bill,
i'm surprised by your remark. i thought you have been around long enough to understand this. we've been over all this before.
the complexity of that computation is driven by two things - the cardinality of the set of visible nodes in the global topology graph, and the complexity of the topology connecting those nodes. (note that "node" here does not imply a router but whole networks).
of these two things, we cannot readily control the edge topology of the graph, so we are only left with controlling the cardinality of the node set of the graph if we wish to influence that complexity.
of course, you can argue that we don't need to care about the problem - that somehow processors will keep getting fast enough fast enough to retain reasonable convergence times.
but then you are betting on a race between two exponentials - and are making the bet that the smaller exponent will win.
i know *i* don't want to wager the future on that bet.
-mo
My argument is that the basic premise that the assumptions wrt the cardinality of the set of visable nodes in the topology graph -may- be wrong. Other than that, I agree with you. Chalk it up to a fit of Noel-Stev syndrome. With a bit of rest, I'm sure it will pass. (and I do recommend "Small Forwarding Tables for Fast Routing Lookup" from SigCom'97, which is one more reason why I think these arguments just might be off) -- --bill