Data traffic is "bursty" in nature; it occurs in periods of intense activity followed by (sometimes long) periods of silence. If delivery of the data cannot be immediate, they may be stored in the network and delivered at a later time (how much later depends on the application). Data traffic is not tolerant of errors; any error passed to an application will cause problematic behavior of some kind. The above characteristics lead to the requirement for protocols to package the data, transmit the package through intermediate points, check for and correct errors, and deliver the package to the distant end. By contrast, voice communications are continuous in nature and have a real-time delivery requirement. A varying delay for a voice signal results in totally incomprehensible speech at the receiving end. However, voice signals are robust and tolerant of errors. Speech itself is so redundant that words remain comprehensible even at error rates of 1 in 100 in the digital bit stream. It is clearly more important for voice applications to deliver the signal on time than to deliver it with 100 percent accuracy.
These different requirements have led to different switching techniques for voice and data communications. Circuit switching sets up a path with a guaranteed bandwidth from end to end for each voice call. While this would also work for data communications, it would be inefficient since the reserved bandwidth would be unused most of the time. A technique known generically as packet switching was developed to effectively share transmission resources among many data communications users. When we refer to data networks in this paper, we are talking about networks employing some form of packet switching.
The final concept we need to explore is the concept of shared versus dedicated networks. In the early days of telephony, all customers used (shared) the public switched network for voice communications. The backbone of the public switched network was composed of large circuit switches located on the telephone companies' premises. By the 1960s, manufacturers began producing economical switches designed for use on customers' premises to lower costs for large users. While these were primarily for local service, it was soon discovered that large customers could connect these premises switches with private lines and create private networks for long distance voice communications. Public switched service rates at the time were high enough that these private networks were economical for large corporations (and the federal government).
As packet switched data networks came into being in the 1970s, the private network alternative was the only one available to customers. There was no public packet switched data network, nor was there a large demand for one. Private data networks grew alongside the private voice networks, with packet switches on customer premises and private lines connecting the switches. Computer processing technology limited the capacity of packet switches to that required by a single large customer; little penalty was paid in not locating switches on carrier premises where they could be shared by many customers.
In the 1980s, two forces converged to spell the end of the private voice network. Divestiture created a very competitive interexchange market, and computer-controlled switch technology evolved to the point where the partitioning of a large network in software became feasible. In this case, the large network was the public switched network of each interexchange carrier that was serving the general population. Over time, competition drove the unit prices being offered to a wide range of customers down to levels consistent with the large total volume. Volume ceased to be a discriminator for price beyond a level of one-tenth of the federal government-wide traffic. This service, which now dominates the voice communications market, is called virtual private network (VPN) service.
The current approach to data communications for large customers is still the private network. There are two shortcomings to this approach: economies of scale beyond a single user are never obtained, and the proliferation of switches on user premises does not further the development of a national infrastructure. Technology such as asynchronous transfer mode (ATM) switches and high-capacity routers is emerging that makes carrier-premises switching feasible. At the same time, initiatives in government and in the research and education communities are generating a large future demand for data communications. The consolidated demand of the federal government could create an infrastructure that pushes unit costs well up on the economy-of-scale curve. However, this will only be the case if a shared network is used to satisfy these requirements. We call this network, based on standard interfaces and protocols, the National Data Network (NDN). This paper presents the