B07. Climate Simulations On Summit

Title

A proposal for running long-duration climate simulations on Summit

Authors

Vincent Larson, Gunther Huebler, Brian Griffin, Zhun Guo, John Dennis

Abstract

It would be desirable to take advantage of GPU-based supercomputers, such as Summit, not only to run short-term science simulations, but also to run long-duration climate simulations that include aerosol effects, e.g., a hindcast of 20th-century climate.  However, to parallelize code efficiently on GPUs, the GPU threads must be given a substantial amount of computational work to do.  Furthermore, in order to run simulations of long duration, the throughput must be high and hence the time step must be long.  

To meet both these requirements, we propose the following variant of E3SM.  In order to give the GPU threads lots of work, we propose to use subcolumns (i.e., to call the physics parameterizations many times per grid column).  If the subcolumns call microphysics (and possibly radiation), the GPU threads will be given lots of work to do.  Moreover, subcolumns are highly parallel to each other.  To allow high throughput, we propose to use low resolution (100 km), which enables long time steps.  

This poster outlines our early steps to develop such a variant of E3SM.