Using Machine Learning and In-band Network Telemetry for Service Metrics Estimation

Abstract

Data plane programmable devices used together with In-band Network Telemetry (INT) enable the collection of data regarding networks’ operation at a level of granularity never achieved before. Based on the fact that Machine Learning (ML) has been widely adopted in networking, the scenario investigated in this paper opens up the opportunity to advance the state of the art by applying such vast amount of data to the management of networks and the services offered on top of it. This paper feeds ML algorithms with data piped directly from INT - essentially statistics associated to buffers at network devices’ interfaces - with the objective of estimating services’ metrics. The service running on our testbed is DASH (Dynamic Adaptive Streaming over HTTP) - the most used protocol for video streaming nowadays - which brings great challenges to our investigations since it is capable of automatically adapting the quality of the videos due to oscillations in networks’ conditions. By using well established load patterns from the literature - sinusoid, flashcrowd and a mix of both at the same time - we emulate oscillations in the network, i.e., realistic dynamics at all buffers in the interfaces, which are captured by using INT capabilities. While estimating the quality of video being streamed towards our clients, we observed an NMAE (Normalized Mean Absolute Error) below 10% when Random Forest is used, which is better than current related works.