Response-Time Bounds for Concurrent GPU Scheduling

Graphics processing units (GPUs) have been receiving increasing attention in the real-time systems community as a potential solution for hosting workloads like those found in autonomous-driving use cases that require significant computational capacity. Allowing multiple programs to access a GPU concurrently can enable the GPU to be more efficiently utilized, if each individual program is incapable of occupying all GPU resources. In this work, we summarize the basic scheduling rules for concurrent GPU scheduling in NVIDIA GPUs. We define a task model for GPU scheduling based on these scheduling rules. In ongoing work, we are attempting to obtain response-time bounds for tasks under this model.