Software Defined Networking - Computerphile

Software Defined Networking takes the control away from basic protocols and gives it to the programmers. Cutting edge companies like Google use it to manage their global Network. Dr Richard Mortier of the University of Cambridge Computer Laboratory explains.

This video was filmed and edited by Sean Riley.

[steveowen]

A very intelligent man, who finds a way to simplify a very complex concept, so that someone very unintelligent (me), can understand it.

Thank you sir.

[MarkCidade]

I love the 3D visualization of the network

[deeef]

Great into to the world of SDN. There's definitely enough material to do a mini series about this (SDN protocols like OpenFlow, POF, P4, issues related to centralization, security issues, applications to cloud computing, future internet protocols, alternatives like Contrail, NFV, and etc.). Judging from the comments I've seen so far people are itching for these to be discussed.

[paulzapodeanu]

As this guy briefly mentioned, with the STP example, each bridge is essentially operating independently. A network isn't a "thing" it's a loose collection of things cooperating, and in the good days giving the illusion of order. The main reason that a network runs protocols like STP or OSPF is that you can't possibly have a centralized controller, and each device must be able to operate relatively independently.

What happens if the controller breaks? Make it redundant? Fine, what happens when the controllers lose connection between them and each of them thinks the peer is dead and assumes master role? You have a drawing there where the controller connects to the switches, that in itself is a network. How does that work? Is it a different network, or is it overlayed on top of the regular network? What happens when it fails? What happens when a centrally controlled network is split into multiple islands? Do they islands still operate independently? If they have connection to the controller maybe, but what if they don't? What happens when a device that has lost connection to the controller tries to collaborate with devices that are still connected to the controller (or does it give up trying to do anything altogether, that seems lackadaisical but it may be best)? I admit that in a datacenter, it would be possible to build a simple, reliable, robust, somewhat redundant network that would connect the controller to the network equipment, and then you would have simplified network by orders of magnitude, however, how do you do this if you are a service provider? Do you operate another control network to manage your data transfer network or do you overlay your control network over your data transfer network, in which case, what happens when something breaks, losing data transfer capability is one thing but loosing control over your network sounds much worse? In a datacenter it may be that the environment is stable and failures relatively rare, but in a network that spans a continent or more, failures will be a common thing. You can not try to avoid them, you must design a network that can operate despite these failures.

The most simple tasks, become unbelievingly complicated when they can not be solved by one device but rather a loose collection of devices cooperating, and each of them can loose connection, or malfunction in countless ways, not to mention wilful attempts to disrupt a networks operation.

Any one given device can fail, and it will not be the end of the world. If, let's say, there is a hardware failure of a memory module in a server, it is acceptable that that server will find itself in an irresolvable state of inconsistency, in which case it should stop, and possibly try to start again and if following the POST it is deemed that the server can safely operate with the remaining function components, then it should continue to do so. But what if a network controller fails? Then the entire network needs to be rebooted? That's unfathomable even in a datacenter or office network, let alone an ISP network or the entire internet? You probably plan to operate multiple controllers, but how do you ensure accurate and timely takeover of the role of a failed controller? And what do you do when the controllers disagree? Which of them is "right"? Do you run an odd number of controllers and then have them vote democratically? But who is the arbiter of that election? By doing this haven't you just shifted the problem to the arbiter? What happens when this fails?

This distributed nature of networking makes it such a difficult and interesting subject, if it were possibly to centrally control it would be very simple, but I can't for the life of me imagine how you can do that (though the fact that SDN exists means that this is possible and that I'm just lacking in imagination).

Any of the above questions I would like to see answered in future computerphile videos!

[EfroimRosenberg]

YouTube already updated the 301 issue, views are now more in real time than before.

[prof.angelinagokhale]

Thank you very much @Dr. Richard Mortier for that wonderful explanation and a big thank you @Sean Riley for filming all the Computerphile videos. I always recommend them to my students :D

[maishamanarat]

I love Computerphiles explanations. Easy to understand and put onto real life situations.

[MiPh]

I don't understand why people don't yell out "zeroth" on Computerphile in the comments!

[fablungo]

The main advantage I can see for this is stateful load balancing. Even if just implemented within a datacenter (as opposed to across the whole internet) this could be really advantageous.

[andrii]

Thanks for an amazing quality, love the vids with Dr. Mortier

[stoneshou]

Question from 2020, what is considered the current status of SDN? Is a fair amount of public internet traffic running on SDN routers and switches now?

[mohanadhani]

What are the topics or problems in the networks that the researcher can work on in graduate studies?

[JimCullen]

Do the switches typically get connected _directly_ to the controller, as was shown in the diagram, or would they be connected through the network, and only one or two of the switches would be directly connected?

[BlankBrain]

Nice! Now the large corporate ISPs can completely control content!

[NikiHerl]

I recently started re-reading ASOIAF. My body is so so ready for The Winds of Winter ^^

[Nabel-Network-Solutions]

I want to start SDN. I want to know how to start and what is the first course I should study??
Plz guide me 🤝🌹

[ConnorKenway]

Can someone tell me the names of the books on his shelf in the background?

[samlaf]

So can we see the traditional tcp-ip routing mechanism as a specific case of openflow where the SDN algorithm implemented is spanning-tree?

[JonMajorCCIE]

Last I heard Google was using SDN for their B4 network. I don't believe YouTube rides atop that. I'm on the fence, as a lot of network engineers are with SDN. I feel like it may have better PR among Engineers if it didn't set out to immediately try and replace us lol. "SDN, giving the network to programmers!" Which, is a hilariously bad idea. Makes about as much sense as giving devops to the network guys imo. Cisco's play with ACI isn't too shabby, heck in their 2.0 release they brought in some NX-OS like configuration modes. Which is the right move to me.

[tscoffey]

Is the controller not then a single point of failure?