Scott Shenker on OpenFlow and SDN

Brent Salisbury sent me a link to a fantastic OpenFlow/SDN presentation Scott Shenker did @ Stanford University a few days ago. It’s a perfect introduction to the fundamental ideas behind SDN and therefore a must-see for everyone vaguely involved in networking.

Here are some of the highlights (from my highly biased perspective):

  • They got it all wrong in the first iteration (@ 20:00) – it’s rare to see someone being so honest about past misconceptions;
  • Network virtualization is the killer app, and SDN is just a means to an end (@ 25:00);
  • MPLS was right (@ 32:00) – would love to see the reaction of a well-known SDN evangelist and MPLS basher ;)
  • A network should have a complex edge and a simple core, with software switching at the edge (@ 38:00);
  • Today’s networks are full of middleboxes which are already x86-based. The functionality of these middleboxes should be moved to x86-based devices at the network edge (@ 42:00);
  • Network latency (actually intra-switch latency) doesn’t matter at the edge (@ 49:00).

There’s still a lot of handwaving and missing details, particularly as you move from tightly controlled environments (data centers) to large-scale WAN networks where individual components have to operate independently to be failure-resilient, but they’ll eventually get there.

This would be a perfect place to smugly list my blog posts (going all the way back to 2011) talking about exactly these same things ... but it’s sunny outside and the kids are waiting ;)

8 comments:

  1. "They got it all wrong in the first iteration (@ 20:00) – it’s rare to see someone being so honest about past misconceptions;"

    He actually said they got it mostly right, but a few things wrong. Isn't it? ;-)

    Sadiq
  2. What's not clear from Scott's presentation is whether he believes there is still room for Openflow in his "new" SDN world. Your views?
    Replies
    1. In a world with centralized controller and distributed forwarding hardware (or software) - not that I would unreservedly agree with this picture - you need a protocol that downloads the forwarding information into the forwarding devices and pass unknown traffic to the controller. OpenFlow is currently one of the best (if not the only) options.
  3. Hi Ivan, thanks for linking this, I'm glad to have seen the presentation, it's so refreshing to hear straight talk from someone close to the source. Cheers :)
  4. Finally, someone with authority that mentions all the things I've had on my mind for as long as I've been looking at SDN.
  5. So the more I think about this, MPLS is right, but will it be used ? Because there might not be enough hardware deployed that supports it or people with experience deploying it.

    I think Network Virtualization is more similar to Server Virtualization than people realize, with Server Virtualization people didn't buy new hardware at first. And Network Virtualization overlays also means we don't need to buy new hardware.

    Overlays might not be the definitive form, but that is where people will start. It's just makes economic sense.
  6. I would like to add that PfR (OER) has been in IOS for more than 6 years and it's interesting that people do not even know about it :)

    > Do you think that PfR is not SDN?
  7. Scott's presentation is nothing but what Nicira wants to do and they are having hard time convincing customers, plus he is just trying to promote it.
Add comment
Sidebar