Wireless LANs face huge scaling challenges

Philippe Hanset is wondering about the intersection of the Slingbox and the campuswide wireless LAN at University of Tennessee, where he's senior networking engineer.

A vendor has been industriously selling the Slingbox to incoming students, who set it up in their dorm rooms where they have cable TV service. The WLAN then lets them stream TV programs to their notebooks anywhere on campus. Imagine a hit like "American Idol" streaming wirelessly to hundreds of student notebooks.

"This could be challenging," Hanset acknowledges, with masterful understatement.

The development is typical of the new scaling challenges facing WLAN administrators as WLANs continue to grow in size, in number of users, and in more demanding applications. With big WLAN deployments, higher education is a kind of ground zero for many of these issues.

Early WLANs focused on growing the number of access points to cover a given area. But today, many wireless administrators are focusing more attention on scaling capacity.

That focus is a broad one, calling for a deeper understanding of what access points are capable of, and paying more attention to scaling back-end systems, servers and networks.

The rise of high-density WLANs

"We've been used to 20 to 50 wireless users in an area, with another 20 to 50 maybe 50 or 100 feet away," says Brad Noblet, a former college IT director, at Dartmouth and Harvard, who’s now an independent consultant, BN Consulting. The assumption: few users, who wanted just e-mail access or Web searching.

Low-density WLANs are giving way to high-density ones, with new challenges for network administrators. "When we first put this [WLAN] in three years ago, there were few wireless clients," says John Turner, director of network and systems at Brandeis University. "Now everyone has a laptop."

The scaling challenges include ensuring adequate wireless, and wired, bandwidth for the applications being served to wireless users. "These scaling issues are becoming more and more apparent where lots of folks show up and you need to make things happen," Noblet says.

What has to happen is that lots of clients have to associate with an access point, get an IP address, be authenticated, get enough bandwidth (wireless and wired) for their applications, and behave themselves as network citizens.

Page Break

Configure access points for capacity, not access

Noblet urges network administrators to configure access points for performance (or capacity), rather than for access. He's found some access points are configured without any limits on the number of client associations. If a large group of users coalesce around an access point, they'll find slow associations or none at all. "What it's really about is understanding the throughput performance of a particular data stream," Noblet says.

But everyone agrees that capacity planning at the level of the access point is more art than science. "When I speak on this topic, I always emphasize that we, the IT professionals, not the vendors are the ones who best understand the user and application scenarios we'll be dealing with in our deployments," says Dan McCarriar, assistant director of network services at Carnegie Mellon University (CMU).

CMU is about halfway through an 802.11n deployment using gear from two vendors, Aruba for academic areas and Xirrus for residence halls. Xirrus packs a WLAN controller along with four, eight or 16 Wi-Fi radios into a single oversized "smoke detector" like package, called an array, with sectorized antennas to prevent interference. The result lets CMU plug a single array into a high-density area, without having to do complex microcell planning and administration.

Keeping up with DHCP

In some cases, DHCP servers can't keep up with a flood of clients. "We're definitely seeing this," says Turner. But CMU's DHCP servers are able to keep pace. The key is designing the centralized IT infrastructure for these kinds of services, which are used by both wired and wireless clients, so it can scale quickly and easily.

In the future, Turner plans to create a more seamless mobile experience across the campus by tying location and mobility services into DHCP. "The DHCP server is not aware that someone has disconnected," he says. "We might be able to do something between the central WLAN controller and DHCP so we’re not holding addresses for people who are never coming back."

The University of Tennessee at Knoxville has run into a slightly different DHCP problem, says Hanset, from the school's network services group: Some returning student notebook PCs or "rogue" access points in dorms act as DHCP servers themselves, serving out useless DHCP leases to requesting clients. The school blocks these hosts at switch ports or the Aruba WLAN controller.

Page Break

Another scaling issue is that once clients are issued IP addresses, they may end up keeping them far longer than needed, so they can't be reissued to newly arriving clients on the same subnet. In some cases, addresses can be exhausted.

At the University of Tennessee, these address leases are limited to two hours, at Brandeis, to just 30 minutes. Both institutions make use of Aruba's virtual LAN pooling, which associates a pool of addresses to a given VLAN. It's an efficient and effective tool, says Turner, but network administrators still "have to think carefully about this."

Carnegie Mellon currently has a flat WLAN, essentially configured as one campuswide subnet, with one large pool of addresses. The university is planning to carefully segment the network, probably into several geographical zones, once the 11n rollout is completed. Then, address exhaustion could become an issue if not properly managed and monitored, says Scott Ambrose, CMU's manager of network design and development.

Ambrose plans to collect a mass of statistics on such things as average number of devices on the network and peak numbers of users, and the locations of the access points they associated with. All that data will go into planning the size and number of zones, and how to allocate the available IP addresses to each one.

Scaling for multimedia

Multimedia use is surging, and 802.11n is expected to make it surge still more. All these universities are configuring their wireless LANS for multicast support, to minimize bandwidth demands where possible. Users in effect tune into a single multicast stream (analogous to viewing broadcast TV) rather than each one receiving his or her own separate, unicast stream. "You have to look at your application and ask 'what am I trying to serve here?' " Noblet says. "That will dictate whether you have to make use of a unicast or multicast transport."

"We enabled multicast everywhere we can," says the University of Tennessee's Hanset. But with that, to further improve performance, the university also disabled the slowest WLAN data transfer rates, of 1M and 2Mbps. "So every broadcast packet is sent at 5.5Mbps," he says.

Test everything, says Carnegie's McCarriar. "Support for multicasting [by vendors] is all over the map," he cautions.

Page Break

11n is no silver bullet

These users are all adopting or evaluating 802.11n, which promises five to six times more throughput than existing 802.11abg WLANs. The additional capacity will be very welcome, but all agree that 11n won't eliminate the need for careful, thoughtful, end-to-end capacity planning.

"It guarantees faster speeds, but it doesn't prevent one bandwidth hog from taking most of that," says Hanset. "I'd like to have a 'fair-share' mechanism in the WLAN, so that can't happen, especially in high-density areas."

Hanset also cautioned that 11n throughput can be dramatically affected by inadequate upstream links and, in the University of Tennessee's case, power injectors. Tests showed an Aruba 11n access point delivering 80M to 90Mbps, but 160Mbps with a gigabit connection and gigabit-capable power injectors," he says.

In many migrations, enterprises simply are replacing existing 11abg access points with 11n devices. But if the 11n products are making use of the 5GHz frequency, because there are more channels and less radio noise, some tweaking will be needed, warns Brandeis' John Turner. That's because 5GHz signals don't propagate as well as those in the 2.4GHz band.

"We've seen where we've installed an 11n access point and expected great performance in the 5GHz band," he says. "But you walk behind two walls and your signal vanishes. It's not dead-simple."