On learning-based mobile performance guarantee under time-varying resource constraints

Abstract

The development of mobile processors has dramatically improved the performance of applications such as deep learning, AR/VR, and high-performance games on mobile platforms such as smartphones. According to purposes and users, requirements for the end performance of such high-performance mobile applications are increasingly advanced. Thus, it is essential to distribute computing and networking resources to ensure the requirements optimally. In this study, to ensure the performance of high-performance applications such as deep learning and high-definition video analysis on mobile platforms, we present i) optimal mobile computing and networking resource distribution techniques in edge computing. i) To maximize the performance of mobile applications, the maximum allocation of power to mobile processors rather reduces energy efficiency and causes thermal throttling due to overheating, which reduces user experience (QoE). Dynamic Voltage and Frequency Scaling (DVFS) is a technology that dynamically distributes power to each processor to increase the energy efficiency of mobile processors (E.g., CPUs, and GPUs.) and, indirectly, to prevent overheating. However, existing DVFS technology has limitations in operating optimally on mobile platforms. In addition, thermal throttling significantly reduces the performance of networking-based applications such as video streaming by limiting the performance of networking processors. Tackling those challenges, this study presents an adaptive control of processors that considers the operating environment and application of the mobile platform. ii) With the development of networking technology, edge computing can provide opportunities to overcome limitations of on-device computing, such as battery problems and overheating problems. In particular, real-time high-performance video analysis, which is challenging to do on a mobile platform, can be offloaded to an edge server to reduce latency and improve analysis accuracy. This video analysis offloading is operated in the order of data encoding on the mobile platform, network transmission, and analysis on the edge server. Existing offloading techniques used frame-down sampling, filtering, and so on to reduce the size of data to offload, focusing on reducing transmission time. However, the increased bandwidth of wireless networks and the commercialization of high-definition video are accelerating the bottleneck in encoding time on mobile platforms rather than transmission time. Therefore, in real-time video analysis using edge computing, computing resources and wireless networking resources on mobile platforms must be considered simultaneously rather than just reducing networking time to ensure stable performance. This study presents a trade-off newly formed on various wireless network resources and mobile platforms. And then, we present a technique that can adaptively control encoding quality to ensure end-performance by the convergence of mobile computing and networking resources.