ML-based Application Performance Modelling over a SDN Network Using End-to-end and Link Metrics

Abstract:
As higher volumes of new applications are dependant on the network for operational performance, network orchestration must manage each traffic type to maximise application performance. Application specific routing requires identification of traffic type, and subsequently a traffic model to best provision for this traffic. This paper presents an analysis of a link-based modeling approach to application performance prediction in the network to achieve fine-grained application specific Traffic Engineering (TE). Using a Software Defined Network (SDN) framework to collect network metrics and manage traffic flows, Machine Learning (ML) tools are applied to model application performance in the network. The performance model is used in network path selection on an application specific basis. Application performance is measured using application specific KPIs which are machine measurable and can be delivered as a metric to the network manager. Used together with network Quality of Service (QoS) metrics a network performance model for an application is determined. This paper identifies a link-based method for dynamic modeling and performs analysis comparing to End-to-End metric modeling. The paper investigates the trade-off between the two types of models to leverage the efficiency of applying the network metrics for modeling the application’s performance. Using VoIP as a use case, the PESQ quality indicator as a performance KPI and an SDN framework to collect and manage network metrics, a testbed is introduced for the purposes of developing and testing application performance modelling. Link and End-to-End metric performance modelling techniques are compared using this testbed.