Dimensioning bandwidth for elastic traffic

Abstract

In this paper, we discuss the issue of dimensioning Internet access lines for elastic traffic. This is important for Internet service providers (ISPs) because over-dimensioning wastes precious bandwidth resources, while under-dimensioning generally leads to less satisfactory quality of service (QoS) perceived by subscribers. Our discussion is based on the M/G/R processor sharing model which characterizes TCP traffic at flow level. Our analysis demonstrates the impact of a number of key factors (and their relations) on the dimensioning procedure. We consider two dimensioning methods based on different QoS criteria. It is found that the method based on the delay factor is superior in that both the average delay (throughput) and blocking performance targets can be satisfied. Numerical and theoretical analyses also illustrate that significant multiplexing gain can be achieved for elastic flows and this gain increases with burstiness.