Supercomputing

Containerized GEOS: Toward a Portable Climate Model

Image from simulation showing flattened projection of earth's surface with continents in outline and air temperature colored from blue to red
The visualization shows a 15-day Goddard Earth Observing System (GEOS) model simulation of 2-meter air temperature (in Fahrenheit) at 25-kilometer resolution. GEOS ran in a container on 1,035 "Cascade Lake" cores of the NASA Center for Climate Simulation's (NCCS) Discover supercomputer. Colder temperatures are blue, and higher temperatures are red. Matthew Thompson, NASA/Goddard

The NASA Goddard Earth Observing System (GEOS) is an Earth system model used for weather, climate, and other scientific applications. GEOS consists of linked components that can run in various configurations such as atmosphere-only and coupled atmosphere-ocean. Running this model on any new supercomputing system depends on operating systems, compilers, MPI stacks, and libraries being present and correctly configured. To remove that burden from users, this project explores building and running GEOS using Singularity containers—files containing all the needed software dependencies—on both NASA high-end computing systems and commercial cloud computing environments. Ultimately, the goal is for containerized GEOS to make it easier for users outside of NASA to deploy and run the model on any machine.

Quick Facts

Without NCCS resources, this work would not have been possible. The NCCS GitLab Continuous Integration Container building service allowed us to rapidly prototype code and script changes to make containerized GEOS easier. Then, using NCCS Discover, we ran our container on up to 3,870 cores at high model resolution, demonstrating that the concept of containerized GEOS was viable and a path forward to making GEOS a portable model.

Matthew Thompson,
NASA Goddard Space Flight Center
  • The team utilized the NCCS GitLab Continuous Integration Container building service to make the changes needed to build usable Ubuntu Linux and Singularity-based containers for the GEOS model.
  • On the NASA Center for Climate Simulation (NCCS) Discover supercomputer, containerized GEOS performs on par with natively built GEOS over a range of model grid resolutions, but has performance challenges with both MPI and input/output at very high core counts.
  • They also used containers incorporating a portable set of boundary conditions to run GEOS at low resolutions in cloud computing services such as Amazon Web Services (AWS), with the need to install nothing more than Docker container software.
  • Future project directions include debugging the issues with running at very high core counts, mitigating operating system (OS) changes on the supercomputing clusters themselves, testing containers built on other OSs such as SUSE Linux, and simplifying runs of containerized GEOS on cloud services.

Researchers

  • Matthew Thompson, NASA Goddard Space Flight Center
  • Anik Mumssen, NASA Goddard Space Flight Center

More Information