mancalablog

posts to days old
of topics
with text
post

link propagation

m (9 Jan 2005 18:43): The Edge asks 120 experts from various fields what they believe to be true but cannot prove.
zong (10 Jan 2005 18:32): The techies make more interesting readings than the fuzzies, who indeed make very fuzzy statements.
m (12 Jan 2005 15:27): BGP routing in the Internet (maybe a little dumbed down, but not too hindered by this -- clearest explanation of BGP I've come across yet). Shocking that (as of may 2003) the global routing tables for the internet contained only about 123000 routes.
zong (17 Jan 2005 18:42): Shocking that for all the hoopla about the Internet being a redundant network, global traffic basically get routed through a handful of large corridors with ISPs hanging off as branches. The topology is more like a tree than a web.

Shocking that you can (or can you?) route traffic your way just by promising to deliver to a certain range.
Paul (18 Jan 2005 18:18): There must be something more to this. If you could reroute arbitrary traffic to yourself, that would be more crippling than any existing DoS attacks.

As for tree-versus-web... as long as there are many (say, >10) paths from A to B, for any points on the "major corridors", that's reasonably redundant... lack of ISP-to-major-corridor redundancy is worse.
m (18 Jan 2005 19:06): I don't think any of the routers that you personally have a chance of talking to will be running BGP. Ones that do afaik have to be explicitly configured to use your IP as a BGP neighbor. I think they also have a bunch of route redistribution filter rules that they use. The explicit neighbor part seems strange. OSPF does automatic neighbor discovery with some "hello" messages, but BGP just doesn't
zong (22 Jan 2005 1:01): BGP was briefly mentioned yesteraday in a seminar on network utility maximization (with applications to congestion control protocols). BGP was described as a "big mess."
m (22 Jan 2005 14:16): congestion control at the mac layer (like, with aloha and macaw for wireless), or at the transport layer (with TCP Tahoe/Reno/Vegas)?
m (22 Jan 2005 14:18): I think i listed every congestion control protocol I know.
oh - CSMA/CD for media access does the same thing as aloha, I think, with exponential backoff.
zong (22 Jan 2005 15:27): Network utility maximization can be used to analyze both and to guide designs of such protocols.
According to the talk, all three TCP congestion protocols are implicity solving a standard network utility maximization problem over the network, in the sense that, they will stabilize at the optimal solution that maximizes sum of individual utilities over a network with fixed link capacities, where utility is a concave function of the transmission rate. Two of them maximize according to a logarithmic utility function, the other one maximizes according to an arctangent utility function. None of them were ever designed to do this.

If you mix and match protocols, though, you may have more than one equilibrium. But, with some contraints on the utility functions and whatnot, the number of equilibria is almost guaranteed to be odd and probably there is a unique one.

Read more about reverse engineering TCP congestion protocols here.

Read about NUM and optimization in general here.
R (30 Jun 2016 19:49): Ambiguous Cylinder Illusion
m (1 Jul 2016 7:43): CONFUSING
m (1 Jul 2016 7:45): Some clarification?