Vijay Gill <wrath@cs.umbc.edu> wrote:
Scalability on the Internet pretty much means that algorithms should run in O(log(N)**M) where N is the total number of end-points and M is constant. (Note that non-CIDR unicast routing doesn't fit this criterion, but CIDR does).
Reference to this?
I guess it's my private rule-of-thumb. (Antique Kleinrock and Klein, maybe?) Anyway, O(N) seems to be infeasible (as at least some network components will have to scale at super-Moore's law exponential rates), and O(log(N)) - too strict (all dynamic routing algorithms are super-linear, but still work). The design based on O(P(log(N))) algorithms was demonstrated to work in practice (DV of SPF with route aggregation). O(P(N)) (non-hierarchial routing) was demonstrated to fail during the CIDR deployment saga.
The pressure is being applied now. Vendors however had a lock on this market, witness how long it took for cisco to give oc-12 atm interfaces. They didn't move till uunet put the GRF into the core.
The pressure is called competition :) Unfortunately, large switching equipment vendors still didn't get a clue on how to build real fast routers. (Nortel may be an exception, with their stake in Avici; however i have serious misgivings about ability of Avici design to deliver on promises. Hypertorus is not what one wants to use for general permutations. And i find their claim to be inventors of massively parallel routing simply hilarious. Obviously, their marketing department employs Al Gore as a PR consultant. Never mind that the only reference to the name "Avici" i was able to find is the particularly hot kind of hell in Buddhist mythology :) --vadim