Capturing provenance as a diagnostic tool for workflow performance evaluation and optimization

Abstract

In extreme-scale computing environments such as the DOE Leadership Computing Facilities scientific workflows are routinely used to coordinate software processes for the execution of complex, computational applications that perform in-silico experiments. Monitoring the performance of workflows without also simultaneously tracking provenance is not sufficient to understand variations between runs, configurations, versions of a code, and between changes in an implemented stack, and systems, i.e. the variability of performance metrics data in their historical context. We take a provenance-based approach and demonstrate that provenance is useful as a tool for evaluating and optimizing workflow performance in extreme-scale HPC environments. We present Chimbuko, a framework for the analysis and visualization of the provenance of performance. Chimbuko implements a method for the evaluation of workflow performance from multiple components that enables the exploration of performance metrics data at scale.