Multitask Scheduling of Computer Vision Workload on Edge Graphical Processing Units

The increasing urbanization in developing countries calls for more efficient and safer transportation systems. A key technique used to enhance such efficiency and/or safety is to utilize running of computer vision algorithms to identify obstructions that may come up, and notify vehicles in real-time. Such real-time detection and notification requires sufficient computation resources located logically and physically close to the cameras. While utilization of edge compute devices has been proposed in the literature, it is unclear how such devices with heterogeneous processing units can handle real-time detection while multi-tasking. In this work, we profile the performance of a few devices with embedded and desktop-quality GPUs, and show that the performance while multi-tasking can be modeled as a submodular function. We utilize this observation to model load-balancing of camera videos as an instance of a submodular welfare problem, and solve it using a greedy algorithm. Our extensive trace-driven simulations show that our technique outperforms the baseline by over 40%.