Task Affinity-Aware Scheduling for Multi-Core Edge Devices in Autonomous Vehicles

Abstract

The evolution of autonomous driving systems toward cooperative perception has significantly intensified the computational workload on in-vehicle multi-core edge computing devices. Existing real-time scheduling strategies often fail to adequately consider task affinity when dynamically allocating heterogeneous core resources. When dealing with highly concurrent safety-critical tasks, this strategic limitation easily triggers frequent cross-core task migrations and severe cache invalidation. Consequently, this leads to a decline in overall system throughput and uncontrollable end-to-end response latency. In light of these complex challenges, this study draws cross-disciplinary inspiration from communication networks. Specifically, it adapts the scheduling logic that uses directed acyclic graphs to predictively avoid physical occlusion and reconstructs it into a mechanism for avoiding cache affinity disruption within the computational domain. Building upon this foundation, this paper proposes a forward-looking task affinity-aware scheduling framework. By deeply mining the temporal evolution characteristics of future computational workloads, this framework accurately maps the multidimensional decision space of multi-core resource allocation into a directed acyclic graph model. It continuously solves for a near-optimal scheduling path along the time series to maximize the net effective computational volume. Complex co-simulation tests preliminarily demonstrate that this mechanism effectively suppresses unnecessary computational jitter under extreme system load scenarios. Furthermore, it provides a solid underlying guarantee for enhancing the braking response safety of autonomous vehicles. This exploration provides a strong basis to further consider the broad theoretical boundaries and potential collaborative optimization directions for multi-core edge scheduling within future advanced autonomous driving architectures.