Last updated: 4 Feb 2019
Project duration: September 2015 – August 2017 (an extension requested until August 2018)
FTEs (plan/used) 1.1 / 0.695 in COSMO year 2015-2016
[1.6]+0.5 / 0.87 in COSMO year 2016-2017
1.6 / 0.45 in COSMO year 2017-2018
Total FTEs used: 2.015
The goal of this project is an as objective as possible comparison of the dynamical cores of COSMO and ICON mainly due to several tests (tasks 1-3 and 5). It is not intended to carry out substantial model developments if obstacles should be recognized, but to collect experiences and to document them.
In this sense this priority project can serve as a preliminary study before a planned larger Priority Project about comparisons of the complete two models (i.e. including parameterisations and data assimilation).
A large part of this project plan is based on the ‘decision tree’ (Feb. 2010) by M. Baldauf and O. Fuhrer for the former PP CDC. It provides a series of criteria, which have to be fulfilled at first to consider the ICON dynamical core as well suited for the requirements of the COSMO members for their applications.
During the next years, the COSMO consortium as a whole and every single member will face the task to switch at a certain time from the current COSMO model to another modeling framework, the ICON model. The development goals of ICON have primarily been determined by the global scale, whereas the regional modeling aspect came into play only later on. Therefore, it is not a priori clear that the superior properties of ICON on the global scale also hold on the regional scale. It is therefore mandatory to extensively test ICON for the COSMO purposes. It is advisable to first concentrate on the dynamical core and its behaviour in comparison to those of the COSMO model. If there are no obstacles seen, then one can proceed to compare the full fledged model systems; this will be done in an independent priority project.
he core actions are the evaluation of the ICON dynamical core and the comparison with those of COSMO with the help of idealized and semi-realistic test simulations. In the meteorological dynamical core community a certain set of idealized test cases to assess dynamical cores slowly evolves. These will be exercised in task 1.
With semi-realistic tests the overall performance of the dynamical cores with respect to run time and the possible occurrence of artificial signals (e,g, unrealistic gradients, extrema, ...) can be assessed (task 2). Partly technical aspects as scalability and usability for special purposes (climate, environmental applications) will be covered by tasks 3 and 5.
It is necessary that the project members have a good working knowledge with ICON before they start contributing to the project. As one possibility, the ICON training planned at DWD for 12 to 15 October 2015 should be used for this purpose.
to be filled
Task 1 Good performance on a standard set of idealized test cases
These tests check fundamental properties of the simulation of the Euler-equations. The chosen tests should be meteorologically relevant e.g. they check the non-hydrostatic flow regime, the accuracy of the terrain-following coordinate system, well-balancing properties, …and they are applicable to a limited area model.
A test set which seems to be comprehensive and is considered as a certain standard in the atmospheric modelling community consists in:
1. Advection test with nonlinear dynamics (Schär et al. (2002)) NN
2. Atmosphere at rest (Zängl et. al (2004) MetZ) Barbu/ Dumitrache
3. Cold bubble (Straka et al. (1993)) (unstationary density flow) Barbu/Dumitrache
4. Mountain flow tests (stationary, orographic flows) Baldauf
4.1 Schaer et al. (2002), section 5b
4.2 Bonaventura (2000) JCP
4.3 3D-case (dry) Schmidli (?)
5. Linear Gravity waves (Baldauf, Brdar (2013)) Baldauf
6. Warm bubble (Robert (1993), Giraldo (2008)) NN
7. Moist, warm bubble: Weisman, Klemp (1982) MWR NN
8. Advection tests for tracer schemes (solid body rotation, LeVeque test, …) Will (without FTE)
Some of these tests have been performed already with ICON, at least on the sphere. There will be some preparatory work to implement these tests on a plane and also to generate a grid and appropriate boundary conditions to perform 2D simulations.
All of these tests are already available in COSMO and have been carried out in the former CDC priority project. Therefore, the comparison between the dynamical cores should not take much time on the COSMO side (one may even refer to the CDC final report).
Requirements:
Satisfying simulation of all defined tests and no drawbacks compared to COSMO is required. Comparison with published results (for some of these tests even exact analytic solutions are available). Some limits can be deduced, too, e.g. maximum achievable steepness in mountain flow tests. Some test setups (special background profiles, initial distributions) must be implemented first. Those tests using a 2D slice (x-z-) setup need an appropriate grid with periodic boundary conditions in one direction; this must be prepared before.
Deliverables:
A report (preferably a COSMO newsletter article) about the outcome of these tests. Some of the test results have already been published elsewhere; in this case a reference is sufficient. The implementation of the limited area test setups is available to the community.
Ressources:
DWD 0.2 FTE (Baldauf, DWD), 0.3 FTE (Barbu, NMA), 0.2 FTE (NN)
Task 2. Ability to handle real-/semi-idealised cases reasonably well
Of course, a dynamical core alone is not able to produce reasonable forecasts. But it should be able to run real-case simulations (i.e. real orography and more or less real inflow conditions) almost without any physical parameterizations stably at least for a limited time and without showing any serious drawbacks. The only exception might be the need of a turbulence parameterization to stabilize the model against unphysical shear instabilities. Since the turbulence schemes in COSMO and ICON are practically the same, this should not too much influence the dynamical core assessment.
(Side remark: real case simulations with full-fledged parameterisations will be performed in another, larger PP as mentioned above).
The regions and domain sizes should be comparable with currently operationally used setups. Therefore resolutions in the range 7 km …1 km … 0.5 km should be covered.
For these tests an improvement of the user-friendlyness of the ‘ICON regional mode’ would be helpful, nevertheless the currently available ICON version can be used with some drawbacks concerning convenience.
To generate comparable testing conditions, any orography filtering should be similar for both models. Possibly, such a filter function must be implemented before.
Requirements:
Tests with a few proposed test cases: a few selected areas (perhaps only one, e.g. Alps should be included), resolutions, inflow conditions (e.g. a winterly strong wind inflow). No serious drawbacks recognizable in the quality of the simulated results. In particular, a comparison of the computation time due to the different time integration schemes in COSMO and ICON must be done. Apart from the inspection of the meteorological fields, some typical diagnostics should be done, too, e.g. look at kinetic energy spectra.
If serious problems with simulations should occur (e.g. instabilities), the ICON (or COSMO) developers should be informed immediately.
Planned regions:
Alps 0.5, 1.1, 2.2, 7 km resolution (deMorsier)
Sochi-region: 0.5, 1.1, 2.2, 7 km resolution (strong inflow winter case)
Especially one should look at pressure difference field.
It must be clarified if a turbulence scheme + transfer scheme is needed for these semi-realistic test runs (additionally ‘common physics’ needed?) or if a free slip condition is sufficient
Deliverables:
a report (preferably a COSMO newsletter article) about the outcome of these tests.
Ressources:
0.1 FTE (Baldauf, DWD), 0.2 FTE (deMorsier, MeteoCH) possibly 0.5 FTE (Dumitrache, Barbu, MeteoRomania) 0.2 FTE (NN, NN, …) preferably people from several centers should be involved
?? FTE Roshydromet?
Task 3 Scalability/Performance suitable for operations as well as for future supercomputing platforms
COSMO currently uses MPI parallelization. ICON is able to use both (internode) MPI parallelisation and intranode OpenMP parallelisation.
Both the currently used supercomputing platforms and also near future architectures should be supported. For the time being this means that it should run efficiently on massively parallel machines (IBM, Cray, with several 1000, 10000 … processors).
Side remark 1: vector machines (e.g. NEC) are currently not used by COSMO partners and need not be considered.
Side remark 2 about GPU based machines: COSMO is now ready for such platforms by a combination of openACC directives and the use of the stencil library STELLA for the dynamical core. An openACC-based implementation of ICON is ongoing, however, it is not clear if a comparison on GPU’s is possible during the lifetime of this project.
Requirements:
approximately linear scalability (strong scaling) for at least 2 different machines (IBM/Cray/…) for one realistic model setup (real case number of grid points horizontally and vertically, real case grid spacings, realistic grid stretching, realistic flow situation, …) i.e. take one of the setups used in task 2.
Deliverables:
a report at the COSMO GM is sufficient.
Ressources:
0.1 FTE (Prill, DWD), 0.1 FTE (deMorsier, MeteoCH, for GPU version)
Task 4 Identification of differences in dynamical core formulations and their assessment
It is well known that the currently used dynamical core of COSMO (‘Runge-Kutta’) is based on the compressible, non-hydrostatic, shallow atmosphere equations like the ICON dynamical core. If the extension of the PP CELO towards a compressible EULAG dynamical core will be carried through, then the same holds for the EULAG version of COSMO, too.
However, there are of course differences in the time integration scheme (HE-VI/time-split versus HE-VI/non-split versus semi-implicit) which results in different computing time and different needs of damping mechanisms. An example is the divergence damping which is mandatory for the RK-time-split scheme but not necessarily for the ICON core.
Moreover, it is not clear how far differences in the grid structure (quadrilateral versus triangular elements) could have a detrimental impact compared with COSMO. Here, differences between the advection 5th order versus advection of mostly 3rd order could be visible. Similar differences could occur for the tracer advection schemes (Bott or Semi-Lagrange or MPDATA versus Miura)
It should also be assessed if ICON is able to overcome current shortcomings of the COSMO model. These known shortcomings are (for ‘Runge-Kutta’):
- steep terrain: both the dynamical core becomes unstable for steep terrain and also some tracer advection schemes have problems with it,
- Conservation: no conservation of the dynamical variables mass, momentum, energy or potential vorticity
- Boundary conditions (less dependence from driving model at outflow boundary is desired)
Requirements:
literature study (model documentations); code inspection, consideration of the results of tasks 1-3. The differences should be summarized and assessed. An extra assessment about the benefits/obstacles of ICON dynamical core compared to COSMO (RK and EULAG) for future model requirements (e.g. for scales 1 km to 0.1 km) should be given.
Deliverables:
a report at the COSMO GM (or a COSMO newsletter article)
Ressources:
DWD 0.2 FTE (Baldauf, DWD), 0.1 FTE (Zängl, DWD)
Task 5 Suitability of ICON dynamical core for other applications than NWP (climate, chemistry, ...) compared to the COSMO model
Due to close collaborations of the NWP centres with other communities (climate, aerosol, chemistry groups) the new dynamical core should work in those applications too. It should be efficient for longer time runs (climate) and also well balanced for coarse resolutions. There should not occur any drifts in the model; but this is probably hard to analyse because such drifts often are induced by the physical parameterizations. A restart of runs should be possible.
Requirements:
literature study. Estimations about run time needs in a climate mode. A test in a climate mode to recognize possible drifts.
Climate runs should be reproducible on different platforms in a statistical sense.
Deliverables:
a report at the COSMO GM
Ressources:
0.5 FTE (NN) from CLM community?
Since this is merely an evaluation (or assessment) project and not a development project, there are only minor risks for carrying out the project itself (enough available FTEs assumed). Task 1 bases on the possibility to perform 2D tests, i.e. with periodic boundary conditions in one direction. Although surely only a technical problem, it could take more time than planned due to the completely different way to construct grids for ICON.
| FTEs 2016 | FTEs 2017 | |
|
M. Baldauf (DWD) |
0.1 |
0.1 |
|
M. Baldauf (DWD) C. Barbu (NMA) |
0.2 |
|
|
N.N., M. Baldauf (DWD) |
|
0.2 |
|
M. Baldauf (DWD) G. deMorsier (MeteoCH) R. Dumitrache, C. Barbu (MeteoRom) N.N. |
|
0.1 0.2 0.5(?)
0.2 |
|
F. Prill (DWD) |
0.1 |
|
|
G. deMorsier (MeteoCH) |
|
0.1 |
|
G. Zängl, M. Baldauf (DWD) |
0.3 |
|
|
N.N. (CLM-community?) |
|
0.5 |
Baldauf M. and S. Brdar (2013): An analytic solution for linear gravity waves in a channel as a test for numerical models using the non-hydrostatic, compressible Euler equations, Quart. J. Royal Met. Soc., 139, 1977-1989.
Giraldo, F. X., Restelli, M. (2008): A study of spectral element and discontinuous Galerkin methods for the Navier-Stokes equations in nonhydrostatic mesoscale atmospheric modeling: equation sets and test cases, J. Comput. Phys., 227/8, 3849—3877.
Robert, A. (1993): Bubble convection experiments with a semi-implicit formulation of the Euler equations, J. Atmos. Sci., 50, 1865–1873.
Schär, C., D. Leuenberger, O. Fuhrer, D. Lüthi, and C. Girard (2002): A New Terrain-Following Vertical Coordinate Formulation for Atmospheric Prediction Models, Mon. Wea. Rev., 130/10, 2459-2480.
Straka, J. M., R. B. Wilhelmson, L. J. Wicker, J. R. Anderson, K. K. Droegemeier (1993):Numerical solutions of a non-linear density current: a benchmark solution and comparisons, Int. J. Num. Meth. Fluids, 17, 1-22.
Weisman, M. L. and J. B. Klemp (1982): The dependence of Numerically Simulated Convective Storms on Vertical Wind Shear and Buoyancy, Mon. Wea. Rev., 110, 504-520.
Zängl, G., L. Ganthner, G. Hartjenstein, H. Noppel (2004): Numerical errors above steep topography: A model intercomparison, Met. Z., 13/2, 69-76.