Understanding Interference Channels in NVIDIA GPUs

Recent advances in Artificial Intelligence (AI) and Graphics Processing Units (GPUs) have made it possible to build truly intelligent cyber-physical systems, such as autonomous vehicles. However, these systems are often safety-critical and are required to satisfy hard deadlines. Such hard real-time requirements rely on robust timing analysis of all tasks in the system, which requires bounds on worst-case execution times (WCETs) that are robust despite the many interference channels. Unfortunately, limited work has been done on the generation of safe WCET bounds for GPU-based real-time systems, as GPU architectures are complicated and different from CPU architectures. In addition, the details of the drivers and hardware of the commonly used GPUs (NVIDIA GPUs) are considered proprietary and lack documentation, which makes it even harder to understand the underlying interference channels. This work serves as the first step towards the understanding of interference channels in NVIDIA GPUs and obtaining safer estimations of WCETs for GPU-based safety-critical real-time systems. Through experiments and case studies on synthetic and real GPU applications, we identify potential interference channels and provide methods for consistently generating such interference in order to get safe WCET estimations.Bachelor of Scienc