Model-Based Evaluation and Optimization of Dependability for Edge Computing Systems

Edge computing moves part of the computing tasks to the edge of the network to improve service capabilities while reducing latency. It has been successfully applied in Internet of Things (IoT) and mobile computing systems. With the increasing popularity of edge computing, the ability of an edge computing system continuously providing services to users without interruptions and failures, which is also known as the dependability, has become an important issue. However, the evaluation and optimization of dependability attributes of an edge computing system still remains an largely unexplored problem. In this paper, we study this issue from a model-based viewpoint. We propose an atomic dependability model of a server and provide quantitative analyses of dependability attributes with Markov chain techniques. In order to facilitate the analyses of multiple attributes in large-scale environments, we adopt a state aggregation method for model simplification, and present its corresponding theoretical proof. Considering the edge-cloud collaboration, we put forward the dependability model of an edge computing system, and provide an evaluation approach using the state aggregation technique. Furthermore, taking task offloading as an example, we formulate the dependability optimization as a continuous-time Markov decision problem (CTMDP), and propose an efficient approach of solving the problem with reinforcement learning. Finally, we use a real-world dataset to conduct simulation experiments, and the experimental results validate the efficacy of our approach.