Living on the Edge

As use of networks continues to grow and change, carriers will have to re-think their congestion control strategies

The explosive growth in customer access capacity - combined with the parallel buildout of core/backbone capacity through increased fiber deployment and transport technologies such as ATM and dense wave division multiplexing - is creating a serious pressure point at the edge of carrier networks. High volumes of many different applications will create a bottleneck at the network's edge. If carriers can't devise an effective strategy for dealing with this increased congestion and complexity at the network edge, they simply won't be able to capture the potential revenue and market share associated with next generation network applications.

Of course, enterprise customers, dot-coms, application service providers and other network users also have a stake in this game. If they want to use or provide these next generation applications, they also need to ensure that the problems associated with high volume, multiservice transport are resolved. It is instructive to look at how carriers and customers are approaching the problem - and how their different approaches may offer clues about how life at the network edge really can be improved.

A pipe dream?

On the carrier side, proposed solutions based on industry-standard quality of service (QOS) have dominated the congestion/control discussion for years. The operative theory is that by defining various classes of traffic and by enabling network devices to detect and route or switch that traffic in accordance with pre-defined policies, networks could be empowered to deal with many diverse application flows.

On paper, this approach seems logical. Set voice at a high priority so it doesn't get hit with quality-destroying latency. Set HTTP sessions at a low priority because no one cares if it takes a Web page an extra second or two to load. Problem solved.

Well, not really. The idea of standardized QOS originated when no one imagined that networks would support as many different services as they do today. Networks aren't just about voice, data and video. They're about thin- and thick-client applications. They're about consumer Web browsing and mission-critical Web-based transactions. In fact, as the business of networking has evolved during the last decade, it is evident that - in addition to a multiplicity of applications - carriers probably will end up selling multiple classes of service for each application. In other words, carriers need to prioritize and differentiate traffic flows that are an order of magnitude higher than anything anticipated when ATM QOS or DiffServ for IP were formulated.

Figure 1 shows how network congestion occurs. At the edge, customer requests for high-volume data streams can exceed edge capacity. In this case, local access links are not congested, but the network cannot support the total incoming edge data, slowing down service to customers, or a single customer may request too many high-volume data streams, slowing down all of that customer's services.

In addition, the networks themselves have become far more complex. Applications-based QOS was conceived at a time when everyone thought the core of the network would be the most critical congestion point. The network core is comprised of a relatively small number of hierarchically similar devices. So administering prioritization policies across those devices is not a highly intensive task, but trying to manage those same types of policies across multivendor, multi-tier access and core devices is.

Another flaw in the QOS approach is that it doesn't account for link capacity, meaning it doesn't help to prioritize traffic types A, B and C over types X, Y and Z; if the edge connection is already saturated by the volume of type A and B traffic, C still won't get through. In addition, technical requirements and business priorities often are two different things. Technically speaking, voice traffic may theoretically require better performance than an extensible markup language (XML) transaction. But if that XML transaction is worth $1 million, then it should take precedence over a call to the deli for take-out.

This last example dramatizes a basic fact of networking in the new millennium. It's not about protocols anymore. It's about content and customer priorities. And QOS simply isn't designed to prioritize traffic based on its content-based business value.

Despite years of work by the finest minds in the industry, there still isn't such a thing as an industry QOS. Many technologies today were not even imagined five years ago. Although it has been discussed for as many years, end-to-end standardized QOS across the Internet is still just a dream. Can network service providers really afford to postpone implementing next generation business models while standards bodies continue to try to convert lead into gold?

Here today, here tomorrow

While carriers and standards bodies wrestle with these issues, dot-coms and other value-added content providers already are moving ahead and making money on the Internet. They're doing this by delivering content, information, streaming media, transaction services and other things for which consumers and business customers are willing to pay. And they're doing it without waiting for network service providers to figure out QOS or any other enhancements to content delivery. As a result, there's no reason for them to share any of their revenue with their network service providers - because they don't perceive those providers as offering any significant value-add. In fact, the only thing they pay their providers for is the commodity of bandwidth.

It's interesting to note that, in the absence of QOS controls on the carrier network, these content providers have developed their own strategies for managing traffic and resource utilization. One of the main elements in these strategies is load balancing, or "virtual resource management."

Load balancing can be contrasted with QOS in many dimensions. For one thing, it's situated on the customer premises, close to the servers and storage areas that house content. For another, it addresses systems and software bottlenecks - not just network bottlenecks.

But perhaps the most intriguing aspect of load balancing from the carrier perspective is that it's content- and user-aware. That is, load balancers today don't just look at a group of similar CPUs and direct Web page requests to the least-burdened server. They can look at incoming packets and determine whether the browser is looking for a static page or is about to click a "buy" button. They can determine whether a request comes from a casual user or a person who subscribes to a premium level of service. In fact, by classifying URLs and IP address and using a variety of other mechanisms, premises-based load balancers have become highly sophisticated, content-driven traffic management platforms.

Take a look at how successful dot-coms are designing their technology infrastructures. A number of questions come to mind: What do dot-coms know that network service providers don't? How can service providers apply this knowledge technologically? How can such a technological solution lead to an improved business model?

One thing content providers clearly understand is that content is king. The Internet isn't about HTTP or FTP or encapsulated Systems Network Architecture. It's about financial reports and music and transactions. People aren't paying for packets. They're paying for the content that those packets contain. So if the infrastructure operates in a content-blind mode, then the carrier operates in a revenue-blind mode, too. Any traffic management solution that isn't content-aware can never support a next generation business model.

Content providers also have demonstrated that to effectively manage traffic on a content-aware basis, a device must be at the network's end-points. For dot-coms, those devices are the load balancers at the headend, and the browser (which is turned into a provider-operated device through the use of cookies) at the subscriber end. Content-based management simply can't be done through the use of arbitrarily defined classes of service as formulated by a group of standards gurus locked in a room somewhere. Content classification, in fact, is unique to every content provider and actually defines each content provider's unique value proposition. An access device is needed on-site to directly link content resources to seamlessly integrate the traffic management platform with those resources.

Does this mean that QOS can't play a role in enabling carriers to support and prosper in the new digital economy? Not at all. As noted earlier, traffic must be prioritized - especially at the network edge. Carriers, therefore, must develop QOS/prioritization solutions that enable them to support large volumes of highly differentiated traffic flows without having to overbuild the network - and burn vast amounts of capital in the process.

But these next generation prioritization solutions must meet a number of criteria (Figure 2). They must be content-based, include a premises-based component, include an edge device that can implement content-based policies and include a policy-based management console that can effective manage and coordinate premises, edge and core infrastructure devices.

It's important to note that such a platform also must support industry-standard QOS mechanisms. This is essential for ensuring prioritized delivery over Internet infrastructure that's not under the direct control of the carrier - which is likely to implement some combination of MPLS, DiffServ and the like. Standards support is also necessary for leveraging the carrier's own existing equipment, which is likely to include legacy routers, non-Layer 7 switches and other core-level devices.

There's another compelling argument for this type of content-aware platform: content-awareness also can be used for metering and billing. This has been demonstrated in the field to be a major value-add that network providers can offer to content providers.

In fact, if utilization of differentiated services can be tracked, they also can be distributed. That's a pretty good way to make money. The cable industry offers an economic model for this. Consumers don't pay money to the Cartoon Network for the right to watch the Powerpuff Girls. They pay their cable company, which acts as a distributor of the content. Network providers, by contrast, carry "pay-per-click" content every day and never see any of the resulting revenue. With a content-aware platform, that can change (Table 1).

The network edge is going to become congested quickly during the next few quarters. It's a problem that carriers will have to address without delay. But, in their haste to relieve that congestion, they should be careful not to implement solutions that solve the technical problem without solving the deeper underlying business problem - how to make money carrying packets.

That money isn't just going to be made by adding capacity and intelligent traffic management. It's going to be made by content awareness.