RE: network controls are necessary

["Tony Li" <Tony.Li@procket.com>]

|   For example, we can say that, with CIDR, the size of the 
|   routing tables
|   in the DFZ routers should scale as something between O(P), 
|   which is the
|   number of DFZ providers, which itself is expected to scale with the
|   number of sites as something like either O(log S) or O(sqrt S),
|   depending which opinion you believe. On the other hand, 
|   without CIDR,
|   the routing tables would scale as O(S). That makes the 
|   comparison quite
|   clear.


Christian,

For M multihomed sites, CIDR will give us growth of O(P + M), where
M is likely to come to dominate (and overwhelm).  We need an
architecture where we have some confidence that in the long term, the
M factor is either removed from the equation or controlled.  Certainly
something like O(P + log M) might be reasonable, but I see no way of
doing that other than something like the geographic approach and the
issues there are considerable.

I can see a way to get to O(P), but we're a long way from consensus
on that.

In any case, the real scalability requirement is that

	d(DFZ)
	------  <<  Moore's law
	  dt

O(P+M) does not meet this requirement, IMHO.

Tony