Most numerical weather prediction models rely on a terrain-following coordinate framework. The computational mesh is thus characterized by inhomogeneities with scales determined by the underlying topography. Such inhomogeneities may affect the truncation error of numerical schemes. In this study, a new class of terrainfollowing coordinate systems for use in atmospheric prediction models is proposed. Unlike conventional systems, the new smooth level vertical (SLEVE) coordinate yields smooth coordinates at mid-and upper levels. The basic concept of the new coordinate is to employ a scale-dependent vertical decay of underlying terrain features. The decay rate is selected such that small-scale topographic variations decay much faster with height than their large-scale counterparts. This generalization implies a nonlocal coordinate transformation. The new coordinate is tested and compared against standard sigma and hybrid coordinate systems using an idealized advection test. It is demonstrated that the presence of coordinate transformations induces substantial truncation errors. These are critical for grid inhomogeneities with wavelengths smaller than approximately eight grid increments, and may overpower the regular-grid truncation error of the underlying finite-difference approximation. These results are confirmed by a theoretical analysis of the truncation error. In addition, the new coordinate is tested in idealized and real-case numerical experiments using a nonhydrostatic model. The simulations using the new coordinate yield a substantial reduction of small-scale noise in dynamical and thermodynamical model fields.
Convection‐resolving models allow to explicitly resolve deep convection at horizontal grid spacings of O(1 km). On current supercomputers, refining the grid spacing to the kilometer scale is computationally still extremely demanding, and therefore, climate simulations at this resolution have so far largely been limited to subcontinental computational domains. However, new supercomputers that mix conventional multicore CPUs and accelerators possess properties beneficial for climate codes. Exploiting these capabilities allows expansion of the size of the computational domains to continental scales. Here we present such a convection‐resolving climate simulation, using a version of the COSMO model, capable of exploiting GPU accelerators. The simulation has a grid spacing of 2.2 km, 1536 × 1536 × 60 grid points, covers the period 1999–2008, and is driven by the ERA‐Interim reanalysis. An assessment of the 10‐year‐long simulation is conducted using a wide range of data sets, including several rain gauge networks, energy balance stations, and a remotely sensed lightning data set. Substantial improvements are found for the 2 km simulation in terms of the diurnal cycles of precipitation. This confirms results found in studies using smaller computational domains. However, the continental‐scale simulations also reveal deficiencies such as substantial performance differences between regions with and without strong orographic forcing. Analysis of the statistical distribution of updrafts and downdrafts shows an increase of the amplitude in seasons with convection and a pronounced asymmetry between updrafts and downdrafts. Furthermore, the analysis of lightning data shows that the convection‐resolving simulation is able to reproduce important features of the annual cycle of deep convection in Europe.
Abstract. The best hope for reducing long-standing global climate model biases is by increasing resolution to the kilometer scale. Here we present results from an ultrahigh-resolution non-hydrostatic climate model for a near-global setup running on the full Piz Daint supercomputer on 4888 GPUs (graphics processing units). The dynamical core of the model has been completely rewritten using a domain-specific language (DSL) for performance portability across different hardware architectures. Physical parameterizations and diagnostics have been ported using compiler directives. To our knowledge this represents the first complete atmospheric model being run entirely on accelerators on this scale. At a grid spacing of 930 m (1.9 km), we achieve a simulation throughput of 0.043 (0.23) simulated years per day and an energy consumption of 596 MWh per simulated year. Furthermore, we propose a new memory usage efficiency (MUE) metric that considers how efficiently the memory bandwidth – the dominant bottleneck of climate codes – is being used.
Currently major efforts are underway toward refining the horizontal resolution (or grid spacing) of climate models to about 1 km, using both global and regional climate models (GCMs and RCMs). Several groups have succeeded in conducting kilometer-scale multiweek GCM simulations and decadelong continental-scale RCM simulations. There is the well-founded hope that this increase in resolution represents a quantum jump in climate modeling, as it enables replacing the parameterization of moist convection by an explicit treatment. It is expected that this will improve the simulation of the water cycle and extreme events and reduce uncertainties in climate change projections. While kilometer-scale resolution is commonly employed in limitedarea numerical weather prediction, enabling it on global scales for extended climate simulations requires a concerted effort. In this paper, we exploit an RCM that runs entirely on graphics processing units (GPUs) and show examples that highlight the prospects of this approach. A particular challenge addressed in this paper relates to the growth in output volumes. It is argued that the data avalanche of high-resolution simulations will make it impractical or impossible to store the data. Rather, repeating the simulation and conducting online analysis will become more efficient. A prototype of this methodology is presented. It makes use of a bit-reproducible model version that ensures reproducible simulations across hardware architectures, in conjunction with a data virtualization layer as a common interface for output analyses. An assessment of the potential of these novel approaches will be provided.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.
hi@scite.ai
10624 S. Eastern Ave., Ste. A-614
Henderson, NV 89052, USA
Copyright © 2024 scite LLC. All rights reserved.
Made with 💙 for researchers
Part of the Research Solutions Family.