Doctor's Theses (authored and supervised):
"Predictive Approaches for Resource Provisioning in Distributed Systems";
Supervisor, Reviewer: S. Schulte, V. Cardellini, S. Tai;
Institute of Information Systems Engineering, Distributed Systems Group,
oral examination: 2020-09-03.
Modern distributed systems, such as cloud computing infrastructures or data stream processing engines, perform resource provisioning tasks such as resource allocation, task scheduling, or scaling. This decision-making substantially influences the systems´ performance, and therefore, the manner of reaching these decisions is crucial to the systems´ operation with regard to cost efficiency, performance, reliability, and adherence to service level agreements. Currently, many approaches to resource provisioning in distributed systems are reactive, i.e., they measure the systems´ state, analyze it, and perform necessary actions. The main downside of reactive approaches is that effectively, such systems perform resource provisioning based only on past observations. In a highly dynamic environment with rapidly changing demands for computational resources, this can lead to delayed reactions, which increase cost, degrade performance, and reduce reliability. This thesis proposes the use of predictive technologies for performing resource provisioning tasks in modern distributed systems. As a foundation, methods stemming from research in the field of machine learning are used to improve target metrics like system performance oroperational cost. In contrast to traditional, reactive approaches, the proposed methodology of predictive decision-making is able to perform operational tasks ahead of time, such as scaling out in advance for a predicted increase of demand. We show how to use predictive methods in various domains of distributed systems, namely cloud computing, business process management systems, data stream processing, and blockchains. We propose approaches to solving challenges in designing predictive methods, such as metric prediction, failure prediction, or data filtering and estimation. We evaluate the impact of the proposed methods on the system using various quantitative methods, including testbed evaluation and simulation, as well as formal and qualitative analysis. Our results show that employing predictive approaches in these domains of distributed systems significantly improves performance attributes such as response time or adherence to service level agreements.
Predictive Systems / Machine Learning / Cloud Computing / Business Processes / Resource Provisioning / Data Stream Processing / Blockchain
Created from the Publication Database of the Vienna University of Technology.