Evaluating Cloud Migration Options for Relational Databases

Migrating the database layer remains a key challenge when moving a software system to a new cloud provider. The database is often very large, poorly documented, and used to store business-critical information. Most cloud providers offer a variety of services for hosting databases and the most suitable choice depends on the database size, workload, performance requirements, cost, and future business plans. Current approaches do not support this decision-making process, leading to errors and inaccurate comparisons between database migration options. The heterogeneity of databases and clouds means organisations often have to develop their own ad-hoc process to compare the suitability of cloud services for their system. This is time consuming, error prone, and costly. This thesis contributes to addressing these issues by introducing a three-phase methodology for evaluating cloud database migration options. The first phase defines the planning activities, such as, considering downtime tolerance, existing infrastructure, and information sources. The second phase is a novel method for modelling the structure and the workload of the database being migrated. This addresses database heterogeneity by using a multi-dialect SQL grammar and annotated text-to-model transformations. The final phase consumes the models from the second and uses discrete-event simulation to predict migration cost, data transfer duration, and cloud running costs. This involved the extension of the existing CloudSim framework to simulate the data transfer to a new cloud database. An extensive evaluation was performed to assess the effectiveness of each phase of the methodology and of the tools developed to automate their main steps. The modelling phase was applied to 15 real-world systems, and compared to the leading approach there was a substantial improvement in: performance, model completeness, extensibility, and SQL support. The complete methodology was applied to four migrations of two real-world systems. The results from this showed that the methodology provided significantly improved accuracy over existing approaches.