The Weakest Link Theory

By Kevin Tolly
Network World, 08/09/99

For as long as I've been testing boxes, vendors have been telling me that I need to consider all the elements that constitute the "end-to-end solution."

Now that I've launched a major research project on campus quality-of-service (QoS) products, some vendors are steering me to evaluate just a single box. ("Don't look at our ABC100, the XYZ400 is the one to do the job.") Allow me to respond by offering my Weakest Link Theory of product testing.

Simply stated, my theory posits that the weakest link in an end-to-end package - such as that needed in QoS - deserves special focus because it is the most likely to fail. An end-to-end QoS offering by definition includes not only powerful core switches, but also the humble edge switch. The capabilities of both need to be evaluated to consider the overall bundle.

A useful approach is to simply assume that the spare-no-expense core switches from different vendors will have roughly similar capabilities. Instead look at what is offered in the edge switch. It is in this price-sensitive area that vendors are likely to cut corners - jettisoning features usually means lowering cost and raising margins. Doing away with fancy QoS features could mean lower research and development costs, lower processing requirements and lower memory needs. Implementing sophisticated QoS support at the edge will likely raise all of the associated costs.

So how valid is it when a "solutions vendor" claims that the company's core switch, rather than its apparently flawed edge switch, should be the primary focus of a QoS study? I say that approach is invalid. Doesn't the performance of both boxes depend upon the sturdiness of the weakest link? If that weak link fails, the whole thing fails. If priority traffic gets clobbered at the edge because of a crude prioritization scheme, traffic will never make it into the sophisticated traffic handling mechanism of the core switch.

Even common sense tells you that edge behavior matters. Few core switch backplanes are oversubscribed whereas most edge switch configurations are.

Most edge switches deployed today are pure Fast Ethernet. The price/performance of Fast Ethernet has made it the de facto standard desktop technology. That fact means countless switches out there are configured with 24 Fast Ethernet client ports and a single Fast Ethernet uplink into the core. This amounts to a 24-fold oversubscription rate on the link into the core. It doesn't take much activity in this configuration to cause congestion and invoke the QoS features.

Some customers might aggregate several Fast Ethernet uplinks to serve as a fractional Gigabit uplink. In the best case, such a switch would be outfitted with a full Gigabit Ethernet pipe. While this improves the ratio dramatically, it still leaves the network manager with a possible 2.4 gigabits of edge traffic vying for a 1G bit/sec uplink.

Today, thanks to low prices for Fast Ethernet network interface cards and switch ports, congestion is more of a concern at the edge than in the campus core. Ironically, the QoS qualities of your inexpensive edge switches may have more impact on your end-to-end service than what takes place at the core.