Preemption Aware Task Scheduling for Priority and Deadline Constrained DNN Inference Task Offloading in Homogeneous Mobile-Edge Networks

TL;DR

This work tackles deadline-constrained DNN inference offloading in homogeneous edge networks by introducing a preemption-aware scheduler with two priority-aware algorithms. The system is implemented on a four-node edge testbed (Raspberry Pi 2) with a central controller that allocates time-slotted communication and computation, leveraging horizontal partitioning of YOLOv2 and throughput estimates to satisfy deadlines. Key findings show that preemption enables up to $99$ percent high-priority task completion and yields a $3$–$8$ percent increase in frames fully classified end-to-end compared with baselines, albeit with some reductions in low-priority task completions due to increased network saturation. The results demonstrate the practical viability of preemption-aware scheduling for edge DNN inference in homogeneous MEC, while highlighting hardware and networking bottlenecks that motivate future improvements toward more powerful devices and more efficient capacity estimation.

Abstract

This paper addresses the computational offloading of Deep Neural Networks (DNNs) to nearby devices with similar processing capabilities, to avoid the larger communication delays incurred for cloud offloading. We present a preemption aware scheduling approach for priority and deadline constrained task offloading in homogeneous edge networks. Our scheduling approach consists of two distinct scheduling algorithms, designed to accommodate the differing requirements of high and low priority tasks. To satisfy a task's deadline, our scheduling approach considers the availability of both communication and computational resources in the network when making placements in both the current time-slot and future time-slots. The scheduler implements a deadline-aware preemption mechanism to guarantee resource access to high priority tasks. When low-priority tasks are selected for preemption, the scheduler will attempt to reallocate them if possible before their deadline. We implement this scheduling approach into a task offloading system which we evaluate empirically in the real-world on a network of edge devices composed of four Raspberry Pi 2 Model B's. We evaluate this system under against a version without a task preemption mechanism as well as workstealing approaches to compare the impact on high priority task completion and the ability to complete overall frames. These solutions are evaluated under a workload of 1296 frames. Our findings show that our scheduling approach allows for 99\% of high-priority tasks to complete while also providing a 3 - 8\% increase in the number of frames fully classified end-to-end over both workstealing approaches and systems without a preemption mechanism.

Figures (10)

• Figure 1: System design
• Figure 2: Frame completion
• Figure 3: High-priority low-complexity completion
• Figure 4: Low-priority completion by mechanism
• Figure 5: Low-priority task completion per request.