Scientific workflow execution reproducibility using cloud-aware provenance

Scientific experiments and projects such as CMS and neuGRIDforYou (N4U) are annually producing data of the order of Peta-Bytes. They adopt scientific workflows to analyse this large amount of data in order to extract meaningful information. These workflows are executed over distributed resources, both compute and storage in nature, provided by the Grid and recently by the Cloud. The Cloud is becoming the playing field for scientists as it provides scalability and on-demand resource provisioning. Reproducing a workflow execution to verify results is vital for scientists and have proven to be a challenge. As per a study (Belhajjame et al. 2012) around 80% of workflows cannot be reproduced, and 12% of them are due to the lack of information about the execution environment. The dynamic and on-demand provisioning capability of the Cloud makes this more challenging. To overcome these challenges, this research aims to investigate how to capture the execution provenance of a scientific workflow along with the resources used to execute the workflow in a Cloud infrastructure. This information will then enable a scientist to reproduce workflow-based scientific experiments on the Cloud infrastructure by re-provisioning the similar resources on the Cloud.Provenance has been recognised as information that helps in debugging, verifying and reproducing a scientific workflow execution. Recent adoption of Cloud-based scientific workflows presents an opportunity to investigate the suitability of existing approaches or to propose new approaches to collect provenance information from the Cloud and to utilize it for workflow reproducibility on the Cloud. From literature analysis, it was found that the existing approaches for Grid or Cloud do not provide detailed resource information and also do not present an automatic provenance capturing approach for the Cloud environment. To mitigate the challenges and fulfil the knowledge gap, a provenance based approach, ReCAP, has been proposed in this thesis. In ReCAP, workflow execution reproducibility is achieved by (a) capturing the Cloud-aware provenance (CAP), b) re-provisioning similar resources on the Cloud and re-executing the workflow on them and (c) by comparing the provenance graph structure including the Cloud resource information, and outputs of workflows. ReCAP captures the Cloud resource information and links it with the workflow provenance to generate Cloud-aware provenance. The Cloud-aware provenance consists of configuration parameters relating to hardware and software describing a resource on the Cloud. This information once captured aids in re-provisioning the same execution infrastructure on the Cloud for workflow re-execution. Since resources on the Cloud can be used in static or dynamic (i.e. destroyed when a task is finished) manner, this presents a challenge for the devised provenance capturing approach. In order to deal with these scenarios, different capturing and mapping approaches have been presented in this thesis. These mapping approaches work outside the virtual machine and collect resource information from the Cloud middleware, thus they do not affect job performance. The impact of the collected Cloud resource information on the job as well as on the workflow execution has been evaluated through various experiments in this thesis. In ReCAP, the workflow reproducibility isverified by comparing the provenance graph structure, infrastructure details and the output produced by the workflows. To compare the provenance graphs, the captured provenance information including infrastructure details is translated to a graph model. These graphs of original execution and the reproduced execution are then compared in order to analyse their similarity. In this regard, two comparison approaches have been presented that can produce a qualitative analysis as well as quantitative analysis about the graph structure. The ReCAP framework and its constituent components are evaluated using different scientific workflows such as ReconAll and Montage from the domains of neuroscience (i.e. N4U) and astronomy respectively. The results have shown that ReCAP has been able to capture the Cloud-aware provenance and demonstrate the workflow execution reproducibility by re-provisioning the same resources on the Cloud. The results have also demonstrated that the provenance comparison approaches can determine the similarity between the two given provenance graphs. The results of workflow output comparison have shown that this approach is suitable to compare the outputs of scientific workflows, especially for deterministic workflows