Automatic deployment and reproducibility of workflow on the Cloud using container virtualization

Abstract

Cloud computing is a service-oriented approach to distributed computing that has many attractive features, including on-demand access to large compute resources. One type of cloud applications are scientific work ows, which are playing an increasingly important role in building applications from heterogeneous components. Work ows are increasingly used in science as a means to capture, share, and publish computational analysis. Clouds can offer a number of benefits to work ow systems, including the dynamic provisioning of the resources needed for computation and storage, which has the potential to dramatically increase the ability to quickly extract new results from the huge amounts of data now being collected. However, there are increasing number of Cloud computing platforms, each with different functionality and interfaces. It therefore becomes increasingly challenging to de ne work ows in a portable way so that they can be run reliably on different clouds. As a consequence, work ow developers face the problem of deciding which Cloud to select and - more importantly for the long-term - how to avoid vendor lock-in. A further issue that has arisen with work ows is that it is common for them to stop being executable a relatively short time after they were created. This can be due to the external resources required to execute a work ow - such as data and services - becoming unavailable. It can also be caused by changes in the execution environment on which the work ow depends, such as changes to a library causing an error when a work ow service is executed. This "work ow decay" issue is recognised as an impediment to the reuse of work ows and the reproducibility of their results. It is becoming a major problem, as the reproducibility of science is increasingly dependent on the reproducibility of scientific work ows. In this thesis we presented new solutions to address these challenges. We propose a new approach to work ow modelling that offers a portable and re-usable description of the work ow using the TOSCA specification language. Our approach addresses portability by allowing work ow components to be systematically specifed and automatically - v - deployed on a range of clouds, or in local computing environments, using container virtualisation techniques. To address the issues of reproducibility and work ow decay, our modelling and deployment approach has also been integrated with source control and container management techniques to create a new framework that e ciently supports dynamic work ow deployment, (re-)execution and reproducibility. To improve deployment performance, we extend the framework with number of new optimisation techniques, and evaluate their effect on a range of real and synthetic work ows.