Page Comparison

As part of the efforts in the CMDV project, interfaces to integrate the MOAB unstructured mesh library with the TempestRemap remapping tool have been undertaken. Detailed information on the algorithmic and implementation aspects of this effort have been written in a manuscript submitted to Geoscientific Model Development [1]. This work has led to the development of a new offline remapping tool called mbtempest, which exposes the functionality to compute the supermesh or intersection mesh between two unstructured source and target component grids, in addition to using this supermesh for computing the remapping weights to project solutions between the grids. This functionality is part of the critical worflow with E3SM, where the generated remapping weights in the offline step are consumed by MCT at runtime to seamlessly transfer solution data between components (atm↔ocn, atm↔lnd, etc).

...

Partitioning meshes with the "inferred" strategy for better performance

A more recent update /wiki/spaces/ED/pages/2208170181 to the mbpart tool is to use the concept of inferred partitions such that the geometric locality on the source and target grids are preserved as much as possible to minimize communication at runtime during intersection mesh computation. This strategy has been shown to provide considerable speedup in the intersection mesh computation, and is now our preferred partitioning strategy in offline workflows, especially when one of the grids has topological holes (OCN mesh). In order to generate the inferred partitions, we usually choose the target mesh as the primary partition and the source mesh as the secondary partition. Then, the source mesh partitions are "inferred" based on the target mesh partition RCB tree. The commands to generate the inferred source partitions are shown below.

...

Note: the serial executables need to be launched with mpiexec -np 1, if on login node.  (otherwise Otherwise you may get encounter an error like :as shown below. 

> mbpart -h
Abort(1091087) on node 0 (rank 0 in comm 0): Fatal error in PMPI_Init: Other MPI error, error stack:
MPIR_Init_thread(136): 
MPID_Init(950).......: 
MPIR_pmi_init(168)...: PMI2_Job_GetId returned 14 )

...

The environment settings for running mbtempest on comps are listed below, and stored in the file: /compyfs/software/mbtempest.envs.sh for reference. 
This is version 5e41106dc9a2 from MOAB (>5.3.0) and version 72df14282a2e9 from tempestremap (>2.1.0)

module load cmake/3.11.4 intel/19.0.3 mvapich2/2.3.1 pnetcdf/1.9.0 mkl/2019u3 metis/5.1.0
export MPI_DIR=/share/apps/mvapich2/2.3.1/intel/19.0.3
export METIS_DIR=/share/apps/metis/5.1.0
export EIGEN3_DIR=/share/apps/eigen3/3.3.7/include/eigen3
export HDF5_DIR=/share/apps/netcdf-MPI/intel/19.0.5/mvapich2/2.3.2
export NETCDF_DIR=/share/apps/netcdf-MPI/intel/19.0.5/mvapich2/2.3.2
export PNETCDF_DIR=/share/apps/pnetcdf/1.9.0/intel/19.0.3/mvapich2/2.3.1
export ZOLTAN_DIR=/compyfs/software/zoltan/3.83/intel/19.0.3
export TEMPESTREMAP_DIR=/compyfs/software/tempestremap/intel/19.0.3
export MOAB_DIR=/compyfs/software/moab/intel/19.0.3

...

1. For the NE30 case, let us generate the CS mesh of required resolution using mbtempest.
   1. Command: $MOAB_DIR/bin/mbtempest -t 0 -r 30 -f outCSMesh30.nc
   2. Here, the type = 0 ( -t 0) specifies that we want to generate a CS grid, with element resolution = 30x30x6 (using -r 30).
2. Next, convert the NetCDF nc file format to a MOAB format, and in the process, also add some metadata for DoF numbering for the SE grid.
   1. Command: $MOAB_DIR/bin/mbconvert  -B -i GLOBAL_DOFS -r 4 outCSMesh30.nc outCSMesh30.h5m
   2. Here, the GLOBAL_DOFS is the tag that stores the DoF numbering for SE grid of order 4. The input "*.nc" mesh and output "*.h5m" mesh is specified as arguments for the format conversion.
3. The next step is to pre-partition the h5m file so that the map generation can be computed in parallel. In this particular example, we will use Metis Zoltan partitioner to generate 1024 128 parts.
   1. Command: $MOAB_DIR/bin/mbpart 1024 -m ML_RB 128 -z RCB outCSMesh30.h5m outCSMesh30_1024p128.h5m
4. Now that we have the ATM grid generated, let us perform a similar conversion on the OCN MPAS file. The MPAS nc file already exists and we will use this input file and convert it to a MOAB h5m file. During this process, unwanted edges and variables are not converted since the mbtempest mapping workflow only requires the actual mesh for computation of the overlap.
   1. Command: $MOAB_DIR/bin/mbconvert -O "variable=" -O "no_edges"  -O "NO_MIXED_ELEMENTS"  oEC60to30v3_60layer.170905.nc oEC60to30v3_60layer.170905.h5m
5. Similar to the CS grid case, let us now pre-partition the grid to 1024 128 parts using the Metis Zoltan Recursive-Bisection algorithm.
   1. Command: $MOAB_DIR/bin/mbpart 1024 -m ML_RB 128 -z RCB oEC60to30v3_60layer.170905.h5m oEC60to30v3_60layer.170905_1024p128.h5m
   We now have fully partitioned MOAB meshes for the CS and MPAS grids and all required inputs for mbtempest is available. Invoke the mbtempest command in parallel to generate the remapping weights after specifying the source and target grids, along with discretization detail specifications.
6. mpiexec -n 16 As mentioned above, better performance can be achieved by using the "inferred" partitioning strategy with Zoltan.
   1. Command: $MOAB_DIR/bin/mbtempest -t 5 -w -l outCSMesh30_1024.h5m -l mbpart 128 -z RCB -b --scale_sphere -p 2 oEC60to30v3_60layer.170905.h5m oEC60to30v3_60layer.170905_1024p128.h5m h5m -m cgll -o 4 -m fv -o 1 -g GLOBAL_DOFS -g GLOBAL_ID -f mapSEFV-NE30.h5m
   2. The particular example above runs on 16 processes, and takes the pre-partitioned input grids outCSMesh30_1024.h5m and oEC60to30v3_60layer.170905_1024.h5m for CS and MPAS respectively
   3. We also specify that the source discretization method is Spectral Element (SE) with continuous representation of DoFs on the element interfaces and the target discretization on MPAS grid is Finite Volume (fv). This option is specified using the -m input parameter, whose default is fv.
   4. The order of the discretization is then specified using the -o options for input and output models. In the above case, we have SE order = 4 and FV order = 1.
   5. Next, we also need to specify the tags in the mesh that contain the source and target global DoF numbers that are stored on their corresponding elements. This will dictate the ordering of the mapping weight matrix that is written out to file.
   6. The final set of argument specifies that the output map file is to -inferred outCSMesh30.h5m
   7. Rename outCSMesh30_inferred.h5m to outCSMesh30_p128.h5m to be consistent
   8. The above command will generate oEC60to30v3_60layer.170905_p128.h5m and outCSMesh30_p128.h5m meshes, which are optimized in term of geometric locality for parallel runs on 128 processes.
      
7. We now have fully partitioned MOAB meshes for the CS and MPAS grids (either from step (3)+(5) or from (6)), and all required inputs for mbtempest is available. Invoke the mbtempest command in parallel to generate the remapping weights after specifying the source and target grids, along with discretization detail specifications.
   1. Command: srun -n 64 $MOAB_DIR/bin/mbtempest -t 5 -w -l outCSMesh30_p128.h5m -l oEC60to30v3_60layer.170905_p128.h5m -m cgll -o 4 -g GLOBAL_DOFS -m fv -o 1 -g GLOBAL_ID -i intx_ne30_oEC60to30v3.h5m -f mapSEFV-NE30.nc
   2. The particular example above runs on 64 processes, and takes the pre-partitioned input grids outCSMesh30_p128.h5m and oEC60to30v3_60layer.170905_p128.h5m for CS and MPAS respectively
   3. We also specify that the source discretization method is Spectral Element (SE) with continuous representation of DoFs on the element interfaces and the target discretization on MPAS grid is Finite Volume (fv). This option is specified using the -m input parameter, whose default is fv.
   4. The order of the discretization is then specified using the -o options for input and output models. In the above case, we have SE order = 4 and FV order = 1.
   5. Next, we also need to specify the tags in the mesh that contain the source and target global DoF numbers that are stored on their corresponding elements. This will dictate the ordering of the mapping weight matrix that is written out to file.
   6. The final set of argument specifies that the output map file is to be written out to mapSEFV-NE30.h5m for the NE30 case in parallel.
8. Now that we have the parallel remapping weights generated in h5m format, we need to re-convert it back to a SCRIP nc file in order to be consumed by E3SM. For this purpose, we use a special "serial" tool to convert the h5m file to SCRIP.
   1. $MOAB_DIR/bin/h5mtoscrip -d 2 -w mapSEFV-NE30.h5m -s mapSEFV-NE30.nc
   2. The -w argument takes the input map file in h5m format and the -s argument takes the output SCRIP filename.
9. At the end of this workflow, we now have a SCRIP file containing the weights to compute a solution projection from an input CS NE30 grid with SE(4) discretization to an output MPAS grid with FV(1) discretization.

Building your own version of the mbtempest tool locally

1. At the end of this workflow, we now have a SCRIP file (mapSEFV-NE30.nc) containing the weights to compute a solution projection from an input CS NE30 grid with SE(4) discretization to an output MPAS grid with FV(1) discretization.
2. Exercise: rerun step (7) with source discretization specification: `-m fv -o 1 -g GLOBAL_ID`. This results in a FV-FV map between the NE30 grid and the OCN MPAS grid. You can also switch the order of -l arguments to generate the weights in the reverse direction here i.e., switch source and target grids/discretization specifications for mbtempest.

Building your own version of the mbtempest tool locally

In order to build the MOAB-TempestRemap stack with parallel MPI launch support, we suggest the following list of commands. First define an installation prefix directory where the stack of library, includes and tools will be installed. Let us call this as the $INSTALL_PREFIX environment variable.

...

1. MPI-enabled C, C++, and Fortran compiler wrappers that are exported in the local environment as $CC, $CXX, and $FC.
2. Next, verify installations of dependent libraries such as $HDF5_DIR and $NETCDF_DIR that have been compiled with MPI support using the $CC, $CXX, $FC compilers.
3. Get Eigen3 package from the webpage and untar to the $INSTALL_PREFIX/eigen3 directory with the following command
   1. Download: wget https://bitbucket.org/eigen/eigen/get/3.3.7.tar.gz  OR  curl https://bitbucket.org/eigen/eigen/get/3.3.7.tar.gz -O
   2. Move: mv eigen-eigen* $INSTALL_PREFIX/eigen3
   3. export EIGEN3_DIR=$INSTALL_PREFIX/eigen3

Dependencies and pre-requisites from an existing E3SM use case

It is recommended to use the same dependent libraries as a regular E3SM case. 
E3SM cases save the environment in files like  .env_mach_specific.sh in the case folder. That is a good environment to start building your tempestremap, MOAB or Zoltan dependencies. Or maybe they are already built on your machine.  That environment is created from config_machines.xml or config_compiler.xml files, and these change all the time, as new releases, use cases and tests become available. Problems can appear for MOAB's mbtempest if the HDF5 library that netcdf4 is built on does not have a good MPI support. Or if (gasp!) hdf5 is built in serial. Then you are limited on building mbtempest without parallel support, which means you are better off by just running tempestremap in serial. Do not bother with building MOAB. 
On compy, the netcdf used for E3SM is built with serial hdf5, so it cannot be used for MOAB. This is why on compy we have a separate netcdf, built with parallel hdf5. 

Build

To get the entire (MOAB-TempestRemap) stack working correctly, we need to find parallel-enabled dependency installations for HDF5 and NetCDF that are built with MPI library support for the current architecture

...

If steps (a)-(g) pass successfully, the MOAB libraries and tools, along with interfaces for TempestRemap will be installed in $INSTALL_PREFIX/moab directory. The offline remapping weight computation tool, mbtempest, will also be installed during this process and can then be used standalone to generate the weight files as needed.

References

1 Mahadevan, V. S., Grindeanu, I., Jacob, R., and Sarich, J.: Improving climate model coupling through a complete mesh representation: a case study with E3SM (v1) and MOAB (v5.x), Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2018-280, in review, 2018.