Abstract

With the ever growing popularity of cloud computing, the trend of deploying applications in the cloud is increasing more than ever. Cloud offers computing resources that can be provisioned as required and scaled according to the workload demand. This feature attracts service providers to deploy their applications in the cloud. As users continue to rely more on the services provided by these applications, it is essential to keep the applications running with minimal service outage. Service Availability Forum (SA Forum) has defined a framework called Availability Management Framework (AMF) which can be used to manage service availability. AMF is agnostic to the services provided by the applications. However, it manages the service availability of applications by orchestrating the redundant entities through a configuration called AMF configuration. The design of AMF configurations for a physical cluster based on the functional and non-functional requirements, such as minimum level of service availability, has been proposed in the literature. In these solutions, the number of physical hosts required to deploy an application is given as input and the resource utilization is not taken into consideration. However, for deploying applications in the cloud the number of physical hosts is not fixed and should vary depending on the workload. Therefore, the issue of minimizing the number of physical hosts while meeting the requested level of service availability arises. In particular, the service availability depends not only on the entities involved in providing the service but also on the interferences caused by the collocation of entities. To minimize these interferences, the collocated entities can be grouped into fault isolation units such as VMs. This in turn may increase the number of resources required.
In this thesis, an approach to generate AMF configuration for the cloud is proposed. In this approach, a novel method is used to calculate the number of AMF entities that meets the availability and resource utilization requirements. In addition, a method to estimate service availability is proposed. It aims to predict the availability of service by considering the potential factors that affect availability, including the interferences due to collocation. Furthermore, an approach to deploy AMF applications in the cloud is proposed. As a proof of concept, a prototype that demonstrates the generation and deployment of AMF configurations in an OpenStack cloud has been developed. This prototype includes the existing Monitoring and Elasticity Engine, previously developed in the MAGIC project.