Computational reproducibility of scientific workflows at extreme scales

Abstract

We propose an approach for improved reproducibility that includes capturing and relating provenance characteristics and performance metrics. We discuss two use cases: scientific reproducibility of results in the Energy Exascale Earth System Model (E3SM—previously ACME) and performance reproducibility in molecular dynamics workflows on HPC platforms. To capture and persist the provenance and performance data of these workflows, we have designed and developed the Chimbuko and ProvEn frameworks. Chimbuko captures provenance and enables detailed single workflow performance analysis. ProvEn is a hybrid, queryable system for storing and analyzing the provenance and performance metrics of multiple runs in workflow performance analysis campaigns. Workflow provenance and performance data output from Chimbuko can be visualized in a dynamic, multilevel visualization providing overview and zoom-in capabilities for areas of interest. Provenance and related performance data ingested into ProvEn is queryable and can be used to reproduce runs. Our provenance-based approach highlights challenges in extracting information and gaps in the information collected. It is agnostic to the type of provenance data it captures so that both the reproducibility of scientific results and that of performance can be explored with our tools.