ML-based Video Streaming QoE Modeling with E2E and Link Metrics

An increasing variety of network services and applications have led to a demand for service specific network management. QoE routing, routing service traffic on an individual basis, has been applied to target this demand. Learning tools have also been applied to automate and tailor management approach in real time within the network. A network manager can evaluate routing decisions to determine if expected performance was reached, and make adjustments to the routing model if not. The difficulty with this approach remains in collecting and evaluating the network state and service performance in real time to enable learning in the network. Such metrics must also be suitable for developing or adapting a routing model. This paper expands on a framework for real time feedback supported management. The aim of the paper is to identify and evaluate a suitable real time mechanism to collect network state data and a suitable application feedback metric. The metrics are evaluated for use in a routing model. The solution is unique as it provides a framework for a general service given a suitable feedback metric for that service. The paper examines application KPI metrics as suitable feedback metrics for two services, video streaming and VoIP, with APSNR and PESQ used as respective feedback metrics. The paper defines and evaluates link metrics as a mechanism for real time state determination. The framework is implemented and evaluated on an emulated SDN testbed.