On-line Packet Loss Rate Estimation

A key requirement for ease of migration from legacy to ambient networks is the elimination of dependencies between functionalities. Currently, in the case of designs for QoS and mobility in IP networks, it is apparent that there is a strong coupling between the two functions. One way to reduce this coupling is to remove the need to have QoS state within the network. This can be done if applications are able to manage the QoS parameters themselves. The basic idea is that the application is responsible for keeping the network in a congestion-free state in order to minimize the loss and delay experienced by the application, much as TCP does today for best effort traffic. On-line estimation of end-to-end packet loss can be used to monitor wireless link performance and to help adaptive applications to make better use of network resources. Existing work has focused on measuring and modeling packet loss in the Internet but most of these technologies do not address end-to-end path performance. This paper proposes an on-line stripe packet-pair probing approach to estimate the packet loss rate that applications may suffer. The approach uses 2-order Markov chain to model the loss rate and loss burstiness. To reduce the computational complexity, we employ maximum entropy to estimate the parameters. The paper validates the model via existing loss traces