Reliability-Aware Personalized Deployment of Approximate Computation IoT Applications in Serverless Mobile Edge Computing

Over the past few years, the integration of mobile edge computing and serverless computing, known as serverless mobile edge computing (SMEC), has garnered considerable attention. Despite abundant existing works on SMEC exploration, there remains an unaddressed gap in guaranteeing dependable application outputs due to ignoring the threat of both soft and bit errors on SMEC infrastructures. Furthermore, existing works fall short of accommodating the personalized requirements and approximate computation of Internet-of-things (IoT) applications, thereby resulting in holistic quality-of-service (QoS) degradation of SMEC systems typically provisioned by limited edge resources. In this paper, we investigate the reliability-aware personalized deployment of approximate computation IoT applications for QoS maximization in SMEC environments. To this end, we propose a hybrid methodology composed of offline and online optimization phases. At the offline phase, a decomposition-based function placement method is devised to accomplish function-to-server mapping by integrating convex optimization, cross-entropy method, and incremental control techniques. At the online phase, a lightweight reinforcement learning scheme based on proximal policy optimization (PPO) is developed to handle the inherent dynamicity of IoT applications. We also build a simulation platform upon the real-world base station distribution in Shanghai Telecom and the practical cluster trace in the Alibaba open program. Evaluations demonstrate that our hybrid approach boosts the holistic QoS by 63.9% compared with the state-of-the-art peer algorithms.