Edge Scheduling Framework for Real-Time and Non Real-Time Tasks

This paper presents a two-stage edge scheduling framework that maps the tasks of a real-time artificial intelligence (AI) application across a collection of edge computing resources. The first stage is global and it creates schedules with execution slots for tasks with real-time constraints. The second stage is local and it uses the schedules from the first stage and places non real-time tasks in the free slots. By creating global schedules for time-critical tasks, the two-stage design allows a group of such tasks to run in a coordinated manner across edge computers while providing the local autonomy to execute other tasks according to a local schedule. We implemented the framework over a heterogeneous collection of machines and measured its performance under different conditions. Results show that the two-stage architecture is better because the flexibility offered by the architecture can be used by the edge servers to obtain higher overall performance (i.e., increase the batch and interactive execution rates or deadline compliance rates).