Predicting Quality of Delivery Metrics for Adaptive Video Codec Sessions

Abstract

Predicting video quality will continue to be an active area of research given the dominance of video traffic for years to come. Network service practitioners that are poised to handle the strain on the existing limited bandwidth constraints are better placed to be SLA-compliant. The dynamic and time-varying nature of cloud-hosted services require improved techniques to realize accurate models of the systems. To address this challenge: (1) we propose Codec-aware Network Adaptation Agent (cNAA), an online light-weight data learning engine that achieves accurate and correct predictions of quality of delivery (QoD) metrics, namely jitter for video services. cNAA achieves this prediction accuracy by leveraging the available network information in the face of congestion and adaptive codecs; (2) we highlight the short-comings of some baseline machine learning techniques that fail to capture network dynamics and demonstrate their failure in comparison with cNAA; and finally, (3) we demonstrate the efficacy of cNAA under varying network and codec conditions and provide evidence showing that machine learning approaches that incorporate network dynamics are better placed to realize accurate and correct predictions.