Search

Cloud-centric systems face challenges in meeting the requirements of real-time Internet of Things (IoT) applications. IoT applications are integrated with Artificial Intelligence to facilitate real-time analytics. An Edge computing paradigm is established that improves the performance of applications in a decentralised approach by moving services closer to the end user. However, resources at the edge are heterogeneous, resource-limited, unreliable, and have intermittent availability. As a result, the application's performance may fluctuate at runtime, affecting the service availability. Therefore, assigned services to the edge resources should adapt to runtime environment to maintain the application's performance. However, it may take time for the services to adapt quickly to the varying runtime conditions. Another problem is that the availability of edge services is impacted by factors such as system overload, network failure, or user mobility, which often occur on edge at runtime. To ensure availability of edge services, the impact of service interruptions caused by failures must be minimised. This thesis aims to develop strategies to cope with the above-mentioned issues for deploying and managing edge services by adapting to runtime conditions and failures. This thesis introduces a framework EDGESER that presents three strategies to improve the performance efficiency of edge services. First, this thesis presents a strategy that models and measures the service delay of available offloading approaches at the edge. Furthermore, the strategy identifies the parameters that impact the service time of different offloading approaches by considering the computation, communication and offline parameters such as the size of the data etc. The second strategy focuses on efficiently adapting to runtime conditions by minimising the downtime incurred while redeploying edge services. The strategy is based on an approach that uses secondary edge-cloud pipelines to process user requests when reassigning edge services. The proposed approach minimises downtime compared to an existing baseline approach, making it suitable for latency-constrained applications. This strategy also considers the trade-off between downtime of edge service and the amount of memory needed for the proposed approach and baseline approach. The third strategy rapidly responds to service failures on edge by redeploying services on a different edge server. The salient feature of the proposed strategy investigates three techniques based on the characteristics of Deep Neural Networks (DNNs) to adapt to service failure. If edge services are recovered and redeployed on edge servers selected solely to mitigate failures, QoS violations, such as service latency violations, may occur. This highlights the significance of optimising service failures and user requirements jointly. The strategy also considers the trade-offs of the techniques, such as accuracy and latency, to choose the best technique based on user-defined objectives (accuracy, latency, and downtime thresholds) when a service failure happens. This enables the edge service providers to balance the trade-off between service quality and service failures.