Proactive and autonomic IoT service auto-scaling in constrained edge computing environments

Abstract

The Internet of Things (IoT) has greatly developed with the widespread use of sensors in various areas of modern life, such as healthcare, and construction. The increasing number of connected devices generates massive amounts of data that negatively impact system performance, especially the response time, which is very important for latency-aware applications. Edge computing, which delegates cloud tasks to edge nodes closer to data sources, has been adopted to alleviate this issue. However, IoT devices at the edge are resource-constrained and operate in a highly heterogeneous environment, which is addressed by using containers as a lightweight virtualization technique for deploying microservices. The limitation of resources poses a significant challenge to deploying services at the edge of IoT networks. Effective resource management of edge devices is essential to meet user requirements (e.g., response time) and optimize resource usage, which increases the deployment capacity. Service auto-scaling is an interesting solution that improves resource utilization by dynamically adjusting the number of service instances to match the workload. However, current auto-scalers, such as Kubernetes, mostly rely on threshold-based reactive approaches, which are difficult to configure and less efficient in dealing with complex workloads. These reactive approaches result in wastage due to over-provisioning of resources and performance degradation during resource releases. On the other hand, proactive auto-scaling that anticipates future workload still needs improvement in forecasting accuracy to optimize system performance and effectiveness. The main objective of this work is to propose an approach for deploying IoT services at the edge that is adaptive to performance requirements, resource availability, and workload dynamics. Our approach to service deployment adaptability is based on the MAPE-K (Monitor-Analyze-Plan- Execute over a shared Knowledge) loop. Therefore, each phase of the loop presents a step in the realization of each contribution of our work. Each step must include a literature review, study, and use of existing techniques and their improvement or even proposal of new techniques, implementation, experimentation, and validation. Our first contribution aims at addressing the limitation of threshold-based solutions. Based on our proposed deployment model, encoded in predicate logic, our solution automatically generates rules for different phases, such as analysis and planning. The second contribution proposes a proactive auto-scaling solution. It uses Long short-term memory (LSTM) for workload forecasting thanks to its accuracy and prediction speed. For further accuracy improvement, our approach adds an automatic featurization phase that extracts features from time-series workload data to improve the workload prediction accuracy. It also addresses, in an original way, the oscillation issue caused by the frequently generated scaling actions. The extracted features issued by the featurization phase are used as a grid to mitigate the oscillation issue. The third contribution aims to improve further the accuracy of workload forecasting while considering the resource limitation of IoT devices. We have investigated a variety of known prediction algorithms according to the metrics of accuracy and prediction time. Our proposed Dynamic Ensemble learning approach effectively reduces outliers and maintains high workload forecasting accuracy and auto-scaling performance. In all our contributions, we have validated our approach with several proof-of-concept implementations. We also conducted the necessary experiments by comparing our results with related work. Overall, the validation shows the feasibility and effectiveness of our different contributions. Finally, we highlighted several research perspectives and proposed potential future work within the scope of the presented study.

Date	24 Jul 2023
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Abdelouahed Gherbi (Supervisor)