Re: draft-azinger-cidrv6-00

[Fred Baker <fred@cisco.com>]

On Jul 22, 2010, at 1:28 AM, Brian E Carpenter wrote:
> On 2010-07-22 11:03, Nick Hilliard wrote:
>> In addition, until such time as there is a viable alternative for end-user multi-homing in ipv6, I'm not sure if making a recommendation like section 10 is actually going to achieve anything - other than raising eyebrows.
> 
> Other eyebrows are already raised by the prospect of ten million BGP4 routes. It's really odd that the RIR constituencies aren't more worried about that.

Well, there are mitigations. They know how to aggregate routes, they now have routers with actual memory in them, links and processing are fast enough that they don't see delays in convergence induced from that source (convergence delays continue to happen, but other reasons are more dominant).

I think your average ISP tends to think that his vendors will solve that.

Speaking from the perspective of one of the obvious vendors, I'm certainly willing to sell them memory as well; that's "just capex", expensive and something they love to hate, but a one-time purchase. And, by the way, not a primary driver of the cost of equipment - there are other things much more expensive than memory.

What is driving the RIR discussions is the desire of edge networks to be independent of their providers, not captive to them. They don't want to have to adjust (renumber, reroute) their networks when they change providers. The fundamental failing of shim6 (and for that matter, of lisp) is that it moves a lot of complexity from the transit networks, who know how to manage it, to the edge networks, who don't. Yes, it minimizes the number of routes in the transit domains by making them provider-allocated; it maximizes the routing for the edge network, because there are multiple instances of an almost-identical route for each subnet in the edge domain. The edge networks see a provider tie and more-complex routing as net negatives.

That is what is driving the discussion of ILNP etc. GSE enables the transit domains to have the relative simplicity of a PA network while giving the edge networks the relative simplicity of a PI prefix, and all the while enabling any system in the network to reliably address any other system in the network (unlike IPv4/IPv4 NAT) using a stateless (and therefore load-sharable and stable) prefix translation. There are some applications one doesn't run across prefix translation boundaries (anything that carries and cares about the prefix) and some things sensible people do in applications (SIP and HTTP refers use names, not addresses, for example); if you obey the rules, it works, and if you don't, you created your problem and you know how to solve it.