At 07:48 AM 5/5/2006, Peter Cohen wrote:

On 5/4/06, Aaron Glenn <aaron.glenn@gmail.com> wrote:

...

On 5/4/06, bmanning@vacation.karoshi.com <bmanning@vacation.karoshi.com> wrote:

...

why would anyone do that?

--bill

Some companies feel entitled to charging more for their routes than they would for simple transit.

aaron.glenn

John: Hopefully this comes out clearly, as writing can be more confusing than speaking... Are you getting at Inter AS /SLA/QOS that you would get from transit vs. best effort peering? Even that has some issues, the one that jumps out to me is hopefully clearly stick figure-diagrammed below:

AS#x $--SLA-->Transit ok... But... AS#x $--SLA-->Transit <-(second hop)--Customers/Peers---No Qos/SLA--->

My point is it is hard to do anything beyond the first AS# for any SLA that you would be paying, since after that the packet switches to no money packets on a paid connection, pushing out the issue for things sent down that pipe...

Peter Cohen

It was not about the SLA, although in theory, buying transit should give the provider more incentive to help. The off-list discussion was more about avoiding the dependency problem of peerings. A "good" peering involves multiple points of geographically diverse interconnections. The number and location of these interconnections would depend on the unique combination of architectures of the two peers. If an AS does not have the traffic levels to justify multiple connections into a neighboring AS, relying on a single interconnection point is a problem. Even if the interconnection does not go down, it might not be a good way to reach particular networks in the other AS. Instead, it might be wiser to "tune" traffic via a different neighbor using transit. In other words, it gives you the best of both worlds. Most traffic travels directly to/from the SFP provider that serves the corresponding networks (like a peer). However, one can use the transit option at will for particular routes. And, one can use transit via the other SFPs should any transit to an SFP fail (fiber cut, etc.) Given that transit is pretty cheap, it seems more cost effective, at lower traffic levels, to purchase single transit interconnections to all the SFPs than attempt true peering at a much larger number of interconnections to those same SFPs. This is getting pretty theoretical, but I was curious if such a business model was attempted. The original SAVVIS did this in part long ago, but to just three neighbors. (I think they are now part of C&W now...I can't keep track of all these mergers.) It sounds like Internap is pretty close to this model, although I don't believe they have transit to all nine (if my SFP count is correct). John