Inferring the Scheduling Policies of an Embedded CUDA GPU

Embedded systems augmented with graphics processing units (GPUs) are seeing increased use in safety-critical real-time systems such as autonomous vehicles. Due to monetary cost requirements along with size, weight, and power (SWaP) constraints, embedded GPUs are often computationally impoverished compared to those used in non-embedded systems. In order to maximize performance on these impoverished GPUs, we examine co-scheduling: allowing multiple applications concurrent access to a GPU. In this work, we use a new benchmarking framework to examine internal scheduling policies of the black-box hardware and software used to co-schedule GPU tasks on the NVIDIA Jetson TX1.