The Case for a Network Calculus Heuristic: Using Insights from Data for Tighter Bounds

Abstract

Deterministic network calculus offers a framework for providing guaranteed bounds on end-to-end delay and buffer usage in computer networks. Various network analysis methods have been proposed in order to reduce the impact of burstiness or multiplexing and provide tight performance bounds. Yet, the choice of which analysis method to use given a network to analyze is not straightforward as it has been shown in the literature that corner cases exist leading to poor tightness. We propose in this paper to take a new look at this question using insights from data and confirm that there is no clear winner when deciding which method to use. Based on those first results, we make the case for a network calculus heuristic in order to predict the bounds produced by a given network analysis method. Our main contribution is a heuristic based on graph-based deep learning, which is able to directly process networks of servers and flows. Via a numerical evaluation, we show that our proposed heuristic is able to accurately predict which analysis method will produce the tightest delay bound. We also demonstrate that the computational cost of our heuristic makes it of practical use, with average runtimes one or two order of magnitude faster than traditional analysis methods.