[Back]

D. Grygorenko:
"Cost-based Decision Making in Cloud Environments using Bayesian Networks";
Supervisor: I. Brandic, S. Farokhi; Institut für Informationssysteme, Distributed Systems Group, 2014; final examination: 2014-10-07.

Cloud computing providers experienced fast growing trend in recent years by offering highly available and scalable services. However, increasing customer resource demand and competitive market force the providers to enlarge their data centers that leads to huge power consumption or apply more economical resource provision plans that degrades quality of service. Hence, the cloud providers need the solutions to decrease energy costs and improve resource provisioning.
In this thesis, we propose a new approach for a cost-aware cloud power management that enables effective placement of customer´s virtual machines (VMs) in the cloud data centers. This approach consists of two phases. First, we create a model of the cloud infrastructure using Bayesian Networks (BNs). BNs are graphical models that represent variables of interest and probabilistic dependencies among them. They allow to apply knowledge about domain in order to find hidden and causal relationships between different parts of cloud infrastructure. Second, we make decisions about effective VM placement in order to save cloud provider costs. On
this step Multi-criteria decision aid (MCDA) is applied. This technique helps to quantitatively measure the benefit of a certain decision. It implies the usage of utility function that is calculated
based on ranking set of reasonable factors and criteria defined by us.
The related work that has been done in this field of study does not cover all aspects of the problem. In particular, our study focuses more concretely on the geo-distributed data centers experienced
frequent power outages, operating in different time zones and in constantly changing outdoor temperatures.
Additionally, we developed a simulation toolkit and used it to validate our algorithm by conducting a performance evaluation. The results of experiments proved good performance and applicability of the proposed model and its high potential to operate in various real-world scenarios.