Enhancing Deadline Miss Early Detection for DAG Tasks with Variable Probabilistic Thresholds

Autonomous driving systems require rigorous real-time performance to ensure safety. The entire process, from sensor data acquisition to vehicle control command generation, must meet strict end-to-end deadlines. If a deadline miss occurs, the system must promptly transition to a fail-safe state. To enhance safety, this paper proposes a method for early detection of deadline misses. This method models the autonomous driving system as a directed acyclic graph (DAG) comprising timer-driven and event-driven nodes, assigning dynamic time constraints to each node based on the end-to-end deadline and remaining workload. Conventional approaches depend on worst-case execution time (WCET) assumptions, which often lead to overly pessimistic deadline miss detection by neglecting the variability in execution times. In contrast, the proposed method incorporates probabilistic execution times to model variability, dynamically adjusting thresholds based on workload. By progressively tightening thresholds for jobs closer to the exit node, the method balances detection sensitivity with false alarm reduction. Experimental results demonstrate that our probabilistic method not only reduces the pessimism of the conventional WCET-based early-detection approach, which corresponds to the current state-of-the-art configuration when the probability threshold is set to 1.0, but also generalizes this baseline by incorporating execution-time variability and workload-aware threshold adaptation.