You'd have thought that, by now, we would have settled the WLAN architecture wars. You'd think one architecture or another would have been declared the winner and that we'd all be buying basically the same product from a small number of vendors - just as is the case with Ethernet switches.

That hasn't happened yet, and may not for some time, in part because you can't really compare wireless LAN architectures directly without quantifiable results of what a given architecture/implementation combination can do. And such results are very hard to produce.

When Symbol Technologies (now part of Motorola) announced the Mobius Axon wireless switch back in 2002, I wrote that the switched (or centralised) architecture would become the most influential in wireless LANs for the foreseeable future (It's true - here's an old link). And that's what happened. Vendors such as Airespace, which was acquired by Cisco, Aruba, Chantry, which was acquired by Siemens, Extricom, Legra, Meru and Trapeze were all formed around the idea of a centralised wireless LAN architecture.

The WLAN switch has largely evolved into a controller with intermediate switching allowed, but a direct connection between the access point (AP) and switch is still required in some cases. One can argue that such a Layer-2 approach actually helps performance, while the flexibility of a Layer-3 strategy eases configuration.

Which is the best approach remains an academic question - and that's not the only one. Other questions include:

  • Thinness of AP. What functionality belongs at the new network edge? Security? Are standardised protocols such as CAPWAP required? Should APs self-reconfigure in response to changing network conditions, like Motorola's new Adaptive AP products (descended from those first Symbol devices)?
  • Direct forwarding. Why should all the traffic between a client and an AP also have to be tunneled back to a switch on a controller? Shouldn't an AP be allowed to forward packets as expeditiously as possible? On the other hand, aren't security and control enhanced if we send all the traffic to the controller? What's another couple of microseconds moving traffic over a wire, as Keerti Melkote of Aruba has noted, if one has just spent milliseconds sending the data over the air? But won't the path to the controller - and, indeed, much of the rest of the network - require more bandwidth than in the direct-forwarding case?
  • Channel assignments. Should radio channels be allocated in the traditional cellular manner, or should all APs be on the same channel, as is the case with Extricom's and Meru's architectures?
  • The control plane. WLAN architectures are often described in terms of planes, describing where required functionality is located. The data plane moves data and the management plane handles policy, configuration, exceptions, and reporting. The control plane makes real-time decisions about how best to operate at a given moment. As with direct forwarding in the data plane, should control be distributed or centralised?

All these questions ignore the fact that traditional controller-less (or distributed) architectures also continue to sell quite well. Aerohive and Proxim have centralised management and mesh capability but are otherwise fully-distributed architectures. Cisco's Aironet product family, augmented by the fundamental enhancements implemented in their Wireless LAN Solution Engine (WLSE) controller is, at its heart, based on traditional APs.

All of the companies mentioned in this column report strong sales and that they are taking advantage of the fundamental expansion of the WLAN industry. The knee in the hockey-stick demand curve has apparently arrived, perhaps thanks to 802.11n, and system architecture remains perhaps the primary differentiator for buyers.

It's important to point out, however, that wireless-LAN architectures can only be truly evaluated by their real-world behaviour in specific installations, and this is where things get complex. Benchmarking WLANs is difficult because of the fundamentally statistical nature of radio itself. Large-scale tests are not feasible, and reproducible benchmark results are usually not possible.

Moreover, current benchmark applications often don't represent the reality that needs to be modelled before the big installation takes place, although Veriwave's WiMix (no relation to WiMax) tool looks interesting. Nonetheless, the fundamental complexity of measuring real-world performance will keep the architectural debate going, with theoretical arguments playing a major role in purchasing decisions.

Craig J. Mathias is a principal with Farpoint Group, an advisory firm specialising in wireless networking and mobile computing. He can be reached at [email protected] This article first appeared in Computerworld.