A Hybrid Approach to Refine WCRT Bounds for DAG Scheduling Using Anomaly Classification

Motivated by the performance demands and stringent timing requirements of safety-critical systems like avionics and autonomous vehicles, research has focused on providing timing guarantees for the scheduling of Directed Acyclic Graph (DAG) tasks in multicore systems. The structural complexity and timing anomalies make this problem challenging. Existing methods bound the Worst-Case Response Time (WCRT) of tasks through static analysis, but these bounds are complicated, difficult to validate, and often remain pessimistic for many scheduling scenarios. Runtime intervention can be effective in eliminating timing anomalies and providing timing guarantees; however, it is ineffective for anomaly-free scheduling scenarios, leads to non-work-conserving schedules, and incurs additional overhead. This paper proposes a hybrid approach to identify timing anomalies in DAG scheduling scenarios within a system, providing tighter WCRT solutions. The static analysis first offers a sufficient anomaly test to directly identify some anomaly-free DAG scheduling scenarios. Leveraging a wide range of scheduling data collected from the running system or its simulator, we then apply a machine learning approach to train a binary classification model, achieving an accuracy of 99.5%. Identifying the anomaly status enables the application of more precise WCRT bounds for different scheduling scenarios, leading to improved system performance. Specifically, we shorten the WCRT bounds for anomaly-free DAG scheduling by an average of up to 21.58%, with a maximum reduction of up to 55.47% compared to the state-of-the-art method.