GNN-MiCS: Graph Neural-Network-Based Bounding Time in Embedded Mixed-Criticality Systems

In mixed-criticality (MC) systems, each task has multiple WCETs for different operation modes. Determining WCETs for low-criticality modes (LO modes) is challenging. A lower WCET improves processor utilization, but a longer one reduces mode switches, maintaining smooth task execution even with low utilization. Most research focuses on WCETs for the highest-criticality mode, with fewer solutions for LO modes in graph-based applications. This letter proposes GNN-MiCS, a machine learning and graph neural networks (GNNs) scheme to determine WCETs for directed acyclic graph applications in LO modes. GNN-MiCS generates test sets and computes proper WCETs based on the application graph to enhance system timing behavior. Experiments show our approach improves MC system utilization by up to 45.85% and 22.45% on average while maintaining a reasonable number of mode switches in the worst-case scenario.