Automatic Generation of Distributed Runtime Infrastructure for Internet of Things

Ph. D. ThesisThe Internet of Things (IoT) represents a network of connected devices that are able to cooperate and interact with each other in order to reach a particular goal. To attain this, the devices are equipped with identifying, sensing, networking and processing capabilities. Cloud computing, on the other hand, is the delivering of on-demand computing services – from applications, to storage, to processing power – typically over the internet. Clouds bring a number of advantages to distributed computing because of highly available pool of virtualized computing resource. Due to the large number of connected devices, real-world IoT use cases may generate overwhelmingly large amounts of data. This prompts the use of cloud resources for processing, storage and analysis of the data. Therefore, a typical IoT system comprises of a front-end (devices that collect and transmit data), and back-end – typically distributed Data Stream Management Systems (DSMSs) deployed on the cloud infrastructure, for data processing and analysis. Increasingly, new IoT devices are being manufactured to provide limited execution environment on top of their data sensing and transmitting capabilities. This consequently demands a change in the way data is being processed in a typical IoT-cloud setup. The traditional, centralised cloud-based data processing model – where IoT devices are used only for data collection – does not provide an efficient utilisation of all available resources. In addition, the fundamental requirements of real-time data processing such as short response time may not always be met. This prompts a new processing model which is based on decentralising the data processing tasks. The new decentralised architectural pattern allows some parts of data streaming computation to be executed directly on edge devices – closer to where the data is collected. Extending the processing capabilities to the IoT devices increases the robustness of applications as well as reduces the communication overhead between different components of an IoT system. However, this new pattern poses new challenges in the development, deployment and management of IoT applications. Firstly, there exists a large resource gap between the two parts of a typical IoT system (i.e. clouds and IoT devices); hence, prompting a new approach for IoT applications deployment and management. Secondly, the new decentralised approach necessitates the deployment of DSMS on distributed clusters of heterogeneous nodes resulting in unpredictable runtime performance and complex fault characteristics. Lastly, the environment where DSMSs are deployed is very dynamic due to user or device mobility, workload variation, and resource availability. In this thesis we present solutions to address the aforementioned challenges. We investigate how a high-level description of a data streaming computation can be used to automatically generate a distributed runtime infrastructure for Internet of Things. Subsequently, we develop a deployment and management system capable of distributing different operators of a data streaming computation onto different IoT gateway devices and cloud infrastructure. To address the other challenges, we propose a non-intrusive approach for performance evaluation of DSMSs and present a protocol and a set of algorithms for dynamic migration of stateful data stream operators. To improve our migration approach, we provide an optimisation technique which provides minimal application downtime and improves the accuracy of a data stream computation