Scream : A New Performance-Portable Global Storm-Resolving Atmosphere Model

The Energy Exascale Earth System Model (E3SM) was born from the U.S. Department of Energy’s desire to leverage its world-leading supercomputing power to answer important questions about climate change. In order to take advantage of the next generation of high-performance computing platforms, the E3SM model must parallelize over millions of computing elements. This can be accomplished by running at very high resolution. The Simple Cloud Resolving E3SM Atmosphere Model (SCREAM) project will propel the E3SM model into a growing class of storm-resolving global climate models, with a target resolution of 3km globally.

SCREAM will represent a nearly complete rewrite of atmospheric physics, providing a unique opportunity to approach the design and evaluation of physics parameterizations focused on both scientific accuracy and the application of software best practices. This effort has resulted in a suite of physics parameterizations that are robustly tested and performance portable.

The atmospheric driver (AD), which is responsible for coupling the processes in the atmosphere model, has been completely redesigned. The new driver infrastructure has an emphasis on simplicity and flexibility. Under the current E3SM infrastructure, atmosphere processes are buried under multiple layers of abstraction. This makes it difficult to add/remove processes or make any adjustments to how they are coupled within the model. The new SCREAM-AD addresses these deficiencies by separating processes from the driver by one interface layer and treating all processes as members of a single atmospheric process class.

From a development perspective, each new process to be included in SCREAM simply needs an interface layer which collects the variables needed by the process, passes them to the parameterization itself, and post-processes the results depending on the coupling mechanism desired. As a result, the adoption of a new process in SCREAM requires just the development of an interface layer. This also serves the inter-operability of parameterizations since they are not directly dependent on any SCREAM infrastructure.

By defining a single atmospheric process class, we are able to choose the set the suite of processes at runtime, which includes the order and the coupling mechanism.

This presentation will introduce the SCREAM design and implementation. There will be discussion of the development and evaluation of the full SCREAM atmospheric physics suite which includes customized versions of the Predicted Particle Properties (P3) microphysics and the Simplified Higher Order Closure (SHOC) macrophysics and turbulence. The coupling of these parameterizations with the rest of SCREAM will be used to illustrate the development of the new SCREAM-AD. Finally, we will demonstrate how robust unit and convergence testing has been used to ensure maximum quality of the SCREAM code base.

This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344 IM Release Number LLNL-ABS-XXXXXX