performance of a semi-implicit, semi-lagrangian dynamical core for high resolution nwp over...
DESCRIPTION
Computational gridTRANSCRIPT
![Page 1: Performance of a Semi-Implicit, Semi-Lagrangian Dynamical Core for High Resolution NWP over Complex…](https://reader035.vdocuments.site/reader035/viewer/2022070616/5a4d1bea7f8b9ab0599e3f33/html5/thumbnails/1.jpg)
Performance of a Semi-Implicit, Semi-Lagrangian Dynamical Core
for High Resolution NWP over Complex Terrain
L.Bonaventura D.Cesari
![Page 2: Performance of a Semi-Implicit, Semi-Lagrangian Dynamical Core for High Resolution NWP over Complex…](https://reader035.vdocuments.site/reader035/viewer/2022070616/5a4d1bea7f8b9ab0599e3f33/html5/thumbnails/2.jpg)
Outline of discretization approach
• Semi-implicit, semi-lagrangian two time level discretization (2d tests: Bonaventura JCP 2000)
• Finite volume approach for the divergence• Weakly nonlinear system Ax+f(x)=b, symmetric
and well-conditioned for any orography• 2nd order in time, 4th order in space SL
advection scheme • Improved interpolation at the boundary for SL
advection (joint work with G.Rosatti, University of Trento)
![Page 3: Performance of a Semi-Implicit, Semi-Lagrangian Dynamical Core for High Resolution NWP over Complex…](https://reader035.vdocuments.site/reader035/viewer/2022070616/5a4d1bea7f8b9ab0599e3f33/html5/thumbnails/3.jpg)
Computational grid
![Page 4: Performance of a Semi-Implicit, Semi-Lagrangian Dynamical Core for High Resolution NWP over Complex…](https://reader035.vdocuments.site/reader035/viewer/2022070616/5a4d1bea7f8b9ab0599e3f33/html5/thumbnails/4.jpg)
Application to open channel flow
![Page 5: Performance of a Semi-Implicit, Semi-Lagrangian Dynamical Core for High Resolution NWP over Complex…](https://reader035.vdocuments.site/reader035/viewer/2022070616/5a4d1bea7f8b9ab0599e3f33/html5/thumbnails/5.jpg)
2D Gallus-Klemp lee wave test case
Vorticity=O(10e-4)
![Page 6: Performance of a Semi-Implicit, Semi-Lagrangian Dynamical Core for High Resolution NWP over Complex…](https://reader035.vdocuments.site/reader035/viewer/2022070616/5a4d1bea7f8b9ab0599e3f33/html5/thumbnails/6.jpg)
3D linear lee waves
![Page 7: Performance of a Semi-Implicit, Semi-Lagrangian Dynamical Core for High Resolution NWP over Complex…](https://reader035.vdocuments.site/reader035/viewer/2022070616/5a4d1bea7f8b9ab0599e3f33/html5/thumbnails/7.jpg)
Setup of tests for comparison of numerical efficiency
• 180*180*40 =O(1.3e+6) gridpoints• dx=2 km, dz=300 m, dt=40 s• Some 1000 m high mountains• A -8 K temperature anomaly (cold bubble) on top of
each mountain• IBM SP4 of CINECA (48 nodes with 16 CPUs each,
1.5 Ghz for each CPU)• native xlf90 compiler and all standard optimizations
switched on• Comparison against LokalModell of DWD
![Page 8: Performance of a Semi-Implicit, Semi-Lagrangian Dynamical Core for High Resolution NWP over Complex…](https://reader035.vdocuments.site/reader035/viewer/2022070616/5a4d1bea7f8b9ab0599e3f33/html5/thumbnails/8.jpg)
Parallel run with 16 processorsTotal CPU time for 1 hour
CPU time solver
COMM time solver
CPU time advection
COMM time advection
Fastest/slowest ratio
SE 328 s 91 s 13.4 s 156 s 40.8 s 1.07
SI Z
207 s 119 s 13.4 s 65.4 s 7.4 s 1.02
![Page 9: Performance of a Semi-Implicit, Semi-Lagrangian Dynamical Core for High Resolution NWP over Complex…](https://reader035.vdocuments.site/reader035/viewer/2022070616/5a4d1bea7f8b9ab0599e3f33/html5/thumbnails/9.jpg)
Parallel run with 36 processors(3D lee wave test)
Total CPU time for 1 hour
CPU time solver
COMMtime solver
Fastest/slowest ratio
SE 88.95 s 45.03 s 11.95 s 1.4 s
SI Z
56.40 s 26.16 s 5.12 s 1.03 s
![Page 10: Performance of a Semi-Implicit, Semi-Lagrangian Dynamical Core for High Resolution NWP over Complex…](https://reader035.vdocuments.site/reader035/viewer/2022070616/5a4d1bea7f8b9ab0599e3f33/html5/thumbnails/10.jpg)
Comparison of speed up
![Page 11: Performance of a Semi-Implicit, Semi-Lagrangian Dynamical Core for High Resolution NWP over Complex…](https://reader035.vdocuments.site/reader035/viewer/2022070616/5a4d1bea7f8b9ab0599e3f33/html5/thumbnails/11.jpg)
Comparison with terrain following semi-implicit LM
• Difficulty of a fair comparison (different stopping criteria and implementation details)
• Slower convergence of iterative solver for terrain following semi-implicit
Residual 1%of initial value
Residual 0.1% of initial value
Residual 0.01% of initial value
SI iterations 6 21 50SIZ iterations 8 17 21
![Page 12: Performance of a Semi-Implicit, Semi-Lagrangian Dynamical Core for High Resolution NWP over Complex…](https://reader035.vdocuments.site/reader035/viewer/2022070616/5a4d1bea7f8b9ab0599e3f33/html5/thumbnails/12.jpg)
Conclusions
• Semi-implicit, semi-lagrangian method using finite volume discretization of the divergence and improved lower boundary condition
• Gain in efficiency over terrain following coordinates, good parallel performance
![Page 13: Performance of a Semi-Implicit, Semi-Lagrangian Dynamical Core for High Resolution NWP over Complex…](https://reader035.vdocuments.site/reader035/viewer/2022070616/5a4d1bea7f8b9ab0599e3f33/html5/thumbnails/13.jpg)
Development plans
• Further testing (Boulder storm, Scania test)
• ARPA-SMR: development of a full SI-SL NWP model in the framework of the COSMO consortium adapting the LokalModell physics
• MPI: test tube for numerical methods to be used in ICON, the new global nonhydrostatic dynamical core
![Page 14: Performance of a Semi-Implicit, Semi-Lagrangian Dynamical Core for High Resolution NWP over Complex…](https://reader035.vdocuments.site/reader035/viewer/2022070616/5a4d1bea7f8b9ab0599e3f33/html5/thumbnails/14.jpg)