Parallel I/O Performance Tuning for the OpenGeoSys Multi-Physics Solver
Date: Monday, May 22, 2023, 03:00 PM - 05:00 PM
Room: Foyer D-G - 2nd Floor
Type: Research Poster
Description: OpenGeoSys (OGS) is a scientific open source project for the development of numerical methods for the simulation of thermo-hydro-mechanical-chemical (THMC) processes in porous and fractured media. Simulation examples are: groundwater flow, solute transport, heat transport, multiphase flow, and reactive transport of pollutants. The software has been successfully applied in the fields of regional, contaminant and coastal hydrology, fundamental and geothermal energy systems, geotechnical engineering, energy storage, CO2 sequestration/storage and nuclear waste management and disposal. OGS is designed to store results recurrently after a given number of simulation steps. Previous investigations have shown that file I/O is a bottleneck in this process. The problem gets worse when scaling the number of CPUs. To solve the bottleneck, a new distributed HDF5-based approach was implemented, which should, if possible, generate only one result file. During the implementation it became apparent that a number of factors have to be taken into account in order to make the output performant. The work shown here reports on the accompanying I/O analysis with the tool Score-P, names important factors and discusses their impact with numbers and timeline charts. Overall, the OGS I/O performance was improved by a factor of 7.
Links: Official link from ISC 2023