Dynamic Resource Allocation and Computational Offloading at the Network Edge for Internet of Things Applications

Abstract

In the Internet of Things (IoT) era, mobile devices possess powerful hardware and networking capabilities, however they still fall short when it comes to executing compute-intensive applications. Computation Offloading, i.e., delegating resource consuming tasks to servers located at the Network Edge, contributes towards moving to a Mobile Cloud Computing paradigm, which will potentially assist towards alleviating the computational strain from the mobile devices. The motivation for this thesis is to deal with the inherent challenges of computational offloading, the most important of which is resource allocation under constraints, while guaranteeing the Quality of Service (QoS) and Quality of Experience (QoE) delivered to the users. Throughout this thesis, Control Theory concepts are leveraged as this domain offers a plethora of tools that can be utilised to tackle the emerging challenges. Additionally, concepts from Probability Theory are exploited as well. Specifically, in this work, an effort is made to address the following crucial research challenges in resource allocation and computational offloading: i) modeling the heterogeneous entities of the system (i.e., the infrastructure, users, applications resources and the interactions among them), ii) estimating the workload that will be offloaded during a future time window (by predicting the mobile devices' positions), iii) optimizing resource allocation by dynamically allocating (scaling) the available resources at the network edge while respecting the given constraints, e.g., guaranteed execution of the estimated workload, delivering the expected QoS/QoE and minimizing energy consumption of the edge infrastructure and, finally, iv) optimizing computational offloading strategies. The aim of this thesis is to offer solutions to the above-mentioned challenges, which can be combined into frameworks applicable to real-world edge infrastructures that will allow IoT devices and applications to reach their full potential. To this end, a two-level dynamic resource allocation and admission control mechanism for a cluster of edge servers is developed, to offer an alternative choice to mobile users for executing their tasks. At the lower level, the dynamic behavior of edge servers is modeled by a set of linear systems, and linear controllers are designed to meet the system's constraints and QoS metrics, while at the upper level, an optimizer tackles the problems of load balancing and application placement (bundled in Virtual Machines, VMs) towards maximizing of the number the offloaded requests. The evaluation illustrates the effectiveness of the proposed offloading mechanism regarding the performance indicators (e.g., the application's average response time) and the optimal utilization of the computational resources of the edge servers. Then, the main mechanism of this framework is put to a test; a three-level Cyber-Physical Social System (CPSS) for early fire detection is presented, with an aim to assist public authorities to promptly identify and act on emergency situations. In general, a CPSS tightly integrates computer systems with the physical world and human activities. Specifically, at the bottom level, the system’s architecture involves IoT nodes enabled with sensing and forest monitoring capabilities. Additionally, in this level, the crowd sensing paradigm is exploited to aggregate environmental information collected by end user devices present in the area of interest. Since the IoT nodes suffer from limited computational and energy resources, the resource allocation and admission control mechanism, at the middle level, facilitates the offloaded data processing, regarding possible fire incidents. At the top level, a decision-making service deployed on Cloud nodes, integrates data from various sources, including users’ information on social media, and evaluates the situation's criticality. In this part, the dynamic resource scaling mechanism is designed to address the demanding QoS requirements of this IoT-enabled time and mission critical application. The experimental results indicate that the vertical and horizontal scaling on the Edge Computing layer is beneficial for both the performance and the energy consumption of the IoT nodes. Additionally, a switching computational offloading mechanism for Industry 4.0 applications is discussed. These applications rely on mobile robotic agents that execute many complex tasks that have strict safety and time requirements. Under this setting, the Edge Computing service delivery model allows the robotic agents to offload their computationally intensive tasks to a powerful computing infrastructure in their vicinity. In this part, a novel switching offloading mechanism for such robotic applications is proposed. In particular, opportunistic offloading strategies for the path planning and localization services of mobile robots are designed. The offloading decision is based on the uncertainty of the robot's pose, the resource availability at the Edge of the network and the difficulty of the path planning. The proposed switching offloading framework is implemented and evaluated using a robot in a real Edge Computing testbed, where the trade-off between execution time and the successful completion of the robot trajectory is highlighted. Finally, a Markovian Random Field (MRF)-based computational offloading and resource allocation mechanism is developed. The proposed mechanism leverages switching systems for modeling the computational resources at the network edge and allocating them dynamically, while minimizing energy consumption. This mechanism consists of two repeated stages; during the first, a Markov Chain-based technique is used to predict the mobile users' movements and subsequently to estimate the offloaded workload demand. During the second, a novel MRF-based technique undertakes the balancing of the offloaded tasks to the available computational resources. These tasks cannot be executed locally, i.e., on the user devices, under the given energy constraints and for a specific QoS. The proposed framework manages to improve energy consumption in the edge infrastructure, while taking into consideration the additional network delays induced by the MRF-based load balancing. Moreover, the efficiency of the proposed scheme is evaluated via modeling and simulation and is shown to outperform a well-known task offloading solution. Summarizing, in this thesis, dynamic resource allocation mechanisms for computational offloading are proposed, based on workload prediction, horizontal scaling, vertical scaling and workload balancing. Research is conducted, also, on formulating offloading strategies that work in harmony with these mechanisms, in order to guarantee a level of QoS and QoE, and minimize energy consumption, while the frameworks emerging from combining these techniques, are evaluated in the realistic, demanding environments of Industry 4.0, Natural Disaster Management and Smart Environments, with successful results.