Paul A Vixie
paul at vix.com
Mon Jan 27 20:07:08 UTC 1997
I want to at the outset that I don't know what MCI does or why, and if I
did know I would probably be under NDA and so I wouldn't be able to talk
about what I knew. I used MCI as an example in my last note to this list
because the person I was replying to had used MCI as an example. For all I
know, MCI peers with space aliens, or doesn't, or not. But, getting to the
provider independent portion of Jeff Young's recent message here:
> the end of your post is interesting, however. it may come to the point
> where larger carriers are forced to tag local as's local and to peer on the
> 'local' basis. but what benefit for that added complexity? local networks
> would still need to buy transit from someone.
Right now we're sucking down a fair amount of backhaul, symmetric to the
backhaul of a local Mom & Pop's transit provider, carrying traffic which is
both to and from the same local area. If backhaul were free, or even cheap,
or even available in the quantities we need, the engineering simplicity would
win out -- it's easier to manage a network if there are fewer links and fewer
powered boxes inside it. On the other hand we're having a lot of trouble
getting enough backhaul on some paths -- at any price.
So OK, let's assume that transit providers all come to every local area to
pick up local customers. A Mom & Pop ISP buys transit from one of them, on
the hopeful assumption that it will peer with the other locally present big
providers, thus preventing traffic between endpoints 10 miles away from
taking a 500 (or 2000) mile loop through some more distant private peerage
or exchange point. This saves on the rarified wide area backhaul, and it is
certainly better than what a lot of local markets have now.
But if we're assuming the existence local links between big ISP's in each
local market, what's wrong with the local exchange concept (since, as I said
in an earlier message, sum(1 .. N-1) links are more expensive to build than
N links and a GIGAswitch, for reasonable values of N)? And if you're doing
a local exchange, why not let Mom & Pop's also connect there, so that their
transit link can be a private 10Mb/s or FDDI wire between cages -- saving the
end users money indirectly by cutting down the number of bypass carriages?
And finally-- if each of the previous steps make sense and you get this far--
why not have big and little ISPs peer directly, sharing only routes which do
not result in assymetric wide area expense? (Route segregation between local
and wide area was already necessary for an earlier step.) This saves router
backplane expense, which while easier (in 1997) to buy than wide area backhaul
is still in shorter supply than some people would like. Some L2's are more
reliable than some L3's -- and are almost always cheaper per bit carried.
This changes "we'll only peer with you if you have a network topology similar
to ours" into "we'll be happy to peer with you but be aware that we only send
local routes when we peer at public exchanges, and if you want a full routing
exchange it'll take 6 T3's worth of private peering -- can you afford it?"
I think the network will work better and scale better when this kind of
"peering" becomes the norm. If wide area telecom costs are the reason big
guys don't like to peer with little guys, then by gum let's see peering and
charging all take place exactly where the economics require it.
More information about the NANOG