A novel approach to bottleneck analysis in networks

In this paper, we devise a novel method for bottleneck analysis of UDP networks based on the concept of network utility maximization. To determine the losses on the links in a UDP network, we propose an optimization problem (geometric program) for which we find and prove conditions under which it accurately determines the true losses. We further extend this analysis to stochastic rates using stochastic optimization techniques and provide a new metric to flag bottleneck links. This method does not rely on time-consuming packet-level simulations, but is instead based on robust mathematical models. Alternatively, one could determine the losses by solving a fixed point problem and extend it to random rates using a Monte Carlo simulation. However, lack of knowledge of convergence makes it difficult to predict the end of such simulations. Our method is more advantageous as it involves solving an optimization problem, the solution to which can be numerically determined to the desired accuracy. Also, compared to a black and white approach between worst-case analysis and average-case analysis, our method offers network managers the flexibility of choosing the shades of gray in between.