InformationWeek

Last fall, a number of network vendors unveiled plans to offer Gigabit Ethernet switches for twisted-pair wiring. The first such products are starting to make their way to market, although given our recent tests, network administrators may want to give the vendors another six months to work out the kinks.

Gigabit Ethernet over twisted-pair cabling promises to reduce the cost and complexity of high-speed networking. In theory, companies should be able to run Gigabit Ethernet over standard Category 5 (Cat 5) twisted-pair wiring. But our recent tests of gigabit switches from Alteon, Intel, and Lucent Technologies, conducted at the Advanced Network Computing Lab of the University of Hawaii, discovered a number of problems with the new standard. Cat 5 cabling proved marginal for gigabit speeds. Category 5 enhanced (Cat 5e) is the recommended specification because of its superior ability to protect traffic from signal degradation. Many existing business connections will need to be reterminated and, in some cases, new cabling installed to support the new gigabit switches. Conversely, recent innovations in fiber-optic cable, such as 3M Corp.'s Volition, may make fiber cheaper to install and faster to terminate than Cat 5e.

We were originally slated to test five new switches, but several vendors failed to participate, indicating problems with their first-generation products. Of the vendors that did take part-Intel and Lucent-only Intel brought production equipment. Lucent brought preproduction blades for their existing Cajun P550 gigabit switch.

We also experienced problems getting the gigabit copper network interface cards (NICs) to work, because of immature software.

But our test results weren't all negative. Once we got through the problems caused by the immature technology, we found that the products worked as well as the optical-fiber versions of the switches, and they should cost less and offer higher port densities as the products evolve.

The principle of the testing was pretty straightforward: capture traffic from the University of Hawaii's production network and simulate that traffic, using traffic generators, including the AX/4000 from Adtech, the Ixia 1600 from Ixia Communications, and the SmartBits 6000 from Netcom Systems.

We also used Chariot from Ganymede Software Inc., which generates traffic using PC workstations. It generates application-level traffic rather than just packets, but is limited in the amount of traffic it can generate by the systems it is running on and the total number of workstations available.

We also used video encoder/decoders (codecs) running a video stream across the network to provide a real-world test of prioritization.

Our first step was to run Cat 5e cabling throughout the lab, verifying it with the Fluke DSP4000. The existing Cat 5 patch panels were retained, but proved to be the most marginal part of the network. Administrators hoping to use existing wiring for Gigabit Ethernet should test it carefully using a gigabit-ready cable tester before they make any plans. The test specification called for four devices: two core switches and two edge switches. The core switches had at least six copper Gigabit Ethernet ports, trunking and failover backbone capability, and four 10/100 ports. The edge switches were required to have at least 12 10/100 ports, plus a Gigabit Ethernet uplink port.

The traffic load tests simulated data streams consisting of a mix of 3-D graphics remote rendering, large amounts of Web traffic, some SAP application traffic, terminal emulation, and file transfers, in addition to high-priority video streams. The amount of data was intended to reflect a total user population of about 30,000 users at the primary campus and 10,000 users at the remote campus.

The first test was a load test using Chariot to simulate traffic at relatively low overall network load. We then used the traffic generators to bring the total network load up to the capacity of the switches, looking for dropped traffic when loads were high. After that, we started video streams, using the Minerva codecs to play a DVD movie across the network. We then oversubscribed the backbone links, using the traffic generators to create more traffic than the switches could handle, to see if they could protect the video traffic during a simulated network overload.

While this might seem like an excuse to watch movies, it's surprisingly easy to detect network problems by observing video quality. We caught problems that didn't show up any other way.