Aeras Project Develops Next-Generation Atmosphere Model

The goal of the Aeras LDRD project is to develop a next-generation atmosphere model suitable for a global climate model, with advanced capabilities such as performance portability and embedded uncertainty quantification (UQ). Performance portability will allow us to run our code very efficiently on a diverse set of current and future computer architectures (such as GPUs or threaded CPUs) without a laborious porting process. Our particular approach to embedded UQ takes advantage of running ensembles of simulations to achieve speedups.

Recently, we completed initial development of the 3D hydrostatic equations, which are the target equations for the atmosphere in climate, and a hyperviscosity implementation, which is the preferred method for stabilizing our spectral element discretization. These implementations are currently undergoing rigorous and methodical testing.

The 2D shallow water equations (applied to air, not water) for the global atmosphere were implemented previously and have already demonstrated the project’s next-generation capabilities. We see a roughly 9 times performance increase using threaded CPUs and up to 24 times performance increase with GPUs when the amount of computational work is sufficient. Our approach for computing ensemble samples concurrently can add up to 2 times performance increase when running as few as four samples concurrently. We expect to see similar results for 3D when the hydrostatic equations are ready for performance analysis.

Timings for Aeras finite element assembly of the shallow water equations on serial, threaded (using OpenMP), and GPU (using CUDA) architectures. Note that these three implementations were achieved with a single piece of code enabled by the Kokkos package and programming model. Figure (a) represents total finite element assembly time, while (b) isolates the computational time by removing copy times to and from the device.
Timings for Aeras finite element assembly of the shallow water equations on serial, threaded (using OpenMP), and GPU (using CUDA) architectures. Note that these three implementations were achieved with a single piece of code enabled by the Kokkos package and programming model. Figure (a) represents total finite element assembly time, while (b) isolates the computational time by removing copy times to and from the device.
Speedup of the Aeras shallow water equations using concurrent sampling. Each data point represents the timing for the given concurrent ensemble size run successively to obtain 32 simulations, divided by the time to run 32 independent simulations without concurrent sampling.
Speedup of the Aeras shallow water equations using concurrent sampling. Each data point represents the timing for the given concurrent ensemble size run successively to obtain 32 simulations, divided by the time to run 32 independent simulations without concurrent sampling.
Contact
William Spotz, wfspotz@sandia.gov

April 1, 2016

News story url: https://www.sandia.gov/ccr/news/aeras-project-develops-next-generation-atmosphere-model/