Re: about draft-baker-v6ops-l3-multihoming-analysis-00

[Fred Baker <fred@cisco.com>]

On Nov 21, 2006, at 11:59 AM, Fred Baker wrote:
> It would have to discuss traffic engineering, from the perspective  
> of saying what requirements were reasonable and solvable and what  
> were not.

Let me put hands and feet on that. One of the ISPs involved in the  
IAB workshop mentioned a need to provide for traffic engineering that  
would, among other things, allow one ISP to cause another ISP to load  
balance traffic across multiple intervening ISPs. In other words,  
imagine this scenario:
                          ,-.
                         /   \
                        /ISP B\
             ,-.       /       \       ,-.
            /   +-----+---------+-----+   \
           /     \     \       /     /     \
          /       +-----+-----+-----+       \
  The    ;         :     \   /     ;         :  ISP D
Internet |  ISP A  |      `-'      |  ISP D  |
         |         |               |         |  Customers
         |         |      ,-.      |         |
         :         ;     /   \     :         ;
          \       +-----+-----+-----+       /
           \     /     /       \     \     /
            \   +-----+---------+-----+   /
             `-'       \       /       `-'
                        \ISP C/
                         \   /
                          `-'

The ISP D has access to some number of undersea cables owned by two  
other ISPs (B&C) and needs to get access to "the Internet" (whatever  
that is) beyond the ISP A. The total bandwidth available is larger  
than the total traffic load, but total headroom is perhaps 30%  
(enough to handle the occasional loss of one cable but not two or  
three). For whatever reason, lets presume that a direct customer  
relationship between A and D doesn't work - maybe neither owns the  
cables. To a certain extent, ISP D can accomplish its goals by  
advertising different sets of sub-prefixes on each cable. However,  
the interests of the other ISPs differ from its, which means that  
ISPs B&C might accept its sub-prefixes, convert them to the actually- 
allocated prefixes, and advertise those north, with the effect of  
completely screwing up the attempted traffic engineering.  
Alternatively, ISP A might view ISPs B and C as primary and  
alternate, with the same effect.

I can think of a number of fairly simple solutions for this class of  
problem that can be effected by the ISPs themselves, but it seems to  
me to be borderline intractable as a general case. One ISP-specific  
solution might be for ISP D to write an agreement with ISP A about  
how it uses its links through B and C for traffic destined to D, for  
example, and ISP A might implement it by its internal routing  
regardless of the prefixes that B and C advertise to it. But in the  
general case, I not only have to have two or four ISPs agree, but all  
ISPs agree, and that seems (ahem) dubious. You heard it here first :-)

So we have requirements being posed to us from the ISPs, but I'm not  
sure that all of them are problems that we can solve in the general  
case. We need to tease the ones we can solve apart from those that  
are probably just too hard.