Résumé
Computational Science and Engineering (CSE) workflows are large-scale, requireHigh Performance Computing (HPC) execution, and have the exploratory nature ofscience.During the long run, which often lasts for hours or days, users need to steerthe workflow by dynamically analyzing it and adapting it to improve the qualityof results or to reduce the execution time. However, to steer the workflow, userstypically perform several interactions (called user steering actions), which need tobe tracked. Otherwise, users find it harder to understand how and what needs to besteered, they can steer in a misleading way, it can be difficult to explain the resultsthat were consequences of their actions, and it can be impossible to reproduce theresults. This thesis addresses this problem by proposing an approach that defines thefundamental concepts for user steering action; introduces the notion of provenanceof steering actions; and contemplates a W3C PROV-compliant data diagram tomodel steering action data with provenance. Also, the approach presents systemdesign principles to enable the management of steering action data by capturing,explicitly relating the actions to the rest of the workflow data, and storing thesedata efficiently. Two instances of this approach were designed and built: one isa lightweight tool to be plugged into parallel scripts and the other is to be usedwithin a Parallel Workflow Management System, which are the two typical ways toconduct CSE experiments in HPC. Using real use cases in the Oil and Gas industry,the experiments show that the proposed approach enables users to understand howtheir actions directly affect the workflow results at runtime and that the systemdesign principles were essential to add negligible overhead to the HPC workflows.