RE: Ungodly packet loss rates

I phoned Sprint asking them to look into it - and I do have to say they
were very responsive and conducted extensive testing - and the bottleneck
turned out to be, surprise surprise, MAE-WEST.

Actually, I'd be more than willing to bet money that it's NOT actually
MAE-WEST itself that is the problem. What has been happening is that
the major NSP's have been redesigning thier networks to try and get a
handle on the incredible growth rates. The "big boys" have gone to a
private interconnection type of arrangement that allows the traffic
to flow between them at multiple points with much less congestion.
(Jeff and Jack from MCI and Hank or Kurt from Sprint could possibly
elaborate upon this topology) The one outstanding problem that results
from this is that the resources of thses companies are no longer
directed at the meet points as an important (allow me some license here
guys) part of thier network. So now all of the traffic originating from
peers at the meet points moves happily across the meet point to the
NSP's router, and then hops onto a DS-3 to travel from the meet point
router to thier backbone. There are two achilles heels here. The port
at the meet point (a switched FDDI port still can only handle 100Mbps
simplex), and the link between the NSP's meet router and thier backbone.
I think in most cases one will find that these are the locations of the
problem, and it's not really a problem that can be fixed by adding more
hardware or new technology at the meet points. It's more a policy and
financial decision that must be made by the NSP's.

Since all the discussion on what backbone (tier 1) providers can do to
improve interconnection seems to have done something - intra-provider traffic
seems to be much better nowadays than it was last year - can we start doing
the same with major exchange points such as MAE-WEST and MAE-EAST, etc?

Again, I think one of the guys I mentioned above could better express
thier ideas here...

If the problem is that the routers/gigaswitches are overloaded, perhaps
some sort of intelligent traffic analysis and division might be in order?

Now this definately has merit. Strong planning (a point that I'm weak in)
by both the Exchange operators and the NSP's could prevent a large part of
the congestion that shows up.

But again, you must look at things from the other side. If I give peering
to someone (say a regional ISP), I have to invest money into my network to
handle the traffic load they send me. If I sell a circuit to that regional,
then there is a revenue stream to maintain the network. Peering has become
a financial decision in addition to a network engineering decision. This
is really where some type of settlement method needs to be established if
peering as we know it today is to remain a viable method of interconnecting
the networks.

Chris A. Icide
Senior Engineer
Nap.Net, L.L.C.