Favorites (Re: UUNET peering policy)

Leo Bicknell bicknell at ufp.org
Wed Jan 17 18:17:47 UTC 2001


On Mon, Jan 15, 2001 at 10:37:18AM -0800, Sean Donelan wrote:
> The traditional arguement is a network composed mostly of a few large data
> centers, with lots of servers sending traffic is getting a "free ride" on
> the network which built out nationwide and has POPs in every LATA.

	Note, one of the ways AboveNet addresses this issue is by 
honoring meds.  This does not work with all networks, but I will
put forth an example:

                            LAX             NYC
Net 1 (eyeballs):            1---------------2-----user
                             |               |
                             |               |
                         PeeringA         PeeringB
                             |               |
                             |               |
Net 2: (content):  server----3---------------4

	In a traditional (closest exit) setup the user will make a request
of the server, and traffic will flow as follows:

user->server: 2->4->3
server->user: 3->1->2

	Since there is generally much more traffic (bits) server->user
the "eyeball" network incurs a higher bit/mile cost (the long haul from
1->2) in this case.

	In the pure case (100% eyeball network peering with 100% content
network, closest exit) it is safe to say the eyeball network carries
a higher cost.

	Consider what AboveNet does.  By honoring meds the following
paths occur (abovenet as net 2 in the picture):

user->server: 2->4->3
server->user: 3->4->2

	The bit/mile cost just shifted to the 3->4 link, paid for
by the "server" network.  At this point the server network is actually
incuring more cost (assuming the pure case, again).

	This all gets muddied very quickly in the real world, as no
network is "100% eyeballs" or "100% content", and there are many other
geographical issues involved etc.

	This is but one tool that can be used to fix the problem of
one network incuring more cost than the other to interconnect.  To
refuse flatly based on ratio, or bit/mile cost without looking for
relatively easy alternatives to balance the cost is a bit short sighted.

	Peering is about "equal value".  Consider the pure content provider
co-located witht he pure eyeball provider, where neither have a long haul
network (eg local providers, similar bit/mile cost on each side).  The
costs are equally distributed, both get similar "value" from the
interconnection (and indeed, can't exist without each other in the pure
case), but the ratio may well be 10:1.  

-- 
Leo Bicknell - bicknell at ufp.org
Systems Engineer - Internetworking Engineer - CCIE 3440
Read TMBG List - tmbg-list-request at tmbg.org, www.tmbg.org




More information about the NANOG mailing list