Online Task Scheduling for DNN-Based Applications over Cloud, Edge and End Devices.

As a combination of artificial intelligence (AI) and edge computing, edge intelligence has made great contributions in pushing AI applications to the edge of the network, especially in reducing delay, saving energy and improving privacy. However, most of researchers only considered the computation approach of end device to edge server and ignored the scheduling of multi-task. In this paper, we study DNN model partition and online task scheduling over cloud, edge and devices for deadline-aware DNN inference tasks. We first establish our mathematical model and find the model can not be solved directly because the solution space is too large. Therefore, we propose the partition point filtering algorithm to reduce the solution space. Then by jointly considering management of the networking bandwidth and computing resources, we propose our online scheduling algorithm to meet the maximum number of deadlines. Experiments and simulations show that our online algorithm reduces deadline miss ratio by up to \(51\%\) compared with other four typical computation approaches.