Send in your ideas. Deadline June 1, 2024

Last update: 2004-03-31

End: 2003-01

Atom-Based Routing; Answers to Questions

Improving global internet routing by implementing atom-based routers.

6. Answers to Questions

6.1. What about IP Version 6?

At the moment, applicability of atom-based routing to IPv6 is an open-ended question. The routing architecture for IPv6 is not well-defined. If, as seems likely, momentum or inertia is the rule, then the IPv6 routing architecture will be fundamentally similar to the IPv4 architecture (including CIDR and its limitations) and the results of this project will transpose cleanly.

IPv6 does introduce a more structured allocation of addresses, apparently intended to allow better aggregation. However under CIDR, multihomed ASes will still introduce prefixes that their providers cannot aggregate (Section 2.3). Also, current engineering tactics (such as load balancing across prefixes) will still be prevalent in an IPv6 world. In addition, the increased length of IPv6 addresses will have a negative effect on router table size, and on the complexity of (distributed) computations.

However, the combination of atoms and IPv6 produces the interesting idea of placing an atom id in IPv6 addresses (at the time of allocation of the address). Routers would not need to maintain an explicit mapping between atoms and address prefixes covering such addresses. The number of entries in the tables covering such addresses would be equal to the number of atoms rather than the number of prefixes (which is substantially less).

Although the address format of an IPv6 address has pretty much been defined, if an implementation of atoms were done and proved successful, there would be motivation for the IETF to look again at the address bits and the way that they are allocated, and perhaps carve out some for an atom id.

6.2. Who Benefits from Atom-Based Routing?

Keeping the Internet backbone routing tables small is beneficial to all users of the Internet. It will help reduce the cost of the infrastructure, simplify the hardware technology that must be deployed and generally lengthen the life of the Internet address space.

6.2.1. Current Growth

The one thing everyone agrees on is that growth continues. Had it been left unchecked (e.g. by CIDR, controls by registry allocation and provider annoucement policies), it would have already quickly outstripped Moore's law. Even now, with checks in place, growth is uncomfortably fast, since providers have to continually upgrade memory sizes and processors in Internet backbone routers. `Coping' in this way is not a good strategy and bears considerable long term risk.

6.2.2. Future Growth

The most recent backbone routers are within a factor of four of the most current processor clock rates and will continue to close quickly. However, the microprocessor industry is less motivated to develop faster processors as the desktop PC demand curve has flattened. If microcomputer industry continues to slow, router industry may actually have to drive the computing industry or have to use multiprocessor computers just to process the control plane. Neither is an appealing prospect, either from a technology point of view or from a robustness perspective. Furthermore, it seems unlikely that the router industry will be able to drive the computing industry: not even the entire router market is large enough to sway the microprocessor industry to accelerate R&D sufficiently.

6.3. Aren't Routing Vendors Tackling these Problems?

The ability to advertise multiple prefixes with common attributes as part of a single BGP update message is a known capability and is not overly difficult. However, the further ability to take that set of prefixes and perform some form of proxy aggregation is not currently a well-understood capability. This functionality would have great returns as it would decrease the prefix loading behind the domain that is performing the atomicity/aggregation computations. A key point to remember is that today, many domains advertise unnecessarily specific prefixes to effect the optimal local routing policy. This research points to a way to recapture the hierarchical topology that the providers originally deployed. This would return a multiplier far greater than the factor of two that has previously been discussed. Further, while such techniques have been discussed in the hallways of the IETF, they have never been fully analyzed and put into practice. There are undoubtedly issues here that will need to be addressed and pushing the research in this direction would hold great promise.

Also note that vendors tend to pay attention to research work more if a prototype implementation and some measurement on it has been done. For example, when Van Jacobson came up with the RED idea (Random Early Detection) for congestion management, he implemented it in the NS simulator and had data for the router vendors to look at. They took his code and algorithms and implemented it. If it had just remained a technical report or a SIGCOMM paper it would not have made it into our Internet.

Next: 7. Related Ideas