https://www.mdu.se/

Clouds are large computer centers that offer remote access to computing and storage resources, making them popular for business and web applications. They are now being considered for use in safety-critical applications such as factories, but lack sufficient time predictability, which makes it challenging to use them in these time-sensitive applications. To overcome this limitation, an intermediate layer, the fog layer, is introduced to provide computational resources closer to the network edge. However, this new computing paradigm faces its own challenges in resource management, scalability, and reliability due to resource constrained nodes. Lightweight virtualization technologies like containerization can solve the performance-reliability dichotomy in fog computing and provide built-in fault tolerance mechanisms. By studying a robotic use-case, we realized the critical importance of persistent data storage for stateful applications, such as many control applications. However, container-based solutions lack fault-tolerant persistent storage. In this thesis, we identify new challenges associated with leveraging container-based architectures, particularly the importance of persistent storage for stateful applications. We investigate the design possibilities for persistent fault-tolerant storage and propose a solution adapted to container-based fog architectures and tailored for stateful applications. The solution provides scalability, auto recovery, and re-integration after failures at application and node levels. Key elements are a replicated data structure and a storage container, using a consensus protocol for distributed data consistency and fault tolerance in case of node failures. The fault tolerance and consistency of the solution are modeled and verified, and its timing requirements evaluated. We use simulation to evaluate the timing performance of our solution in larger set-ups. The results of our study show that although adding a consistency protocol introduces a timing overhead, the solution still meets timing requirements for the studied use-case even in presence of a set of relevant faults. By leveraging a four-dimensional approach, we also conduct a comparative analysis of our solution with other approaches from various perspectives, indicating that our solution can be applied in a broader context than initially intended.