CLOUD RESOURCE MANAGEMENT USING A HIERARCHICAL DECENTRALIZED FRAMEWORK

Abstract

Cloud providers (CPs) build and operate large scale data centres that contain numerous computing resources that are typically virtualized and require a level of orchestration of the shared resources. With an increased demand for cloud computing resources at a lower cost by end-users, CPs need to increase the efficiency of their infrastructure usage. To achieve this, CPs aim to increase resource utilisation and lower operational costs, typically the energy to administer, run and cool computing resources. A promising approach to increase the efficiency of infrastructure usage is to adapt the assignment of resources to workloads. This can be used, for example, to apply a policy that conserves energy by combining workloads and enabling CPs customers to meet their performance objectives. The mapping of workloads to data centre resources can be viewed as being carried out by two abstract components, Management Algorithm (MA) and Management Framework (MF). The MA is responsible for deciding how workloads are assigned to infrastructure resources. At the same time, the MF enables the MA to execute by providing standard functionality, such as the scope of the in- infrastructure being managed and aggregation of metrics that will allow the MA to make decisions. Several architectural solutions have been presented for MFs. However, these tend to be centralized and may suffer in their ability to run the MA at scale and support data centres with thousands of physical nodes. Decentralized approaches solve the scalability problem but have a limited view of resources across the data centre, which reduces the opportunity to remap resources across a larger scope of the infrastructure. Several techniques are used for MAs, with heuristics being a common choice. However, heuristics’ performance depends on multiple factors, including the statistical patterns of workload demands, and if the underlying scenario changes, heuristics may start to perform poorly. This thesis is grounded on the hypothesis that solving the scalability challenge in mapping workloads to resources starts by addressing scalability in the MF. We propose a novel scalable hybrid MF and demonstrate this to improve the ability to meet performance objectives and provide a global view of the infrastructure through empirical evaluation. To address the challenge with heuristic MAs, we propose a reinforcement learning-based MA that can learn a policy to dynamically balance achieving Service Level Agreements, achieve high CPU utilization, and remove the need to use defined CPU utilization thresholds. We combine the proposed MF with the proposed MA and demonstrate this outperforms heuristic approaches in reducing service level agreement violations and provides high CPU utilisation through empirical evaluation.

Date of Award	31 Dec 2022
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Norman Paton (Supervisor) & Rizos Sakellariou (Supervisor)