HAMAM: HAM and ART Acceleration for Many‐Core Architectures

PI: Dominik Brunner (Empa) 

Co-PIs: Ulrike Lohmann, Jennifer Buchmüller, Xavier Lapillonne, Stephan Henne, Sylvaine Ferrachat, Christina Schnadt Poberaj

July 1, 2021 – June 30, 2023

Project Summary

Human activities have altered the chemical composition of the atmosphere in a major way, with wideranging implications for our environment. Increases in the concentrations of long‐lived greenhouse gases are driving global warming, emissions of chlorofluorocarbons and other synthetic gases have threatened the stratospheric ozone layer, air pollutants like ozone and particulate matter are often reaching levels harmful to human health and ecosystems, and changes in aerosols concentrations are affecting climate by altering the radiative balance of the Earth and the formation and properties of clouds.  

Atmospheric chemistry and transport models (ACTMs) have played a key role in advancing our understanding of these complex processes, and have provided critical information guiding policy decisions. They are important extensions of weather and climate models enabling new applications of high societal relevance including air quality forecasts, predictions of the dispersion of hazardous pollution plumes, or projections of climate change accounting for the effects of aerosol‐cloud interactions.  

Our capability to simulate the past and future evolution of weather, climate and atmospheric composition has seen enormous progress over the past decades. A key driver of this evolution, in addition to advances in our understanding of the underlying processes, has been the rapid progress in high‐performance computing (HPC). The finer resolution of today's models allows us to more explicitly represent key processes such as convection that had to be parameterized previously. The higher computing power also allows us to integrate more processes and describe them in greater detail.  

Running ACTMs is typically an order of magnitude more expensive computationally compared to weather and climate models, because a much larger number of tracers have to be transported and many additional processes including complex chemical reactions and aerosols have to be considered. The high computational cost has been a major limitation in the past, restricting the resolution of regional air quality models, for example, to several kilometers, which is far from resolving the true spatial variability in air pollutant concentrations. It is expected that with the further increase in HPC capacity, the resolution of these models will continue to be refined, and new applications will be made possible with better guidance for environmental impact assessment and climate change adaptation measures.  

Because the evolution of microprocessor technology is reaching fundamental limits, emerging HPC architectures increasingly involve the use of heterogeneous many‐core architectures consisting of both CPUs and accelerators, e.g. Graphical Processing Units (GPUs). Because of their high computing performance and energy efficiency, GPUs are currently one of the most promising architectures. Indeed, the current fastest systems in the world (https://www.top500.org/) are based on such an architecture, as is Piz Daint, the largest system in Switzerland.   

Fully exploiting the computational capacity of these systems requires a paradigm shift in the way atmospheric models are designed. The weather and climate modeling community has taken on this challenge and has started adapting existing models and developing new tools and libraries supporting this transition. A pioneering example is the regional numerical weather prediction and climate model COSMO, which has been fully ported to GPUs and is now operationally used by MeteoSwiss. Similar developments are under way for other model systems in Europe, coordinated under the EU projects ESCAPE (Energyefficient SCalable Algorithms for weather and climate Prediction at Exascale) and ESCAPE‐2.  

Similar coordinated efforts in the chemistry‐climate and air quality modeling communities are lacking, which bears the risk that the development of ACTMs will lose connection from their parent models. Adapting ACTMs to emerging architectures requires a major investment, because a large variety of processes with a correspondingly large variety of computational motifs and numerical methods has to be ported.   In this project we aim to start or extend the porting to GPU architectures of two state‐of‐the‐art aerosol and chemistry modules, which have been recently integrated into the Icosahedral Non‐hydrostatic (ICON) atmospheric general circulation model. The modules are (ICON‐)ART and (ICON‐)HAM, which have been extensively used in the past for studying aerosol–climate interactions, atmospheric chemistry processes, and air pollution in various model configurations from the global to the regional scale. The ICON model (in specific configurations) has recently been adapted to GPUs in the PASC project ENIAC (Enabling the ICON model on heterogeneous architectures) building on the tools developed for COSMO. MeteoSwiss is currently porting all additional components required for numerical weather prediction (NWP), namely the so‐called NWP‐Physics and data assimilation as part of the IMPACT project, financed by Deutscher Wetterdienst (DWD), and is planning to run ICON on GPU for operational production in the 2023 time horizon. The adaptation of ICON to future exascale computing will be continued and taken to a next level in the 6‐year project EXCLAIM at ETHZ, but no adaptation of ART or HAM is foreseen in that project. Porting of the HAM module to GPUs is currently addressed in a 1‐year extension of the ENIAC project.  

This activity will be extended in this project by developing more advanced porting strategies.