UC Riverside

Real-time GPU scheduling plays a critical role in meeting the performance and timing requirements of modern cyber-physical and autonomous systems, where timely execution of critical tasks is essential. However, the ever-increasing complexity and heterogeneity of modern computing hardware have introduced new challenges in achieving resource and energy efficiency while delivering the required real-time performance. With the proliferation of GPU-accelerated applications and the rise of power-constrained environments, there is an urgent need to optimize resource allocation, minimize energy consumption, and successfully execute time-sensitive tasks.In this dissertation, we address the challenges in real-time GPU scheduling by proposing novel algorithms system-level solutions. Firstly, we investigate the trade-off between energy efficiency and real-time performance, developing novel approaches that dynamically allocate resources based on task characteristics. This optimization allows us to balance power consumption while meeting strict timing constraints. Secondly, we focus on energy-efficient real-time scheduling in heterogeneous multi-GPU systems. By considering the heterogeneity of GPU architectures and workload characteristics, we introduce strategies to improve both energy efficiency and real-time performance. Lastly, we develop systematic techniques to enable preemptive priority-based scheduling for real-time GPU tasks. Through the utilization of the proposed preemption techniques, we enhance resource utilization, improve responsiveness, and achieve enhanced schedulability in multi-core systems. Through this work, we make significant contributions to the field of real-time GPU scheduling, addressing energy efficiency, performance balancing, and preemptive scheduling challenges.