On 7/5/04 1:18 AM, "Steve Gibbard" <scg@gibbard.org> wrote:
The performance arguments are probably more controversial. The arguments are that shortening the path between two networks increases performance, and that removing an extra network in the middle increases reliability. The first argument holds relatively little water, since it's in many cases only the AS Path (not really relevant for packet forwarding performance) that gets shortened, rather than the number of routers or even the number of fiber miles. If traffic goes from network A, to network A's router at an exchange point, to network C, that shouldn't be different performance-wise from the traffic going from network A, to Network B's router at the exchange point, to Network C. Assuming none of the three networks are underprovisioning, the ownership of the router in the middle shouldn't make much difference. The reliability argument is probably more valid -- one less network means one less set of engineers to screw something up, but the big transit networks tend to be pretty reliable these days, and buying transit from two of them should be quite safe.
I believe that peering does lead to a more robust network and somewhat better performance. Being heavily peered means that when one of my transit providers suffers a network 'event', I am less affected. Also, just because I'm sitting at a network exchange point (and take my transit there) doesn't mean that's where my transit networks peer. Quite often, I see traffic going to Stockton or Sacramento through one of my transit connections to be delivered to a router just a few cages away at PAIX.
The pricing issues are simpler. There's a cost to transit (which is, to some degree, paying some other network to do your peering for you), and there's a cost to peering. Without a clear qualitative difference between the two, peering needs to be cheaper to make much sense. The costs of transit involve not just what gets paid to the transit provider for the IP transit, but also the circuit to the transit provider, the router interface connecting to the transit provider, engineering time to maintain the connection and deal with the transit provider if they have issues, and so forth. Costs of peering include not just the cost of the exchange port, but also the circuit to get to the exchange switch, sometimes colo in the exchange facility, engineering time to deal with the connection and deal with the switch operator if there are issues, and time spent dealing with each individual peer, both in convincing them to turn the session up, and dealing with problems affecting the session. Even if the port on the exchange switch were free, there would be some scenarios in which peering would not be cheaper than transit.
When we established our connection at PAIX, peering bandwidth was a factor of 20 cheaper than transit. Now they're at parity. Unfortunately, some *IX operators haven't seen fit to become more competitive on pricing to keep peering more economical than average transit pricing. $5000 for an ethernet switch port? It makes me long for the days of throwing ethernet cables over the ceiling to informally peer with other networks in a building. In the 'bad' old days of public exchanges (even the ad hoc ones), most of the problems were with the design and traffic capacity of the equipment itself (not a real problem now), not with actual 'operations'.