Software Aging in a Real-Time Object Detection System on an Edge Server

Abstract

Real-time object detection systems are rapidly adopted in many edge computing systems for IoT applications. Since the computational resources on edge devices are often limited, continuous real-time object detection may suffer from the degradation of performance and reliability due to software aging. To provide a reliable IoT applications, it is crucial to understand how software aging can manifest in object detection systems under resource-constrained environment. In this paper, we investigate the software aging issue in a real-time object detection system using YOLOv5 running on a Raspberry Pi-based edge server. By performing statistical analysis on the measurement data, we detected a suspicious trend of software aging in the memory usage, which is induced by real-time object detection workloads. We also observe that a system monitoring process is halted due to the shortage of free storage space as a result of YOLOv5's resource dissipation. The monitoring process fails after 24.11, 44.56, and 115.36 hours (on average), when we set the sizes of input images to 160px, 320px, and 640px, respectively, in our system. Our experimental results can be used to plan countermeasures such as software rejuvenation and task offloading.