A Solution to the Bandwidth Allocation Problem for the Internet

• Activity Number: 432
• Type: Invited
• Date/Time: Thursday, August 7, 2003 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistical Computing
• Author(s): William S. Cleveland*+ and Jin Cao and Don X. Sun
• Companies: Bell Laboratories and Bell Labs, Lucent Technologies and Bell Laboratories
• Keywords: nonlinear time series ; long-range dependence ; point process ; queueing ; internet traffic engineering ; quality of service

Abstract

The most basic problem of Internet traffic engineering is determining the bandwidth (b bits/sec) required to carry a traffic load (t bits/sec) offered to a single link and satisfy specified quality-of-service requirements for the traffic. The offered load is packets of varying sizes arriving for transmission on the link. Packets can queue up and are dropped if the queue length is bigger than a buffer size. For today's predominant traffic on the Internet, the applicable quality metrics are the queueing delay (d sec) and the packet loss (w percent). This bandwidth allocation problem has received much attention. While important insight has been gained, the problem has resisted solution due to a lack of comprehensive, validated statistical models for the packet arrivals and sizes. The required bandwidth depends on the queue-length process which, in turn, depends heavily on the statistical properties of the arrivals and sizes. Equipped with recently developed statistical models for arrivals and sizes, we develop a solution, a statistical model for t/b as a function of t, d, and w that is surprisingly simple given the complex time dependence of the traffic.