Available bandwidth estimation and measurement-based admission control in IP networks

Abstract

The capacity of core networks has increased tremendously due to recent technology development in optical transmission and high-speed router/ethernet switches. However, IP networks originally designed to provide best-effort services can not guarantee strict or statistical quality-of-service (QoS) requirements for real-time traffic flows because resources are not reserved and all packets are treated equally in most nodes. Thus, it is very important to monitor network status and manage network resources in order to guarantee QoS for flows with real-time performance requirements. Since the delay performance strongly depends on the available band-width of the path among many network resources, this dissertation is concerned with monitoring the available bandwidth and proposing an admission control scheme of internet flows based on the estimated available bandwidth. First, a new mechanism is proposed to estimate the available bandwidth of a queueing system, whose service rate and the load of input traffic are not known in advance. In order to estimate the available bandwidth, we propose a probing method called a minimally backlogging method and propose two statistics. The first statistic is based on the delay of each probing packet and the second statistic is based on the amount of probing packets served in a specific time interval. We first show that an M/G/1 queueing system is stable when probing packets are sent to the system according to the minimally backlogging method. We also show that the available bandwidth can be estimated by using either of the two statistics if the probing packets are sent to the queueing system by the minimally backlogging method. Especially, the second statistic can be used to estimate the available bandwidth of a G/G/1 queueing system. We apply the theory developed for a single server. in order to estimate the available bandwidth for a local server as an application. The accuracy of the two proposed statistics is evaluated numerically ...