On Thu, Aug 15, 2019 at 11:29:49AM -0400, Keith Moore wrote: > The answer is no, they cannot, particularly not if either address > assignment or routing are hierarchical. It should be clear to one and all that IPv4 addresses are too small and also that we're not going to do an IPv8, so 128 bits it is and must be. At any rate, the size of the routing tables surely still matters, even if it feels like we're a long way from the CIDR crisis of 1994 (fun times at NANOG). Whatever the size of an IP address, the number of routing table entries needed still needs to be kept below 2^35 (RAM) or so (2^40 with NVMe), no? Now, 128-bit address spaces must be sparse by definition: we can't make even 2^64 addressable nodes with all the stuff on planet Earth, or even on all the planets in our Solar system. But routing cannot be practical with purely randomized and sparse allocation of addresses in a 128-bit address space, so no matter what, our effective average prefix length -that is, log base 2 of routing table size in entries- must be manageable lest we be forced to abandon the idea that routers must have full routing tables. I've long felt that it would be much more scalable to route only based on AS (i.e., IP packets should have addresses, yes, but also src and dst AS numbers), and have an address->AS resolution protocol.... ....then end-nodes would have to first resolve a peer's address to AS number, then route based on AS number. This would allow routing tables to shrink enormously, down to 2^16! Address -> AS resolution services would have to have well-known AS numbers in order to bootstrap, naturally. Though even then, I suspect we'd end up having structure in IP address allocations to make resolution practical. Such a hare- brained scheme is at least 30 years too late to be feasible. Nico --
