Estimating Network Link Characteristics using Packet-Pair Dispersion: A Discrete Time Queueing Theoretic View

TL;DR

This paper introduces a queueing theoretic framework for estimating network link characteristics using packet-pair dispersion, enabling more accurate end-to-end measurements and parameter estimation in network tomography.

Contribution

It develops a novel queueing model and estimation algorithm for link parameters from end-to-end packet dispersion data, addressing non-stationarity in arrival processes.

Findings

Estimation algorithm performs well for two queues in series.

Effective in multicast scenarios.

Works under non-stationary arrival conditions.

Abstract

Packet-dispersion based measurement tools insert pairs of probe packets with a known separation into the network for transmission over a unicast path or a multicast tree. Samples of the separation between the probe pairs at the destination(s) are observed. Heuristic techniques are then used by these tools to estimate the path characteristics from the observations. In this paper we present a queueing theoretic setting for packet-dispersion based probing. Analogous to network tomography, we develop techniques to estimate the parameters of the arrival process to the individual links from the samples of the output separations, i.e., from the end-to-end measurements. The links are modeled as independent discrete time queues with i.i.d. arrivals. We first obtain an algorithm to obtain the (joint) distribution of the separation between the probes at the destination(s) for a given distribution of the spacing at the input. The parameter estimates of the arrival process are obtained as the minimizer of a cost function between the empirical and calculated distributions. We also carry out extensive simulations and numerical experiments to study the performance of the estimation algorithm under the fairly `harsh' conditions of non stationarity of the arrival process. We find that the estimations work fairly well for two queues in series and for multicast.