I-D Action:draft-azinger-scalable-addressing-00.txt

[Tony Li]

Hi Gert,

> Let me try to illustrate this with a specific example.  No complex policies
> for now, just focusing on "shortest path to destination".


Nope, sorry, you don't get to change the game in the middle.  You asked for a path equivalent or better than what a single prefix would give you.

Specifically, routing today does NOT give you the shortest path to the destination.


> There's a source prefix S1 that is routed via ISP 1, and a source prefix 
> S2 that is routed via ISP 2 (both prefixes coming from the respective 
> ISP's PA space, so BCP 38 filters will require that source address 
> selection automatically selects egress ISP).


No, BCP 38 requires that the source address MATCH the egress.  If the egress prefix is applied at the ASBR, that also suffices.


> Destination prefix D1 is connected to peer of ISP 1, while destination
> prefix D2 is connected to something 20 hops behind ISP 2 (assuming that
> "20 hops" means "more latency, less bandwidth", for the simplicity of
> the model).
> 
>    +--- S1 === ISP 1 === ISP X ======== D1 ---+
>  source         (*)                         destination
>    +--- S2 === ISP 2 --- ISP A, B, C -- D2 ---+
> 
> How exactly is the end host supposed to know that "S1->D1" is good, while
> "S2->D2" is bad and "S1->D2" is very bad?  (Assume that "use the longest
> sequence of common bits in the addresses" won't help, which is the normal
> case when crossing RIR regions).


It doesn't know and shouldn't care.  


> If I have a single S and D prefix, the end host does not *need* to know,
> and the network *will know* that the path via ISP 1 to its peer to D is 
> shorter than via ISP 2.


No, it will not.  It will simply use the path selected by BGP.  That could easily be S1 to D2.


> I hope I have succeeded this time in explaining where "we the operators"
> see a gap in the N*M model - and that has nothing whatsoever to do with
> "network admins not being up to the job".   


No, sorry.  What you've shown is that the N*M model doesn't give you a panacea.   It will not read your mind and give you ultimate optimality.

What the N*M model _can_ do, specifically with ILNP, is to present all prefixes to the end host, in the preference order that the destination selects.   Assuming that the destination is only advertising working prefixes, then this gives you equivalent semantics to what you have today.

Further, if the hosts also support MPTCP, you can make use of all N*M paths in parallel.  This is very nice as it not only discovers the path with the optimal throughput, but it also allows you to use ALL of the paths in parallel.  Can't do that with the PI model.

Bottom line: optimal path selection is neither a host nor routing function.  It's a transport function.

Regards,
Tony