338 PART III Routing and Routing Protocols Economic considerations often play a role in routing policies as well. In the old days, there were always subsidies and grants available for continued support for the research and educational network. Now the ISP grid-net has ISPs with their own cus- tomers, and they can also be customers of other ISPs as well. Who pays whom, and how much? PEERING Telephony faced the same problem and solved it with a concept called settlements. This is where one telephone company bills the call originator and shares a portion of the billed amount with other telephone companies as an access charge. Access charges compensate the other telephone companies, long distance and local, that carry the call for the loss of the use of their own facilities (which could otherwise make money for the company directly) for the duration of the call. Now, in the IP world the source and destination share the cost of delivering packets, but the point is that telephony solved a similar issue and the terminology has been borrowed by the ISPs, which are often telephone companies as well. The issue on the Internet becomes one of how one ISP should compensate another ISP for delivering packets that originate on the other ISP (if at all). The issue is compli- cated because the "call" is now a stream of packets, and an ISP might just be a transit ISP for packets that originate in one ISP's AS and are destined for a third ISP's AS. ISP peers have tried three ways to translate this telephony "settlements" model to the Internet. First, there are very popular bilateral (between two sides) settlements based on the "call," usually defined as some aspect of IP packet flows. In this settlement arrangement, the first ISP, where the packet originates at a client, gets all of the revenue from the customer. However, the first ISP shares some of this money with the other ISP (where the server is located). Second, there is the idea of sender keeps all (SKA), where the flow of packets from client to server one way is supposedly balanced by the flow of packets from client to server the other way. So each ISP might as well just keep all of the revenue from their customers. Finally, there are transit fees, which are just settle- ments between one ISP and another, usually paid by a smaller ISP to a larger (because this traffic flow is seldom symmetrical). Unfortunately, none of these methods have worked out well on the Internet. TCP/IP is not telephony and routers are not telephone switches. There are often many more than just two or three ISPs involved between client and server. There is no easy way to track and account for the packets that should constitute a "call," and even TCP sessions leave a lot to be desired because a simple Web page load might involve many rapid TCP connections between client and server. It is often hard to determine the "origin" because a packet and packets do not always follow stable network paths. Packets are often dropped, and it seems unfair to bill the originating ISP for resent packets replacing those that were not delivered by the billing ISP in the first place. Finally, dynamic rout- ing might not be symmetric: So-called "hot potato" routing seeks to pass packets off to another ISP as soon as possible. So the path from client to server often passes through