Provenance Of Dynamic Adaptations In User-Steered Dataflows

Due to the exploratory nature of scientific experiments, computational scientists need to steer dataflows running on High-Performance Computing (HPC) machines by tuning parameters, modifying input datasets, or adapting dataflow elements at runtime. This happens in several application domains, such as in Oil and Gas where they adjust simulation parameters, or in Machine Learning where they tune models' hyperparameters during the training. This is also known as computational steering or putting the "human-in-the-loop" of HPC simulations. Such adaptations must be tracked and analyzed, especially during long executions. Tracking adaptations with provenance not only improves experiments' reproducibility and reliability, but also helps scientists to understand, online, the consequences of their adaptations. We propose PROVDfA, a specialization of W3C PROV elements to model computational steering. We provide provenance data representation for online adaptations, associating them with the adapted domain dataflow and with execution data, all in the same provenance database. We explore a case study in the Oil and Gas domain to show how PROV-DfA supports scientists in questions like "who, when, and which dataflow elements were adapted and what happened to the dataflow and execution after the adaptation (e.g., how much execution time or processed data was reduced)", in a real scenario.