Efficient Gradient-Based Network Calculus for Scalable Synthesis of Network Configurations

Standards aiming for a bounded networking delay are being developed and deployed at an unprecedented pace, in particular in industrial settings such as avionics. Simultaneously, modeling and analysis techniques are being extended to derive formally verified delay bounds for these networks. Network Calculus (NC) has seen much attention, including a first evolutionary step from an analysis method to a synthesis method called Differential Network Calculus (DiffNC). That is, while classical NC requires a fully specified network for model derivation and analysis, DiffNC can find delay-bound-optimizing settings for open configuration parameters. Yet, synthesis of parameter values increases the computational demand drastically such that computational efficiency becomes imperative. We implemented an entirely new tool chain for DiffNC, allowing for parallelization and offering a gradient-based nonlinear optimization algorithm more suited for that task. We demonstrate scalability by optimizing two orthogonal configuration options as well as application to the network of the Airbus A350. In numerical evaluations we show improved delay bounds by up to factor two and runtime reduction by one order of magnitude.