Task Partitioning and Orchestration on Heterogeneous Edge Platforms: The Case of Vision Applications
Dapeng Lan, Amir Taherkordi, Frank Eliassen, Lei Liu, Stéphane Delbruel, Schahram Dustdar, Yang Yang
Abstract
Running computer vision applications, such as 3-D simultaneous localization and mapping (SLAM), on mobile devices requires low-latency responses and a massive amount of computation. Edge computing has been introduced to move Cloud features closer to end users, providing necessary computing and network resources for end devices. The heterogeneous edge devices, with different hardware architectures (e.g., CPUs and GPUs) and runtime environments, provide diverse resources to support processing tasks from end devices, resulting in different costs and quality of services. How to partition these computing tasks and distribute them over these heterogeneous hardware nodes is still an open research question. Considering these inherently heterogeneous hardware architectures, new approaches for service orchestration and task scheduling are required to meet the service-level agreement and reduce the overall cost of the system (e.g., facility utilization cost). This article presents a system framework, EDGE VISION, for computer vision applications partitioning and orchestration on heterogeneous edge computing platforms considering both CPUs and GPUs. EDGE VISION abstracts the heterogeneous hardware resources and the task runtime environments and divides the application into separate tasks to be orchestrated and deployed into the heterogeneous edge nodes. We also propose two scheduling algorithms in our framework, minimum latency task scheduling and minimum cost task scheduling, aiming to minimize the processing latency and the overall system cost. We evaluate our framework by implementing the edge-based 3-D SLAM application in our real testbed with ten heterogeneous edge devices. Evaluations show that EdgeVision can efficiently minimize the processing latency and the system overall cost and achieve up to 30% decrease in task processing latency and 15% more cost saving compared to the State-of-the-Art baselines.