LE3 .A278 2022

2022

Shakshuki, Elhadi

Acadia University

Master of Science

Masters

Computer Science

Computer Science

The cloud computing market is dominated by service-providing companies, including but not limited to Microsoft, AWS, Salesforce, and Google to name a few. These cloud service providers have witnessed tremendous development in their infrastructure to support the performance requests of customers. In order to stand by their service level agreements (SLAs) the computing demands of the customers are met. There’s a pressing need to bring distributed virtual machines to customer devices to cater to the task(s) latency constraints. A task can be classified as high latency, low latency, and ultra-low latency task. High latency tasks are those tasks that require extensive computation time. An example of high latency task is downloading a 4k image on a device that is located far away from the host server. Usually, any task that requires more than 200ms to execute is known as a high latency task [7]. Whereas low and ultra low latency tasks are usually linked with Internet-of-Things (IoT) devices that must converge to completion in a very short duration (a few seconds). An example of such a task includes streaming a live event, as it crucially depends on how low is the end-to-end latency [76]. It is not possible to rely on a cloud data center that is geographically far from the IoT devices for low and ultra-low latency tasks. Unlike the traditional cloud, fog computing and edge computing enable us to bridge the distance between services (virtual machines) and IoT devices, allowing us to handle with tasks that demand faster execution. The emerging fog computing technology is characterized by ultra-low latency tasks that benefit time-sensitive applications and services. Quality of service (QoS) is the description or measurement of the overall performance of a service. Minimizing energy and time while meeting users’ QoS preferences is critical for cloud service providers to maximize profits while maintaining users’ service level agreements (SLA). Allocating user tasks to appropriate virtual machines is an essential topic that contributes in minimizing the time and energy consumed by virtual machines in the cloud. Thus, this thesis proposes clipped double Q-learning, a deep reinforcement learning algorithm by coupling with LSTM for state-action pair forecasting to allocate tasks to their respective virtual machines. Two other algorithms: LSTM-RNN and Deep Q-network base task scheduling, are also proposed in this research to automate the process of allocating tasks to the virtual machines that enhance the performance of the system. The proposed Clipped Deep Double Q-Learning algorithm achieves the desired results by improving the system’s performance by decreasing average service delay, execution and waiting times and several other metrics detailed in this thesis.

The author grants permission to the University Librarian at Acadia University to reproduce, loan or distribute copies of my thesis in microform, paper or electronic formats on a non-profit basis. The author retains the copyright of the thesis.

https://scholar.acadiau.ca/islandora/object/theses:3744