CAPEST: Offloading Network Capacity and Available Bandwidth Estimation to Programmable Data Planes

Abstract

Measuring available bandwidth and capacity represents an essential requirement for a multitude of network applications spanning from traffic engineering and admission control to network security. Measurement techniques frequently presume to know capacity a priori, but this constitutes a weak premise in a number of modern scenarios due to conditions such as abstractions in infrastructure virtualization, dynamic demands in resource sharing and fluctuations in interference, all of which can affect capacity in short time spans. Despite consistent efforts, currently employed techniques struggle to balance accuracy, intrusion and freshness, depending on either substantial intrusion, onerous processing or unfeasible deployment. Recent developments on data plane programmability have breathed new life into this undertaking, allowing observation points to be more efficiently distributed and programmable packet methods to be executed in-situ. This paper proposes CAPEST, a passive capacity and available bandwidth measurement method for the data plane, employing packet dispersion and autocorrelation. The method is evaluated regarding its parametrization sensitivity, its intrusion and freshness in comparison to state-of-the-art techniques and its performance in the real-world application of video routing. CAPEST was found to incur substantially (80%) less intrusion and achieve 10% better accuracy, all the while providing an order of magnitude improvement in freshness.