02 Sep 2004

SIGCOMM Day 1 and Day 2

(note: i'm going to use proper old-english punctuation just because i feel so charitable today)

On Day 1, there were three sessions plus the keynote. The keynote was by Simon Lam who seems to be well respected at SIGCOMM, apparently involved with working on the ALOHA protocol. He talked about some stuff that Sandy Fraser has been devising for a while, which is putting flow labels and making the network layer aware of flows. Of course, this is very controversial in the field at SIGCOMM where everyone loves IP. He panders to those people by saying that this should be built into IP. IP should be aware of flows but be able to assign QoS to it.

As for the sessions, I only attended the first one, which was Network Geometry. This session was quite interesting, the papers aren't as complex as they sound. The most interesting one is the MIT paper on assigning coordinates to nodes based on RTT. The cool thing here is that they have developed algorithms for many nodes to converge to a stable state, and then figured out how to converge quickly and accurately. So for instance, they make progressively smaller steps when they know they are closer to the stable state. The other cool thing they do is to use something called "height vectors" to get around the fact that they can't always plot an optimally stable graph of these nodes on a 2D plane.

The rest of the day we skipped and walked around Washington Park. Will talk about that later.

On Day 2, there were three sessions and a poster session. Nothing really memorable from the talks here. The first session was on wireless networks, and that was the most relevent to what I know. The first paper on some link-level measurements of the RoofNet where they find out (OMG SUPRISE!) that multi-path is the greatest contribution to packet loss in 802.11b. The other papers talk about some new hot topic among networking, which I think is useless, called Delay Tolerant Networks. These are like store-and-forward networks where you oppurtunistically hope your packets get sent by people who may or may not be able to forward your packet. The last paper in this session was especially funny, about using the postal service (by Randy Wang from Princeton) as a high bandwidth high latency network. The things he propose isn't too much different from NetFlix after the intiial bootstrap. New ideas include sending people encrypted volumes of data, and then getting them to use a lower latency low bandwidth link to get the key to decrypt the portion they are entitled to.

The second session was about network security. However, most of the issues were to do with BGP security, which I have no idea about. The last paper in this session was presented by Microsoft Research about "shield." Think of it as anti-virus for network packets. So instead of sending out virus definitions, they propose sending out exploit definitions to act as a stop-gap before a vendor can release a patch. Thats all that I could see they were proposing.

The last session was about network troubleshooting. Also very hard to tune into that because all of it would only be relevant to ISPs and backbone operatos who manage networks.

You can reply to me about this on Twitter: