On the Robustness of Deep Learning-predicted Contention Models for Network Calculus

摘要: The network calculus (NC) analysis takes a simple model consisting of schedulers and data flows crossing them. A number "building blocks" can then be applied to capture the without imposing pessimistic assumptions like self-contention on tandems servers. Yet, adding pessimism cannot always avoided. To compute best bound single flow’s end-to-end delay thus boils down finding least contention models for all in – an exhaustive search easily become very resource intensive task. literature proposes promising solution this dilemma: heuristic making use machine learning (ML) predictions inside NC analysis.While results work were terms quality computational effort, there is little no insight when prediction made or if trained algorithm achieve similarly striking networks vastly differing from its training data. In paper, we address these pending questions. We evaluate influence features accuracy, impact scalability. Additionally, contribute extension method by predicting n alternatives order increased robustness application outside Our numerical evaluation shows that good accuracy still achieved large although restrict are two orders magnitude smaller.