Mapping issues for high-res meshes

Owned by Mark Taylor
Last updated: Feb 14, 2020
6 min read

Document issues with our mapping file generation tool chain at high resolution. Instructions on using MBtempest: Offline remapping workflow with mbtempest

Status

For PG2 grids, we will mapping FV-FV. For monotone conservative linear maps, the only choice is cell integrated using a piecewise constant reconstruction. Options for producing these maps include ESMF, TempestRemap (TR), Moab with TR agorithms (MBTR) and NCO. All codes produce similar max and l2 errors when evaluated with the analytic vortex test field.

As of Jan 20, 2020

  • NCO: good results, fastest code. ~1.5H using 4 cores (4 threads) Cori-Haswell node. oRRS18to6 → ne1024pg2

  • TR: good results with code from Jan 17, 2020 and with the “--correct_areas” option. Very slow: 15H (1 core) oRRS18to6 → ne1024pg2

  • MBTR: maps are bad

  • ESMF: maps for ne1024 are bad.

     

For NP4 grids,:

  • TR from Jan 15, 2020: most maps are good, some combinations still produce crashes.

  • MBTR: maps are bad. is producing bad weights (NaNs or >> 1 ).

Analysis Mapping files

Mapping files are needed for remapping native grid output to lat/lon grids for conventional analysis tools. Often exact conservation is not required making TR’s integrated bilinear a good option: it’s accurate, has built in downscaling and is monotone so it wont introduce spurious oscillations making it good for plotting.

  • TR “intbilin”

    • several issues with loss of precision documented on https://acme-climate.atlassian.net/wiki/spaces/NGDNA/pages/831554328

 

SE GLL to SE GLL

These mapping files are mostly used to generate initial conditions. Remapping the cam.i file from a spunup NE30 simulation will generate a reasonably well spunup initial condition file for higher resolutions.

  • TR “mono” (with code fixes from Jan 15, 2020 )

    • ne30np4 → ne256np4

      • weights are good. row sum error 1.3e-15, conservation error: 5e-14.

      • Status: GOOD

    • ne30np4->ne512np4

      • Status: GOOD

  • MBTR “mono”

    • ne30np4->ne256np4

    • --monotonic 1: mapping weights have NaNs

    • --monotonic 3 --noconserve (“intbilin” option) weights are bad, max = 2.5

    • Status: NOT USABLE

  • ESMF “bilin”

    • ne30np4->ne256np4

      • weights between [0,1]. row sum error 2.2e-16. conservation error: 3.4e-1 (bilin is not conservative)

      • mapping file frac_a=0 (BUG?), but frac_b=1

      • Status: GOOD

    • ne120np4->ne512np4

      • Status: GOOD

    • ne120np4->ne1024np4

      • 10M out of 120M weights are bad, >> 1.0. max row sum = 5.0

      • Status: NOT USABLE

  • ESMF “aave”

    1. We dont consider these maps. We are trying to move away from these maps since they require the “dual grid”, which is difficult to produce and unnatural for GLL grids.

 

MPAS → np4 grids

The Jan 15, 2020 TR improvements suggest we should take another look at “mono” maps for MPAS/EAM coupling. For these maps, testing reported on Recommended Mapping Procedures for E3SM Atmosphere Grids recommended using the transpose of the np4->MPAS mono map. So we test the original map here:

  • TR “mono”

    • ne256np4 → oRRS18to6 ( with TR from Jan 17, 2020 )

      • weights: good. row sum error: 1e-15 . Conservation error: 4e-14

    • ne512np4 → oRRS18to6

      • ~4h to compute overlap mesh

      • ..EXCEPTION (src/LinearRemapSE0.cpp, Line 1016) Target grid must be a subset of source grid:

        Input mesh area (7.245674993333319e-06) Target area (2.898288183419045e-05)

    • ne1024np4 → oRRS18to6

MPAS → pg2 grids

We expect SCREAM will ultimately be running on the ne1024pg2 “phys grid”. To create these compsets, we first need to create a domain file which needs the monotone map from the MPAS ocean grid to the pg2 grid.

Note that these maps are FV to FV maps and thus avoid the issue of having to construct GLL dual grids even for ESMF maps.

Note2: these maps have a “mask_a,mask_b” and “frac_a,frac_b” variables. It appears “mask” is used to mask out point not used by that component model. For a global atmosphere grid, mask_b=1 for all points. For an MPAS grid, mask_a=1 for all points since the grid only contains ocean points. A POP grid is probably the only example where mask would contain both 0 and 1 values. The MPAS grid has frac_a=1 for all ocean points = all points in the grid. The mapping file value of frac_b ranges from 0..1 (we should check this) on the atmosphere grid, with coastlines given by the fractional values.

Note3: for TR FV to FV maps, the --mono options are ignored. monotone conservative maps are constructed using --in_np 1 (piecewise constant reconstruction). higher order maps (--in_np 2) are not monotone.

  • ncremap built in FV to FV map file generator

    • oRRS18to6 → ne256pg2 status: GOOD

      • ~3min with 4 threads on Cori frontend node

      • Weights are good. row sum and conservation error: 6e-13

    • oRRS18to6 → ne1024pg2

      • ~40 min on Compy with 4 threads. 1.5 h on Cori-Haswell node

      • Weights are good. row sum error 2e-12, conservation error 1e-12

  • TR “mono” . (TR from TR from Jan 17, 2020 with new --correct_areas option)

    • oRRS18to6 → ne256pg2 status: GOOD

      • 2.5h on Anvil (single processor)

      • Weights are good. Row sums error: 7e-16. Conservation error 7e-16

      • vortex test field l2/max error: 3.6145e-05 4.3881e-04

    • oRRS18to6 → ne512pg2: status GOOD

      • ~9h to compute overlap map

      • weights: good. row sum error: 7e-16. Conservation error 7e-16

      • vortex:     l2, linf: 1.7867e-04 2.2985e-03

    • oRRS18to6 → ne1024pg2: status: GOOD

      • 15.5h on Anvil, using 20GB

      • weights: good. row sum error: 4e-16. conservation error: 1e15

  • MBTR “mono”. Status: NOT USABLE

    • oRRS18to6 → ne256pg2

    • Weights are good, but 316 MPAS grid points have weight=0. This is incorrect and corrupts other diagnostics.

  • ESMF “aave”

    • oRRS18to6 → ne256pg2. Status: GOOD

      • weights are good. row sum max 8.5e-12. Conservation error 1.5e-11

      • vortex test field l2/max error: 3.6144e-05 4.3881e-04

    • oRRS18to6 → ne512pg2. Status: GOOD

      • weights are good. row sum error 3e-11. conservation error 1e-11

      • vortex:     l2, linf: 1.7866e-04 2.2985e-03

    • oRRS18to6 → ne1024pg2. Status: NOT USABLE

      • 35min on Cori Haswell node

      • ( /global/project/projectdirs/acme/taylorm/mapping/map_oRRS18to6v3_to_ne1024pg2_aave.nc.BAD )

      • weights are bad: max 2.6, row sum max 3.0. conservation error 4e-4

      • same issue with ESMF v8.

 

 

Outstanding issues

  • MBTR FV->FV error - missing source points. It appears MBTR maps have a similar loss of precision as TR maps, but this could be corrupted du to the missing source points.

  • MBTR SE->SE: montone weights are bad, either NaN or >1

  • ESMF “aave” and “bilin” maps are good up to ne512pg2, but both produce too large weights with ne1024pg2 (both SE->SE and FV->SE)

  • AnalyzeMap (Added in TR as of Feb 14, 2020 )

    • request: add min/max of map file frac_a, frac_b data to output

    • request: check for NaNs when printing weight min/max

    • for non global (MPAS) source grid, prints tens of thousands of useless consistency warnings

 

 

Command line options used:

New versions required:

  • TR: code from Jan 18, 2020 or later

  • NCO 4.9.2-alpha05 or newer

#atm grid set NE=1024 set atmgridnp4 = TEMPEST_NE${NE}.g set atmgrid = TEMPEST_NE${NE}pg2.g # used by TR, MBTR set atm_scrip = TEMPEST_NE${NE}pg2.scrip.nc # used by ESMF, ncremap # option "--alt" now the default and has to be removed from newer TR versions: GenerateCSMesh --res $NE --out_format Netcdf4 --file $atmgridnp4 GenerateVolumetricMesh --in $atmgridnp4 --out $atmgrid --np 2 --uniform ConvertExodusToSCRIP --in $atmgrid --out $atm_scrip # ocean grid: set target = ocean.oRRS18to6v3.scrip.181106.nc

Produce conservative monotone mapping:

# NCREMAP: set mapname = map_oRRS18to6v3_to_ne${NE}pg2_nco.nc ncremap -6 --thr_nbr=4 --grd_src=$target --grd_dst=$atm_scrip --map=$mapname ncatted -a grid_file_src,global,a,c,"$target" $mapname ncatted -a grid_file_dst,global,a,c,"$atm_scrip" $mapname #ESMF: set mapname = map_oRRS18to6v3_to_ne${NE}pg2_aave.nc srun -N 1 -c 4 -C haswell ESMF_RegridWeightGen -s $target -d $atm_scrip -i --src_regional \ --method conserve --netcdf4 -w $mapname #TR: set mapname = map_oRRS18to6v3_to_ne${NE}pg2_mono.nc GenerateOverlapMesh --b $atmgrid --a $target --out overlap.g GenerateOfflineMap --out_mesh $atmgrid --in_mesh $target --ov_mesh overlap.g \ --in_type fv --in_np 1 --out_type fv --out_double --out_format Netcdf4 \ --correct_areas --out_map $mapname

To check maps, use “ncks --chk_map”:

  • weights: verify that all weights are between [0,1], none of them are NaNs.

  • row sums: The row sum consistency check measures the ability of the map to preserve the constant function.

    • ocean->atm maps: max value should be close to 1.0. min value will be between [0..1] at land and fractional land points. Error reported above is from the max value.

    • atm->ocean maps: min/max value should be close to 1.0 for all rows with nonzero weights (since atmosphere grid is global)

    • atm->atm maps: min/max values should be close to 1.0

  • area weighted column sums: this measures if the map is conservative.

    • ocean->atm maps: area_b weighted column sum should equal area_a for all columns.

    • atm->ocean maps: area_b weighted column sum should equal area_a at all ocean points, and be between [0,1] at fractional (coastline) points.

    • atm->atm maps: equal for all columns.

 

Older versions of ncremap option is missing global attributes needed by CIME’s domain generation code. they can be added via:

  • ncatted -a grid_file_src,global,a,c,"$target" $mapname
    ncatted -a grid_file_dst,global,a,c,"$atm_scrip" $mapname