maximum ipv4 bgp prefix length of /24 ?

Fri Sep 29 04:54:01 UTC 2023

> 
> In principle, a company could make a business out of announcing a
> large block from a bunch of peering points and then tunneling (vpn)
> parts of it back to customers with sub-/24 assignments. With a broad
> enough selection of peering points, the routing would not be too
> inefficient. And it would divorce the IP addresses from the last-mile
> Internet infrastructure, allowing you to take your addresses with you
> as long as you kept paying the tunnel company.

Actually, such a service would be much easier to stand up as a bunch
of virtual routers running on VPS instances in various cloud providers.
Simple as standing up a VPS running Debian 12 and FRR, then sell
routed blocks to people.

Personally, I think that’s fairly hideous, but someone can probably find a
way to make money doing it.

I know that there are companies charging $ridiculous for “SDN” solutions
that are literally not much more than a tunnel running between two
AWS nodes.

> In practice... there's not enough money in it. If you could ante up
> the cost, you could find a way to qualify for and acquire a full /24.

Given what some of the SDN providers out there are charging, I’m not
so sure that’s true. YMMV.

>> Is it unacceptable; considering most big networks that do full-table-routing also use multi-core routers with lots of RAM?
> 
> You're thinking of DRAM. But that's not the way it works. Some routers
> use heavily parallel routing engines, each of which need enough dram
> to hold the full forwarding information base and which can suffer from
> CPU cache exhaustion even then. Others use an expensive kind of memory
> called a TCAM that's very fast but both expensive and power hungry, so
> generally not sized for huge numbers of tiny routes.

Trio and Later generations of Juniper MX line cards (which are getting fairly
long in the tooth these days) can handle more than 5M routes in the FIB.
Even the old (now ancient) DPCs can handle ~1.5M routes if you don’t
need a boatload of access lists. (Basically you have to steel FIB memory
from the Access List memory partition, but that’s a simple software
command and a reboot of the line card).

Owen