Date of Degree
Computer and Systems Architecture
Edge computing, resource allocation, energy saving, multi-view 3D reconstruction, DNN partitioning, long-term optimization
With current and future mobile applications (e.g., healthcare, connected vehicles, and smart grids) becoming increasingly compute-intensive for many mission-critical use cases, the energy and computing capacities of embedded mobile devices are proving to be insufficient to handle all in-device computation. To address the energy and computing shortages of mobile devices, mobile edge computing (MEC) has emerged as a major distributed computing paradigm. Compared to traditional cloud-based computing, MEC integrates network control, distributed computing, and storage to customizable, fast, reliable, and secure edge services that are closer to the user and data sites. However, the diversity of applications and a variety of user specified requirements (viz., latency, scalability, availability, and reliability) add additional complications to the system and application optimization problems in terms of resource management. In this thesis dissertation, we aim to develop customized and intelligent placement and provisioning strategies that are needed to handle edge resource management problems for different challenging use cases: i) Firstly, we propose an energy-efficient framework to address the resource allocation problem of generic compute-intensive applications, such as Directed Acyclic Graph (DAG) based applications. We design partial task offloading and server selection strategies with the purpose of minimizing the transmission cost. Our experiment and simulation results indicate that partial task offloading provides considerable energy savings, especially for resource-constrained edge systems. ii) Secondly, to address the dynamism edge environments, we propose solutions that integrate Dynamic Spectrum Access (DSA) and Cooperative Spectrum Sensing (CSS) with fine-grained task offloading schemes. Similarly, we show the high efficiency of the proposed strategy in capturing dynamic channel states and enforcing intelligent channel sensing and task offloading decisions. iii) Finally, application-specific long-term optimization frameworks are proposed for two representative applications: a) multi-view 3D reconstruction and b) Deep Neural Network (DNN) inference. Here, in order to eliminate redundant and unnecessary reconstruction processing, we introduce key-frame and resolution selection incorporated with task assignment, quality prediction, and pipeline parallelization. The proposed framework is able to provide a flexible balance between reconstruction time and quality satisfaction. As for DNN inference, a joint resource allocation and DNN partitioning framework is proposed. The outcomes of this research seek to benefit the future distributed computing, smart applications, and data-intensive science communities to build effective, efficient, and robust MEC environments.
Zhang, Xiaojie, "Resource Management in Mobile Edge Computing for Compute-intensive Application" (2023). CUNY Academic Works.