I think that before we decide whether we are for or against a 
locator/identifier split, we need to pick a single definition for what the 
locator is and a single definition for what the identifier is. LISP refers to 
the IP address of a certain router as a "locator" and the IP address of an end 
system as the "identifier", and splits them by making them different layers in 
a tunnel-like (map/encapsulation using an ip/ip tunnel header) overlay network. 
ILNP and other GSE-derivitives make the IP prefix the "locator" and the IP End 
Point Identifier the "identifier", and split them in the sense that the 
identifier is consistent end to end but locator is fungible at stated points in 
the network according to a set of rules and therefore has different degrees of 
locality.

I personally am in favor of the latter (ILNP ideally, 
http://tools.ietf.org/id/draft-mrw-behave-nat66 as a fall-back position which I 
suspect will be easier to deploy as it does not require host changes), and not 
in favor of the former. 

My reasoning is that the ILNP model is a win-win between edge and transit 
networks, giving them both a direct solution to a problem that they have said 
publicly and loudly that they would like to see fixed, while LISP is (in my 
opinion) not.

The problem we set out to solve recognized a fundamental difference of opinion 
being battled out in the RIR communities. Edge networks, as a class, want to be 
independent of their upstream networks. They want to be able to change 
upstreams without a significant level of effort in renumbering or changing 
their routing. They generally don't want a large route table, and generally 
solve this using default routes to their upstreams, potentially tuned with a 
few more-specifics for key business partners. Their solution in IPv4 is either 
IPv4/IPv4 NAT, which has operational problems for several classes of 
applications, or PI addressing, which explodes the route table by enumerating 
the edges. Transit networks, as a class, prefer or don't object to having their 
customers captive to them, and want a route table of a manageable size. For 
them, a PA allocation - allocation that enumerates only the points where the 
edge attaches to the transit core - achieves that goal.

I prefer the GSE model, instantiated in ILNP, because it gives the edge 
networks the relative simplicity of management and independence they seek - 
their network appears to them as if it were PI - while giving the transit 
networks the relative simplicity of management and route table effects of a 
totally or mostly PA network. Both sides win. The route table in the backbone 
enumerates the folks in the world that have AS numbers, whether they are truly 
ISPs or whether they are simple complex edge networks. The edge networks are 
independent of their upstreams with the added benefit of being able to address 
systems in their own and other edge networks statelessly, uniquely, and 
exactly, just as if addressing were truly end to end, and allows for 
predictable load sharing across DMZs in parallel, which IPv4/IPv4 stateful NAT 
does not. 

I don't believe that LISP does that. It does simplify the transit domain - 
addressing in that domain is PA. Outside the transit domain, addressing remains 
PI, and the complexities implied by that are moved from the transit domain 
(which has expertise in handling that) to the edge (which does not). I think it 
makes the network harder for the edge to attach itself to without expert help.

BTW, I have a similar opinion of shim6, although the wording of the argument is 
different. shim6 makes the entire network PA, which is good for the transit 
networks, and makes multihomed edge networks bear the complexity burden of 
maintaining routing for each prefix they use throughout their networks. The 
edge, I fear, lacks either motivation or expertise to offload the complexity 
from the transit core.


As to your proposed questions, I think they make a false assumption - that we 
have four options on the table:
 - one with a hierarchical system and Global/universal characteristics, 
 - one with a hierarchical system and local characteristics,
 - one with a flat system and Global/universal characteristics, and
 - one with a flat system and local characteristics.

I'm not sure the proposals are malleable in that way.


On Jul 25, 2010, at 5:02 PM, Toni Stoev wrote:

> Researchers,
> 
> I think we must not attack each other personally. Nor even each other's views.
> We need to find the best solution each of the design goals.
> 
> Here we talk about identity/location split. Robin is against. Most but seem 
> to be for.
> So with the identity/location split we encounter identifier uniqueness. 
> Whether it has to be global/universal or just local. Some hesitate about 
> this; others, like me, take universal uniqueness for natural.
> Next question: Uniqueness system is to be hierarchical or flat? Any arguments 
> on flatness side? But there is a ready-made solution. Is it reliable?
> Are we going any further with this ambiguity?
> 
> I propose this approach: Let's answer the questions the reverse way.
> 
> 1. Hierarchical or flat uniqueness system?
> 2. Global/universal or just local uniqueness?
> 3. Shall we make identity/location split?
> 
> Good will
> Toni
> _______________________________________________
> rrg mailing list
> [email protected]
> http://www.irtf.org/mailman/listinfo/rrg

http://www.ipinc.net/IPv4.GIF

_______________________________________________
rrg mailing list
[email protected]
http://www.irtf.org/mailman/listinfo/rrg

Reply via email to