Breaking vs. solving: analysis and routing of real-time networks with cyclic dependencies using network calculus

Distributed real-time systems in the aerospace domain require worst-case end-to-end latency analysis methods to provide certification evidence of the correct temporal behavior of critical traffic classes. One such analysis method is the Network Calculus framework. While the Network Calculus analysis is mature enough to be allowed in certification artefacts, it is only applied in networks where there are no cyclic dependencies between communication flows (so-called feed-forward networks). In general topologies, flows can form cyclic dependencies, making it difficult to prove the determinism of a network. There are two approaches to solve this problem: 1) breaking the dependencies in the routing algorithm to study a feed-forward network; 2) solving, i.e., computing the bounds, in the dependency. In this paper, we review the recent improvements of both methods and do a performance analysis of AFDX and TTEthernet networks to compare their impact on the worst-case delay and backlog bounds. Results show that the best method depends on a number of parameters, such as load and dependency length. Using these results, we propose a new routing methodology resulting in the lowest bounds for networks with cyclic dependencies.