Copyright 1997 Nikkei Business Publications,Inc. all rights reserved, permission is hearby given for reproduction, as long as attribution is given and this notice is included.
The future of Internet backbones.
By: Scott Bradner
The Internet is very much a victim of its own success. There are now at least 13 million computers connected to the Internet. Internet traffic is growing far faster than anyone could have imagined just a year ago. The Internet is seen more and more as a key piece of national and international infrastructure. The advent of Internet-based commerce has the potential to maintain this growth almost indefinitely. This explosive growth has strained the Internet infrastructure to the point that people like Bob Metcaffe have started claiming that the Internet is due for a catastrophic collapse in the near future. He has said that even if the Internet as a whole does not collapse, then major Internet service providers (ISPs) may be in real danger of periodic total failure.
Any way you look at it, the Internet infrastructure is in current or potential trouble. Internet traffic at MAE-East, a inter-ISP connection point located near Washington DC, now has daily peaks of over 750Mbps bits per second Traffic on some backbones is increasing by 30% per month. The size of the routing table that the backbone routers must be able to support is now over 36,000, and more importantly, the data in the routing table can change as often as 110 times per second. The fastest of the current generation of routers can support interface speeds of up to 155 Mbps and this is not fast enough to keep up with the existing demand never mind the future demand. (The ISPs that are advertising themselves as having backbones faster than 155Mbps break down the faster links into multiple parallel links, each with a maximum data rate of 155 Mbps.) Routers with faster ports are promised but not yet here.
There is not enough fiberoptic cable in the ground to support the growing demand and it takes a long time to install more. New technologies such as Wave Division Multiplexing which can be used to expand the capacity of fiber links but they do so by adding additional parallel circuits on the same fiber, this means still more router interfaces will be needed.
Currently, the total market for backbone-scale routers is limited by the fact that there are not many very large Internet backbones and by the fact that even the largest of the corporate networks do not present the same level of demand on the routers. The limited potential revenues and the cost of development have meant that here are only a few vendors servicing this market.
The interconnect points such as MAE-East are severally stressed and are seen by many observers as the weakest links in the Internet Infrastructure. Most of the larger ISPs are now bypassing these interconnect points in favor of private direct ISP to ISP connections.
I have painted an unpleasant picture of where we now are in the development of the Internet and a potentially bleak outlook on the future but I am not yet done being negative.
There are a number of people, mostly people with telephone company or mainframe backgrounds who think that if ISPs just start charging for their services in a different way everything will be magically fixed. Most of these people think that if Internet users were charged a usage-based fee then they would change their behavior to minimize their cost. This, tied to the ISPs starting to charge each other a usage-based fee to exchange traffic is, according to these pundits, supposed to moderate the traffic growth. I do not think this is true for two reasons. First, there is currently no way to record the traffic of an individual user in a way that can be used by an organization to give feedback to that individual in the same way that a bill for long distance telephone service can identify each originating telephone extension. In the case of a telephone bill a manager can talk to an employee who talks too much. Second, it is very hard for an Internet user to know in advance what traffic some action will produce. Clicking on one web link might cause a 10KB file to be read, while clicking on the link beside it could cause a 10MB file to be sent.
Enough pessimism. The Internet has a great future we just have to figure out how do things so that we can enjoy it when it comes. Notice that I did not include running out of IP addresses as one of the looming problems. There is less of a problem in this area than there once was. There are a number of reasons. The use of classless inter-domain routing (CIDR) has enabled more efficient address assignments that more closely match the actual needs of organizations than the old class A. B & C addressing did. As long as addresses are assigned following the actual topology of the underlying network then CIDR addresses can be aggregated in a way that reduces the size of the backbone routing tables. The use of address translating firewalls and network address translating (NAT) boxes permits an organization to use RFC 1918 private addresses within their own network and minimize the size of the address block they need for external visibility. Finally, IPv6, with its very large address space will relieve any remaining scarcity as it starts to get deployed.
I have outlined both technology and policy challenges that must be faced. On the technology side there are a number of promising developments. The above mentioned address translation technologies will be of great help in moderating the growth in the size of routing tables and thus reducing the processing load on the routers. In and of itself keeping the routing table growth small is not sufficient since the routing computational load is a combination of the size of the table and the frequency with which that table must be recalculated. Many of the ISPs are currently experimenting with forms of "route damping", in which small fast changes in the routing information are delayed and summarized to reduce the change frequency. The combination should be able to keep the route table processing load growth under the rate of growth in the power of the processors in the routers. This will be the hardest problem to solve to keep us on track for the bigger Internet that is on its way, but with discipline it can be solved.
Part of this discipline is that sites have to renumber their computers when they switch ISPs. One needs to renumber when moving in the network topology so that your address can represent your actual position in the network and be aggregated with the addresses of others. This is needed to slow the growth of the routing table. The use of Dynamic Host Configuration Protocol (DHCP) makes renumbering far easier and the advanced renumbering abilities in IPv6 will make it easier still. Alternative addressing architectures are being explored for IPv6 that could cause the issue of renumbering to become moot because it will be so easy to renumber a site without anyone having to reconfigure routers or hosts. The use of NAT boxes or address translating firewalls can also make the difficulty of renumbering a thing of the past.
We are going to need bigger and faster routers. Routers that can support individual interfaces in the gigabit per second range. But these are being developed now and should be on the market soon. We are going to need WDM to be able to create new bandwidth on existing fiberoptic cables, and the WDM development is well under way.
But I think there is one other thing that will be needed: differential services. Right now ISPs can only offer one product, best effort IP packet delivery. In fact the best level of guarantee that a customer can get from many ISPs is that they agree to accept your packet, they make no claim that they will do anything useful with the packet, but they will accept it. I think that ISPs need to start offering at least two levels of service. In one level a user will get the current "best effort" service which is subject to packet loss due to congestion. The second level of service, which would be more expensive, would include a guarantee that this part of the service would not be oversold and that the ISP guarantee to deliver the packet not just accept it. The delivery might be to another customer of the same ISP or it might be to another ISP which might or might not have an equivalent guaranteed delivery service. Thus implementing differential services is not just a technical issue, the IETF's integrated services work may be able to be used here, but also a policy and ISP relationship issue. This is one place that ISP to ISP payments may actually be reasonable. I think that differential services are needed because I do not think that the ISPs can keep ahead of the traffic growth curve for best effort delivery and the result will be an Internet that will get more and more congested, making it an unsuitable vehicle for the support of much of the projected Internet-based commerce. With differential services someone ruining a commercial web site might be able to pay more to ensure that their customers get reliable connections.
From a technical and business point of view I'm optimistic about the future of the Internet. Things get a bit trickier when contemplating the impact of governments trying to control or tax the contents of the transmissions over the Internet.