Cable & Wireless "de-peering"?!?

Mailman List Mirror (Read Only)
1

Peter writes:

> I do still think UUnet is in a downward spiral, just like C&W. Strong
> peering policies are not good for the Internet.

David replies:

  I'm not so sure that's true. Personally, I think the Internet is better
served by having a smaller number of larger and better maintained meeting
places than by just having a large number of peering points were everyone
connects to everybody else.

  Here are just a few reasons why, for example, it's better if you use
transit to FooNet to reach BarNet rather than BarNet peering with you
directly (assuming you are not too big yourself):

  1) FooNet and BarNet are more likely to keep their peering points scaled to
handle the load than you are. They are more likely to monitor performance
and shut down failures.

  2) FooNet and BarNet will meet at more places than you will meet BarNet,
allowing traffic to get off the source network faster and providing better
fault tolerance.

  3) Fewer BGP sessions means faster convergence and less instability.

  4) You may be more likely to meet BarNet at public peering points while
FooNet is more likely to meet BarNet at private peering points. Your traffic
to BarNet will get the benefit of the higher amounts of effort FooNet and
BarNet will put into keeping their meeting points efficient.

  DS

First, I'm not sure how these reasons relate to UUNET's direction
(downward spiral, up and to the right, quick-quick-slow). Second,
many of them are based on presumptions that relate to past
performance, with no guarantee of future results. This is especially
true if you consider traffic exchange to occupy more than two points
on the spectrum than just Free---Paid Transit.

For example, 1) has historically been true when BarNet is not paid for
taking your traffic; it may value it for other reasons
(content/eyeballs), but unless it is paid it is hard to get resources
for link or equipment upgrades. In paid peering, it is paid to take
your traffic and is likely to take seriously performance degradation
or failure; whether you do is, of course up to you. So it may make
more sense to enter a paid peering agreement with BarNet than to use
transit through FooNet to get to them.

Similarly, 3) has the interesting assumption that you are not using
BGP to talk to FooNet, which will not be the case if you have a
multi-homed transit arrangement aside from the peering to BarNet. 4)
and 2) contain some interesting assumptions about meeting points and
the topological relationships among FooNet, BarNet and the customer
network. 4) seems to assume, for example, that more of BarNet's
effort goes into a private connect to FooNet than to the public
connect at BazNAP. Could be true; could be entirely the other way
around.

As David's post points out, though, there is no easy assumption about
what is going to be best. I think the Internet is actually best served
by having lots of available of choices for how to interconnect. That
way, when your mileage varies, there is something you can do about it.

      regards,
        Ted Hardie

Disclaimer: I am not speaking for my employer. This disclaimer is,
however, included solely for the amusement of Stephen Stuart; any
other use may be a violation of his rights to amusement.
(http://www.irbs.com/internet/nanog/0104/0164.html)

2

Also sprach hardie@equinix.com

This is especially true if you consider traffic exchange to occupy more
than two points on the spectrum than just Free---Paid Transit.

Using this statement as a hopping off place, and not trying to
contradict you in any way as I largely agree with you.

I think what it comes down to is that you get the best connectivity for
the lowest price. Considering in the price calculation the ability to
maintain that connectivity. Radical thought, huh? :slight_smile:

Seriously though, IgLou has transit with uu.net and att.net. Given our
size, we really aren't actively considering any other transit
connections at the moment as our current setup meets our needs. That
being said, if anyone has a great deal on transit...don't not call
because we're not actively considering it. :slight_smile:

What I am looking into some is peering relationships...usually fairly
small scale given the scale of IgLou's network. Particularly at this
point we're considering it primarily with setups that are predominantly
content sources, as we're predominantly content sinks, and this is the
best way to get the most bang for the buck in our peering, by peering
with networks that predominantly source content rather than sink it.
*shrug*

I probably will not get more than one or two significant peering
connections at present, however, because the cost and effort of
maintaining them would outstrip the benefits.

It all comes down to cost-benefit tradeoffs. If you have more transit
or peering connections that you can keep on top of for maintenance and
scaling, then adding more will likely end up degrading your network.
I feel that IgLou is at that point for transit, but could still benefit
from peering connections, and that's really the only thing on my plate
that I'm looking at currently. If/when we get a couple of peering
connections, then there will be enough that I won't be able to keep on
top of them and I won't add anymore until we grow our resources to be
able to manage them better.

Its really just not all that difficult of a concept to grasp.

3

Similarly, 3) has the interesting assumption that you are not using
BGP to talk to FooNet, which will not be the case if you have a
multi-homed transit arrangement aside from the peering to BarNet.

  No, I'm assuming you *must* have BGP sessions to everyone you get peering
or transit from. The more places you peer with, the more BGP sessions you
are going to have.

4)
and 2) contain some interesting assumptions about meeting points and
the topological relationships among FooNet, BarNet and the customer
network. 4) seems to assume, for example, that more of BarNet's
effort goes into a private connect to FooNet than to the public
connect at BazNAP. Could be true; could be entirely the other way
around.

  That's not quite the assumption. It really is that more effort goes into a
private connect to FooNet than to the public connect at BazNAP as it applies
to your particular connection. They might monitor their BazNAP connection
closely, but not be particularly concerned if they notice possible problems
with a smaller peering connection -- especially if they don't charge for it.

As David's post points out, though, there is no easy assumption about
what is going to be best. I think the Internet is actually best served
by having lots of available of choices for how to interconnect. That
way, when your mileage varies, there is something you can do about it.

  Well, I won't argue with that.

  Another possible disadvantage with 'free' peering is you may have a harder
time getting the larger companies to deal with issues that arise surrounding
that peering. They may feel they have no financial incentive to worry about
it, and you really have no contractual leverage to get, for example,
response time guarantees.

  DS

4

  Another possible disadvantage with 'free' peering is you may have a harder
time getting the larger companies to deal with issues that arise surrounding
that peering. They may feel they have no financial incentive to worry about
it, and you really have no contractual leverage to get, for example,
response time guarantees.

Isnt the financial issue and management overhead less (relatively) for a
large company with a large well ordered NOC than that of a small one? So
their only excuse remains the one that they 'think your too small to care'

Steve