Monthly Archive: October 2008

And one more demo…

OpenFlow Demo at the GEC3

OpenFlow Demo at the GEC3

We just finished the OpenFlow Demo at the GENI Engineering Conference, and it was amazing. We showed our new OpenFlow protocol running on switches from Cisco, Juniper, HP and NEC. Our experimental network stretched half way around the globe from Stanford to Tokyo via New York. It used fibers fromĀ  Internet2, CalRen and JGN2plus.

Over this network we showed how we can move around a running game server from one physical host to another without the game even getting interrupted. We demonstrated how you can route a network connection with a simple drag and drop interface (e.g. a TCP flow inside Stanford going via Tokyo and Houston). We even sent a running game server to Tokyo from Stanford, without losing the connection.

Press coverage of the demo included articles English, Japanese, Swedish and Spanish. The OpenFlow web site recieved a few thousand hits, with visitors from every major company in the networking space. All this was made possible by about 40 people from Stanford, Internet2, Cisco, Juniper, HP and NEC had been working on this for months.

As a result of this, OpenFlow is building momentum. NEC announced during the conference support for OpenFlow in their product, and more announcements will follow. By mid next year we are hoping to have pilot deployments at 6-10 universities, and I would hope we will see commercial deployments in that time frame as well. All in all a huge step forward for OpenFlow.

Best Paper Award for Neda, Yashar, McKeown et. al

Congratulations to Neda, Yashar, Monia, Nick and Geoff for their best paper award at the Internet Measurement Conference. Their paper Experimental Study of Router Buffer Sizing tests out recent results onĀ  buffer requirements of high-speed routers that serve highly aggregated traffic. Amongst other things it verifies the C/sqrt(n) result from my thesis as well as my former office mate Yashar Ganjali’s work on very small buffers and find that they hold well.

It is great to see this work getting recognized, but what is even more encouraging is that two router vendors privately confirmed to me that the next generation of some of their products will have substantially smaller buffers. This not only reduces power consumption, but also means that we are less likely to see latency spike whenever peering points or core links are congested.


The project that currently takes up the majority of my time at Stanford is OpenFlow. OpenFlow is a new protocol that we specified and that vendors are now adding to their routers and switches. What OpenFlow allows you to do is remotely control the behavior of a switch from an controller software that runs on a standard server. This has two major advantages:

  1. You can now write your own control software and try out new switch functionality at full line rates. In the past this has been difficult as all major router and switch vendors lack APIs and are typically closed platforms.
  2. If you use a centralized controller now has a unified view of the network. For some applications such as mobility management, virtualized data canters or security this allows you to do things that previously would have been very difficult or impossible.