On the use of supervised machine learning for assessing schedulability: application to Ethernet TSN

In this work, we ask if Machine Learning (ML) can provide a viable alternative to conventional schedulability analysis to determine whether a real-time Ethernet network meets a set of timing constraints. Otherwise said, can an algorithm learn what makes it difficult for a system to be feasible and predict whether a configuration will be feasible without executing a schedulability analysis? To get insights into this question, we apply a standard supervised ML technique, k-nearest neighbors (k-NN), and compare its accuracy and running times against precise and approximate schedulability analyses developed in Network-Calculus.
The experiments consider different TSN scheduling solutions based on priority levels combined for one of them with traffic shaping. The results obtained on an automotive network topology suggest that k-NN is efficient at predicting the feasibility of realistic TSN networks, with an accuracy ranging from 91.8% to 95.9% depending on the exact TSN scheduling mechanism and a speedup of 190 over schedulability analysis for 10^6 configurations. Unlike schedulability analysis, ML leads however to a certain rate "false positives'' (i.e., configurations deemed feasible while they are not). Nonetheless ML-based feasibility assessment techniques offer new trade-offs between accuracy and computation time that are especially interesting in contexts such as design-space exploration where false positives can be tolerated during the exploration process.