least cost, and declare that option "better." A key problem with this approach is that it is difficult to define exactly the needed level of service for an Internet connection. The requirements depend on the applications being run over the connection, but these applications are constantly changing. As a result, so are the costs of meeting the applications' requirements. Until about twenty years ago, human conversation was by far the dominant application running on the telephone network. The network was consequently optimized to provide the type and quality of service needed for conversation. Telephone traffic engineers measured aggregate statistical conversational patterns and sized telephone networks accordingly. Telephony's well-defined and stable service requirements are reflected in the "3-3-3" rule of thumb relied on by traffic engineers: the average voice call lasts three minutes, the user makes an average of three call attempts during the peak busy hour, and the call travels over a bidirectional 3 KHz channel. In contrast, data communications are far more difficult to characterize. Data transmissions are generated by computer applications. Not only do existing applications change frequently (e.g. because of software upgrades), but entirely new categories—such as Web browsers—come into being quickly, adding different levels and patterns of load to existing networks. Researchers can barely measure these patterns as quickly as they are generated, let alone plan future network capacity based on them. The one generalization that does emerge from studies of both local and wide- area data traffic over the years is that computer traffic is bursty. It does not flow in constant streams; rather, "the level of traffic varies widely over almost any measurement time scale" (Fowler and Leland, 1991). Dynamic bandwidth allocations are therefore preferred for data traffic, since static allocations waste unused resources and limit the...
