petascale atmospheric models for the ccsm: new ... · “reference” solutions •evaluation...
TRANSCRIPT
Petascale Atmospheric Models for the CCSM: New
Developments and Evaluation of Scalable Dynamical Cores
Mark Taylor (Sandia)
Jim Edwards (NCAR/IBM), Christiane Jablownowski (Univ. Michigan), Peter Lauritzen (NCAR), Ram Nair (NCAR), Amik St.Cyr (NCAR)
SciDAC PI Meeting, Seattle, July 2008
Office of Science
U.S. Department of Energy
Why Petascale Climate Models?
• DOE BER Scidac Project to build a first generation Earth System Model
– Based on the Community Climate System Model (CCSM)
– Full carbon cycle (and other cycles) to allow for prescribing greenhouse gas emissions instead of prescribing concentrations
– Improved regional resolution to better assess potential impacts of climate change
– These new capabilities will require petascale performance
• Many other modeling centers have similar goals & plans
Outline
• Largest scalability bottleneck in CCSM: – Dynamical core of atmospheric component
• 2008 NCAR ASP Summer Colloquium on Numerical Techniques for Global Atmospheric Models
– International & collaborative efforts to develop petascale-ready atmospheric dynamical cores
– Focus on model inter-comparison using idealized test cases with “reference” solutions
• Evaluation within the full atmospheric model– No convergence under mesh refinement! Use Williamson
equivalent resolution methodology for Aqua Planet
– Results using 90K processors from a cubed-sphere based dynamical core in the CCSM.
Community Climate System Model(CCSM)
Model Components
• Petascale Ready: scaling to O(100,000) processors at target climate resolutions (10km)
• Land and Ice models are petascale-ready
• Ocean component is near petascale-ready
• Atmosphere component is the performance bottleneck
Community Atmospheric Model(CAM)
• Atmospheric Model: Physics component– Mostly subgrid parametrizations and forcings (convection,
precipitation, radiative forcing, etc.)
– Appear as forcing terms in the equations
– Column based, embarrassingly parallel with 2D domain decomposition
– Responsible for most of the uncertainty
• Atmospheric Model: Dynamics component– Solves the Atmospheric Primitive Equations (compressible Euler
equations in a rotating reference frame with hydrostatic and shallow atmosphere approximation, scale selective dissipation)
– Responsible for most of the variability
– Scalability bottleneck!
The Dynamical Core Scalability Bottleneck
•Most dynamical cores in operational models use latitude-longitude grids: – Well proven. Many good solutions to
the “pole problem”: Spherical harmonics, polar filtering, implicit methods
– But these approaches degrade parallel scalability
– However, scalability continues to improve (see SciDAC BER/ASCR Scientific Application Partnership)
The Dynamical Core Scalability Bottleneck
•Petascale dynamical core:– Quasi-uniform grids avoid the pole problem
– Can use full 2D domain decomposition in horizontal directions
– Each column in the vertical/radial direction kept on processor
– Equations can be solved explicitly, only nearest neighbor communication
– Numerical methods that perform as well as lat/lon based methods remains a challenge.
• NCAR ASP Summer Colloquium on Dynamical Cores– June 2-13, 2008, Boulder CO
– Organizers: Lauritzen, Jablonowski, Taylor, Nair
– NCAR/NSF, NASA and DOE Support
– ~40 graduate students
– 12 keynote lecturers (See upcoming Springer Lecture Notes in Computational Science and Engineering)
– 11 models: GISS BQ, CAM Eulerian, CAM FV isentropic, GEOS FV, GEOS FV-cubed, CSU GCM, GME, HOMME, ICON, OLAM, NCEP GEF, MIT GCM
Summer Colloquium on Dynamical Cores
•Students ran 6 dynamical core test cases (some with several variants)
– Steady-state (3 rotations)– Baroclinic instability (3 rotations)– Pure 3D advection with prescribed winds– Rossby-Haurwitz wave– Mountain-induced Rossby waves– Gravity wave tests (4 variants)
•All models tested with identical initial conditions and run in their operational configurations which included their typical diffusion mechanisms. Most models are hydrostatic and based on the Primitive Equations.
•Results shown here are for medium high horizontal resolution (about 1 degree). Vertical resolution varied from 20 to 60 levels.
•1 TB of data collected, to be put on the Earth System Grid
•https://www.wiki.ucar.edu/display/dycores/Home
Jablonowski and Williamson, A Baroclinic Instability Test Case for Atmospheric Model Dynamical Cores, Q.J.R. Meteorol. Soc. (2006)
– Dynamical core only: no atmospheric physics– L2 error in surface pressure as a function of
time shown below– Converges under mesh refinement to
reference solution (uncertainty in reference solution is yellow shaded region)
Test 2: Baroclinic Instability Test
Test 2: Baroclinic instability. Surface pressure at day 9. The tests starts with balanced initial conditions that are overlaid by a Gaussian hill perturbation. The perturbation grows into a baroclinic wave. Some models show cubed-sphere or icosahedral grid imprinting in the Southern Hemisphere. High order methods show spectral ringing in the 1000mb contour.
GISS-BQ CAM EUL CAM-FV isen
GME
OLAMICON
GEOS-FVcubeGEOS-FV
HOMME
Test 3: Pure Advection. Latitude-height cross section of a 3D slotted ellipse tracer distribution after one revolution around the sphere (day 12). The 3D winds are prescribed. The slotted ellipse has followed a trajectory path with three wave cycles in the vertical direction. The test evaluates the diffusion characteristics of the advection algorithm.
GISS-BQ CAM EUL CAM-FV isen
GME
OLAMICON
GEOS-FVcubeINITIAL STATE
HOMME
Test 4: Westward propigating Rossby-Haurwitz wave. Geopotential height is plotted at 500mb at day 15. The wave number 4 pattern moves westward with minimal change in shape.
GISS-BQ CAM EUL CAM-FV isen
GME
OLAMICON
GEOS-FVcubeGEOS-FV
HOMME
Test 5: Mountain induced Rossby wave. Zonal wind at 700mb, day 15. The test starts with balanced and isothermal initial conditions. A 2-km high Gaussian-hill shaped mountain with half-width 1.5 km is placed at 90°E, 30°N. The mountain triggers Rossby waves. The test evaluates the treatment of the orography and reveals numerical noise (especially at later days).
GISS-BQ CAM EUL CAM-FV isen
GME
OLAMICON
GEOS-FVcubeGEOS-FV
HOMME
Test 6: Pure Gravity Wave Test. Longitude-height cross section along the equator of the potential temperature perturbation at day 4. The Coriolis parameter is set to zero. The test reveals the influence of the diffusion and damping mechanisms in the dynamical cores. Divergence damping (in FV models) leads to a significant decrease in the gravity wave amplitude. Diffusion also influences the sharpness of the gradient.
GISS-BQ CAM_EUL CAM-FV isen
GME
OLAMICON
GEOS-FVcubeGEOS-FV
HOMME
Dynamics and Physics Evaluation
Neale & Hoskins, 2000a: A standard test for AGCMs including their physical parameterizations, Atmos. Sci. Lett.Compare dycore behavior in full atmospheric model (moisture and full physics), but still in an idealized settingStresses dissipation mechanisms and moisture advection much more than pure dynamical tests.
Aqua Planet Experiments
HOMME cubed-sphere dycore integrated in CAM, the Community Atmospheric Model component of the CCSM
CAM Eulerian (global spectral model) and CAM-FV results taken from Williamson, Tellus 2008a, 2008b
Follow Williamson equivalent resolution methodology – No convergence under mesh refinement, as expected due to the nature of many of
the subgrid physics parametrizations
– Strong signal with resolution
– However, agreement between dynamical cores allows us to establish equivalent resolutions
– CAM 3.1 Physics with parameters held fixed for all models and all resolutions
– Compare 1 year means (after suitable spinup)
CAM/HOMME Spectral Element Dynamical Core
h-p Finite element method on a cubed-sphere grid
4th order accurate
1st dynamical core in CAM to conserve both Mass and Energy– Conservation via Compatible Discretization
– Energy conservation even for equations in non-conservation form
– Zero dissipation from the discretization
1st dynamical core in CAM to allow full 2D domain decomposition
Tracer advection: consistent with continuity equation, positive preserving, but not monotone
Resolution Viscosity PRECC PRECL CLDTOT TMQ
EUL T42 5m 1.0E+16 1.71 1.11 0.64 20.21
HOMME 1.9 5m 1.0E+16 1.76 1.14 0.66 20.09
EUL T85 5m 1.0E+15 1.59 1.38 0.60 19.63
HOMME 1.0 5.5m 1.0E+15 1.59 1.43 0.61 19.67
EUL T170 5m 1.5E+14 1.44 1.62 0.55 19.13
HOMME 0.5 5m 1.5E+14 1.48 1.62 0.55 19.36
T340 5m 1.5E+13 1.36 1.75 0.50 18.75
Physics dt
Aqua Planet Global Mean Quantities
Compared to the size of the resolution signal, there is a remarkable agreement between CAM/HOMME and CAM/Eulerian
Precipitation PDFsH
OM
ME
FV
& E
uler
ian
1mm bin-size 10mm bin-size
Aqua Planet Experiment: Zonal DataComparison with FV & Eulerian Dycore
HO
MM
EF
V &
Eul
eria
n
Fixed Mesh Scalability CAM/HOMME
•Good scalability down to 1 element per processor (86,200 processors at 0.25 degree resolution). Higher resolutions will easily scale to even more processors
• Integration rates better than 5 simulated years/day at resolutions down to 0.25 degree
•BGL results: 1 processor per node due to memory constraints. BGP results use 4 processor cores per node. BGP is 4x-8x faster per node.
High Resolution Aqua Planet Experiments
CAM 3.5 Physics
14 month simulations
5min physics timestep
Eulerian T85 physics tunings
Simulations on LLNL BG/L system–0.250 degree: 43200 processors ~1 day per
simulation–0.125 degree: 57600 processors ~3 days per
simulation
High Resolution Results - CAM 3.4 Physics
Summary
• Community is committed to developing petascale-ready earth system models
• Atmospheric component: – Unprecedented collaboration between modeling groups on next
generation dynamical cores initiated at NCAR Colloquim
– CAM/HOMME: Petascale-ready cubed-sphere dynamical core integrated into CAM
– CAM/HOMME: aqua planet test case results agree remarkably well with conventional dynamical cores
• Current Focus: (BER SciDAC projects)– CCSM coupling with land, ocean and ice (Mariana Vertenstein, Tony
Craig, Kate Evans)
– Discontinious Galerken advection schemes for HOMME (Ram Nair, Amik St.Cyr, Henry Tufo)