OpenFlow is basically a way to turn a packet network into a rather dumb virtual circuit network. It works something like Tymnet, circa 1971. In Tymnet, all the virtual circuits were set up by a "supervisor" computer, which told each node where each flow was supposed to be forwarded. The supervisor also handled authentication, but data packets didn't have to pass through the supervisor once the connection was set up. That's what OpenFlow does, mostly. The first packet of each new "flow" (IP/port/IP/port set, usually) is sent to Master Control, which decides whether that flow will be allowed. Master Control can also choose to monitor the flow. The implications are obvious.

DOCSIS 3, the cable modem traffic control architecture, can potentially do most of the same things, and offers better control over bandwidth. DOCSIS 3 tends to be run more to control users than to maximize throughput, but that's a marketing issue. (If your cable connection is throttling something, the commands to do it were probably sent to a DOCSIS node.) There's good QoS and fair queuing stuff in DOCSIS 3, but it's not always used intelligently. DOCSIS is less intrusive than OpenFlow; the nodes are sent rules to enforce, but there's no need to get permission of Master Control for every new flow.

The rest of "software defined networking" seems to involve adding another layer of indirection to Ethernet addresses so they can be moved around within the data center. ("There is no problem in computer science that cannot be solved by adding another layer of indirection.") That's a reasonable network management tool, but it's not exactly a profound concept.

Re:dunno

[ppanon]

From what I read of SDN, the idea is to have centralized routing (presumably for use within a data centre, telco, or high-performance campus network) instead of decentralized routing. Instead of having each individual node recalculate routes using tree-based routing algorithms like OSPF, a central node with a holistic view of the network recalculates and redistributes routes using algorithms that allow more fine grained slicing of packet flows for closer to optimal load balancing and congestion management.Unless you're a telco, a co-lo, or have a datacentre with >5 racks steadily generating >50Gb/s of network I/O per rack and needing high availability, it's doubtful that you need to pay the premium for it.

Re:Software Defined Networking

[ppanon]

In theory, the centralized routing agent, by having a global view of the network, can both optimize load across links and adjust to congestion patterns better than a distributed network infrastructure can. So for telcos and other colos, that can mean better uptime and more efficient use of expensive high-end infrastructure. Just as a centralized traffic light control system for a city can identify congestion hot spots and adjust light timing to reduce congestion better than isolated traffic lights can.

For instance, if nets A, B, and C are linked in a loop for redundancy, spikes of traffic between A and C could be load balanced by shifting certain source-destination pairs over the link to B to avoid congestion, even though it would normally be faster to go over the A-C link and that's what an OSPF router would do. Alternatively, in the longer term, a telco could make more effective use of fibre if it could temporarily re-allocate fiber currently not used for voice to deal with a spike in data traffic.