From: Deepak Jain (no email)
Date: Wed Nov 02 2005 - 17:19:35 EST
> I don't understand them, either. However, if you define incoming
> traffic as "bad", it encourages depeering by the receiving side if the
> incoming/outgoing ratio exceeds a certain value, especially among
> close-to-tier-1 carriers: the traffic does not automatically disappear
> just because you depeer. Now suppose that the sending side doesn't
> want to play games and buys transit from one of your other peers.
> Given the tier-1 status, there is some chance that this has a
> measurable impact on the traffic ratio with that other peer.
> Essentially, this is a self-fulfilling prophecy, and it works equally
> well if you define outgoing traffic as "bad".
I was trying to stay out of this. But I think I'm going to chime in here.
I think (originally)... means... whenever the NAP structure was created
and the NSFnet was decommissioned, long haul moving of the bits was very
expensive. Perhaps more so than the local-distance piece because you had
comparatively few of these links and had to cart bits a very long way to
dump them off.
Now I believe that with the influx of large, reasonable colos and 20+
high speed interconnects in a region (or slightly larger area). This
coupled with dramatically reduced long haul costs has shifted the
value/expense ratios in peering.
For example if you are a content network. If you needed to peer in 5
places in the old (long-haul=expensive) model would colo your content in
5 places near your interconnected and the only interconnections between
your colos would be whatever you needed to keep heartbeat and data sync
between them. But you incurred a large cost for this colo and manpower
and other things.
Now... let's just say you don't need to do that... because you can run a
few circuits around the country (or MPLS them) and its pretty "cheap".
However, the last-mile piece for access networks HASN'T moved as much in
the same time period. Lots of networks (Q, LVLT, GBLX, etc) built long
haul networks. Lots of them colo'd in ILEC switch (and tandem
buildings). But this doesn't do ISP type businesses very much good. You
still have to pay interconnect fees to the ILEC or exorbitant colo fees
to them to backhaul the circuit to your DC with all of your equipment.
Means... that even though you control the customer and the CPE, you are
paying fees to many folks that _really_ own the copper (or coax) or
underlying infrastructure and they have little to NO competitive
pressure to be more than slightly price concerned.
This drives you to the idea that actually moving higher PPS is "bad"
while increasing peak speeds is "good". Customers are buying/being
marketed to by peak speed. So you give them large pipes because once
you've paid the underlying tariffs to get to their house/business (say a
dry pair) whatever speed you signal on it is your equipment cost,
nothing else. But actually encouraging them to USE that bandwidth is
expensive because your cost of growing the T3, OC3, OC12 or whatever you
are backhauling from the various COs to your network is BAD and expensive.
This is the model as I understand it today. Formerly the longhaul and
cost of transit was so expensive these costs were more negligible.
Now we have a playing field. Now if you depeer a guy [Cogent] for
example, and you force him to buy transit from someone who doesn't have
a very vibrant transit business [NTT/Verio, I'm being kind] you increase
his costs and force [NTT/Verio] to upgrade their network which may take
time. The depeered guy suffers. Maybe he doesn't suffer much at all
[Like when Cogent could force all of its AOL traffic through its Level3
connection].
But his customers... They want speedy access to your eyeballs. Maybe
some of them will want to reduce the number of networks traversed to get
to your eyeballs. By limiting the number of SFI connections you have...
I would theorize you can force those who can afford it to interconnect
with you directly. Not everyone. But a few. If you perceive your network
as "better" than your SFI peers, then naturally you would assume that
the business of their customers would eventually flow to you.
The problem with this kind of increased instability is that the
perception and tenacity of all the players is very difficult to assess
and is often not as vastly different as the players would hope. But to
the extent you can force others to have to redo their network
interconnections with little impact [at least what you believe when you
are taking the action] the better it is for you. Especially if you are
running out of value-added sales pitches.
Further proving the counterpoint: Large eyeball networks (like Verizon
broadband) that use Level3 (formerly genuity) a lot, didn't get affected
by this depeering. Why? Because they've already added diversity to their
network. Even if the Level3 routes are normally chosen more often than
the other providers [guessing, not saying its true], Level3 forced their
95th percentile to peak with their other providers for a month while
simultaneously showing them a lack of "of my god my performance sucks"
calls from their customers with the new routes.
I think this has been proven out to be universally bad for Level3.
Cogent probably mostly bad. But I think I wanted to raise the point
[above] that while content guys have enjoyed cost improvements over the
last years the access guys have not seen as much improvement and the
areas of their businesses (loops, co colo vs server pricing, server
colo) have become a much larger, more unwieldy part of their businesses.
[guessing]
Comments?
Deepak
|
|
|