Dynamic Resource Orchestration and Management of IoT-based Cyber-Physical Systems

Abstract

With the advent of fifth-generation (5G) wireless networks and the Internet of Things (IoT), a crucial part of enabling modern technologies is optimizing the performance of computational offloading and resource allocation mechanisms. The current paradigm in providing additional resources to devices is Multi-Accessed Edge Computing (MEC) which is characterized by the limited availability of resources. Hence, the devices along with the edge servers form a Cyber-Physical System (CPS) which must be carefully designed to fit each application's and scenario's needs. Thus, in this dissertation, we tackle the problem of realizing efficient offloading strategies that benefit the devices and on the other hand, the task scheduling and dynamic resource allocation for optimal utilization of the edge layer's resources, while achieving high-quality of service and experience. In this PhD dissertation, we considered designing holistic frameworks and co-designing these two fundamental aspects as they are tightly coupled. Specifically, an effort is made to address the following crucial research challenges: (i) optimizing the offloading decision taking into consideration device-specific aspects (e.g., signal strength, the position of devices), (ii) optimizing the resource management of resources by jointly designing task scheduling and dynamic resource allocation mechanism that guarantees certain performance criteria while the resources are optimally utilized, (iii) the performance modeling of dynamic workloads and application to assist in deciding the correct number of deployed resources for each application, (iv) the design of practical computation offloading strategies in the concept of approximate computing to alleviate the computational burden from the devices while finding balance between duration and quality of operations and finally (v) the design of closed-loop controllers taking also into account the current state of devices, the communication overhead and the available resource on the edge side.

To this end, we design a holistic architecture to optimize the offloading decision of users, i.e., users offload their request if the communication overhead is low. The offloaded requests may also be rejected by the edge server if the total amount exceeds the current capacity of the deployed resources utilizing performance modeling for the specific application. Then a central entity jointly solves the task scheduling and dynamic resource allocation problem keeping a certain QoS level for the application under a dynamic workload.

Then, in a similar setting, we enhance the task scheduling and dynamic resource allocation mechanism using modeling from queuing theory to dynamically adjust the number of deployed resources for a computing-intensive application. The framework provides stability and performance guarantees for the resource management problem. This is achieved by optimally distributing the incoming tasks and the performance modeling of resources. More specifically, we utilized machine learning algorithms to map theoretically computed values for the processing rate of the deployed resources along with various monitoring KPIs.

Additionally, a novel switching offloading mechanism for robotic applications is proposed. In the context of Industry 4.0, 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. In particular, practical offloading strategies for mobile robot path planning and localization tasks are designed. The offloading decision is based on the uncertainty of the robot’s pose, the resource availability of the edge server, the quality of the network connection, and the difficulty of path planning.

Similarly, Focusing on the stability guarantees and the convergence of the system, we introduce a set-based estimation offloading framework, for the specific case of the navigation of a unicycle robot towards a target position. The robot is subject to modeling and measurement uncertainties, and the estimation set is calculated using overapproximation techniques that alleviate additional computations. A switching set-based control mechanism provides accurate navigation and triggers more precise estimation algorithms when needed. To guarantee the convergence of the system and optimize the utilization of remote resources, a utility-based offloading mechanism is designed, which takes into account both the dynamic network conditions and the available computing resources at the network edge.

Finally, we study the case of real-time edge-assisted inference and batch scheduling. We find out that the quality of the Edge-assisted inference process and the overall latency of the system are competing metrics. So, we formulate a joint optimization problem to maximize the quality of inference while minimizing the overall latency for the GPU-enabled batch processing of inference applications. To deal with the computational complexity of the optimization problem, we first compute the optimal compression policies for the inference tasks to minimize the transmission time. By carefully examining the results of the compression problem, we identify that compressing the tasks in such a way to arrive simultaneously for remote processing significantly increases the performance of batch processing.

Summarizing, in this thesis, we study the problem of dynamic resource orchestration and management of IoT-based Cyber-Physical Systems. The immense advancement of modern applications along with the stringent requirements for ever-expanding resources requires new sophisticated approaches in the journey of providing high QoS and QoE to users. Although many new architectural concepts have arisen in the past few years, there are still many challenges to be addressed to provide seamless operations and execution of the complex algorithms a CPS may have. The two major challenges are the task offloading strategies and the accompanied resource management mechanisms.