Controlling oversubscription.

The secret to Internet quality of service is out.

With all of the buzz surrounding the topic of Internet quality of service (QoS), it is easy to overlook some of the basic principles for providing what is considered to be a true quality of service. Debates about standards and protocols, such as differentiated services (DiffServ), multiprotocol label switching (MPLS), and resource reservation protocol (RSVP), sometimes divert attention from more fundamental issues such as oversubscription. Oversubscription occurs when a network builder consciously overbooks network resources. It is the way in which service providers manage the delivery of services in a world where it isn't viable to build a network that would support every possible user at his peak level of usage at all times. Properly managing oversubscription is a means to advance the state of Internet QoS.

Within some statistical bounds, bandwidth oversubscription is a legitimate way to increase network use without necessarily compromising the quality of the service. In the circuit-oriented world of voice or Frame Relay/ATM, there has always been some level of end-to-end provisioning implicitly associated with each customer connection. Typically, the level of service guarantee is inversely proportional to the degree to which resources are oversubscribed (i.e., heavily oversubscribed resources usually yield poorer service).

When assessing service quality, it is important to understand what metric is typically being used to calculate the level of oversubscription. Is it the ratio of potential users to actual connected users? Is it the bandwidth offered to connected users vs. the bandwidth actually used? Is it the ratio of the available peak bandwidth to minimum guaranteed bandwidth? For the voice network, the number of possible users far exceeds the number of actual users, which equates to a high oversubscription ratio. However, once the user is connected, the amount of bandwidth available is equal to the amount of bandwidth used, which equals a low oversubscription ratio. This is reflected in the service profile of a voice network. While you may sometimes get a busy signal, when you do get through, the call is high quality.

Now enter the connectionless world of IP where there is no implicit end-to-end provisioning associated with user access to the network. As a result, the oversubscribed Internet has been somewhat uncontrolled. You just show up and take your chances. Performance issues have been addressed in a reactionary way: when a crisis occurs at a network hot spot, it receives attention.

AOL was the first to achieve national visibility for an excessive subscriber-to-dial-port oversubscription ratio. The company embarked on a "$19.95 unlimited access" promotion which boosted the number of dial-up subscribers and lengthened call-hold times to a "near death" busy-signal experience. Fortunately, AOL was able to respond by rolling out an enormous number of additional dial ports and server resources. This brought the subscriber-to-dial-port oversubscription ratio back into the green.

So, just how is Internet QoS going to improve? The clever folks who brought you the commercial Internet in the first place--the Internet service providers--are busy analyzing and understanding which oversubscription metrics are important to the efficiency of their networks and the service quality that they deliver.

Metrics now include number of dial subscribers vs. number of dial ports in a local calling area, speed of access lines into the PoP vs. speed of the PoP uplink to the backbone, and--ultimately--the speed of the backbone in relation to local, national, and international traffic patterns. Managed Web sites are also monitored to understand the simple statistics of supply and demand. Got a hot Web site? Be prepared to either increase the capacity of the site (i.e., more bandwidth, faster servers, load balancing) to better accommodate the demand or accept the penalties of slower response times.

What is the lesson we can we learn from this? Internet QoS is not really a mystery but more a conscious trade off between supply and demand. Perhaps when it comes to QoS, there is no substitute for starting with a good network design. When combined with other specific techniques, such as traffic classification, aggregation, and admission controls, we can selectively engineer the network to achieve the desired level of QoS. Network providers are mastering subscriber- and traffic-growth trends and are responding by balancing their investments in not only more bandwidth but also in a new generation of products and technologies that will enable them to better control the Internet services they offer. While standards, such as DiffServ and MPLS, are important to simplifying the definition and delivery of QoS across the Internet, they must be combined with a solid understanding of how best to trade among cost, performance, and quality-of-service requirements.

Stephenson is chairman of Xedia Corp., Littleton, Mass.