title: The View From Flatland

The View From Flatland

In 1883 Edwin Abbott published "Flatland; Une Aventure a Plusieurs Dimensions" ("Flatland; An Adventure of Many Dimensions") about life in a world of two dimensions. (Two, if you don't count time as a dimension.) It is quite limiting to live in a world of X and Y but no Z. ( http://ofcn.org/cyber.serv/resource/bookshelf/flat10/ )

From the way that some switch sales people talk, one could imagine they were recruited from Flatland. They keep talking about "flattening" the network, by which they mean removing routers. For example, a few years ago I talked to the president of a company building Ethernet to FDDI bridges. (The same box now is called their Ethernet to FDDI switch.) I suggested that he might want to look into adding some routing functionality to his box. He replied that since there was only one router per site, it was not a big concern for him.

Routers are added to networks for a number of reasons. The main ones are managerial compartmentalization, security compartmentalization, and the segmentation of broadcase domains. These reasons are not eliminated in a frame switched environment. One additional reason for routers, traffic segmentation, is effected by a migration to a frame switched infrastructure.

Flattening a network by combining subnets should be done carefully with a clear understanding about what caused the separation in the first place. If a particular LAN in a University network was set up to interconnect the administrative and financial computers it might not be a real good idea to combine it with a LAN that services the undergraduate dorms.

If management structures are set up to assign addresses and set up filters based on many small LANs, they will have to be revised if some of these LANs are to be merged. This is particularly true if the address assignment is delegated to a local support organization for each LAN.

It should be specifically noted that if the LANs support TCP/IP (increasingly likely) that the number of nodes per subnet may become important. Most new address assignments in the Internet are in "C-sized" chunks from the part of the IP address space known as the Class C space. In theory one can assign two adjacent C-sized assignments to the same LAN and get a single address space that could support 510 hosts but theory is ahead of fact in many cases. If the C-sized address assignment is from the Class C space (as all currently are) many hosts will assume that the LAN has been assigned a single Class C address and can support at most 254 hosts. All too many hosts can not do "supernetting" and combine two Class C nets. There are ways around this by having a router which understands that there are more than one subnet on the same wire but this is at best messy.

I do find it a bit inconsistent that many of the same sales people that want you to flatten your network sing the praises of virtual LANs. Somehow they seem to think that the arguments they make in favor of fake LANs do not work with real LANs. But I will admit to having a long-term inability to understand the logic behind sales-speak.

By the way, most of the arguments about a router being a performance problem and a bottleneck are a bit hard to reconcile with the wire-speed performance of many of the routers one might use. In most environments latency is far less of an issue than switch vendors would have you believe, and even then the difference in latency between a modern router and a switch is meaningless when considering overall application-to-application latency.

Bottom line, think before you assume that pulling routers will produce an easier to manage, faster and more secure network.

disclaimer: Harvard does have a copy of the 1883 "Flatland; Une Aventure a Plusieurs Dimensions" but that does not necessarily imply that Harvard is two dimensional, even if it were, these are my opinions.