RMA notes¶

In the RMA version of DART, the state vector is not required to be stored completely on any process. This is achieved using Remote Memory Access (RMA). The RMA programing model allows processes to read (and write) memory on other processors asynchronously. RMA DART supported models:

Before bitwise testing with Lanai please read Bitwise Considerations

NetCDF restarts¶

The programs filter and perfect_model_obs now read/write directly from NetCDF files, rather than having to run converters (model_to_dart and dart_to_model). To facilitate this, there is a new required call add_domain which must be called during static_init_model. It can be called multiple times in static_model_mod, e.g. once for each NetCDF file that contains state variables. There are three ways to add a domain:

From Blank : This is for small models such as lorenz_96 and no NetCDF restarts
- dom_id = add_domain(model_size)
From File : This is for models which have NetCDF restart files
- dom_id = add_domain(template_file, num_vars, var_names, ... )
From Spec : Creates a skeleton structure for a domain ( currently only used in bgrid_solo )
- dom_id = add_domain(num_vars, var_names, ... ) call add_dimension_to_variable(dom_id, var_id, dim_nam, dim_size) call finished_adding_domain

For models without NetCDF restarts, use add_domain(model_size). This is the minimum amount of information needed by DART to create a netdcf file. For models with NetCDF restarts use add_domain(info_file, num_vars, var_names) which lets DART read the NetCDF dimensions for a list of variables from a file (info_file). There are several routines that can be used together to create a domain from a description: add_domain, add_dimension_to_variable, finished_adding_domain. This can be used in models such as bgrid_solo where the model is spun up in perfect_model_obs, but the model itself has variable structure (3D variables with names). See Additions/Changes to existing namelists for how to use NetCDF IO.

Note when using NetCDF restarts, inflation files are NetCDF also. The inflation mean and inflation standard deviation are in separate files when you use NetCDF restarts. See Netcdf Inflation Files for details.

Calculation of forward operators¶

The forward operator code in model_mod now operates on an array of state values. See Forward Operator for more detail about distributed vs. non-distributed forward operators.

In distributed mode the forward operators for all ensemble members are calculated in the same model_interpolate call. In non-distributed mode, the forward oeprators for all ensemble members a task owns (1-ens_size) are calculated at once.

Vertical conversion of observation locations¶

The vertical conversion of observation locations is done before the assimilation. In Lanai this calculation is done in the assimilation as part of get_close_obs if a model_mod does vertical conversion. See Vertical Conversion of Observations for details about this change. Note that not all models do vertical conversion or even have a concept of vertical location, but every model_mod must have the following routines:

query_vert_localization_coord - returns the vertical localization coordiate (or does nothing if there is no vertical conversion)
vert_convert - converts location to required vertical (or does nothing if there is no vertical conversion)

Diagnostic file changes¶

For large models DART format diagnostic files (Prior_Diag.nc and Posterior_Diag.nc) have been replaced with separate files for each copy that would have gone into Prior_Diag.nc and Posterior_Diag.nc.

For Prior_Diag.nc:

Mean and standard deviation: preassim_mean.nc preassim_sd.nc
Inflation mean and standard deviation (if state space inflation is used): preassim_priorinf_mean.nc preassim_priorinf_sd.nc
The number of ensemble members specifed in filter_nml (num_output_state_members): preassim_member_####.nc

For Posterior_Diag.nc:

Mean and standard deviation: postassim_mean.nc postassim_sd.nc
Inflation mean and standard deviation (if state space inflation is used): postassim_priorinf_mean.nc postassim_priorinf_sd.nc
The number of ensemble members specifed in filter_nml (num_output_state_members): postassim_member_####.nc

The num_output_state_members are not written separately from the restarts. Note that restarts will have been clamped if any clamping is applied (given as an arguement to add_domain). This is different to Posterior_Diag.nc which contains unclamped values. Note also that there are 2 more “stages” which might be output, in addition to the preassim and postassim discussed here.

For models with multiple domains the filenames above are appended with the domain number, e.g. preassim_mean.nc becomes preassim_mean_d01.nc, preassim_mean_d02.nc, etc.

Changes to nc_write_model_atts¶

nc_write_model_atts has an new argument ‘model_mod_writes_state_variables’. This is used to communicate to DART whether the model will create and write state variables in Prior_Diag.nc and Posterior_Diag.nc. If model_model_writes_state_variables = .false. DART will define and write state variables to the new diagnostic files. If model_model_writes_state_variables = .true., nc_write_model_vars is called as normal.

Perturbations¶

The option to perturb one ensemble member to produce an ensemble is in filter_nml:perturb_from_single_instance. The model_mod interface is now pert_model_copies not pert_model_state. Each task perturbs every ensemble member for its own subsection of state. This is more complicated than the Lanai routine pert_model_state, where a whole state vector is available. If a model_mod does not provide a perturb interface, filter will do the perturbing with an amplitude set in filter_nml:perturbation_amplitude. Note the perturb namelist options have been removed from ensemble_manager_nml

State_vector_io_nml¶

&state_vector_io_nml
   buffer_state_io         = .false.,
   single_precision_output = .false.,
/

When buffer_state_io is .false. the entire state is read into memory at once if .true. variables are read one at a time. If your model can not fit into memory at once this must be set to .true. .

single_precision_output allows you to run filter in double precision but write NetCDF files in single precision.

Quality_control_nml¶

These namelist options used to be in filter_nml, now they are in quality_control_nml.

&quality_control_nml
   input_qc_threshold          = 3,
   outlier_threshold           = 4,
   enable_special_outlier_code = .false.
/

Additions/changes to existing namelists¶

New namelist variables

filter_nml¶

&filter_nml
   single_file_in               = .false.,
   single_file_out              = .false.,

   input_state_file_list        = 'null',
   output_state_file_list       = 'null',
   input_state_files            = 'null',
   output_state_files           = 'null',

   stages_to_write              = 'output'
   write_all_stages_at_end      = .false.
   output_restarts              = .true.
   output_mean                  = .true.
   output_sd                    = .true.

   perturb_from_single_instance = .false.,
   perturbation_amplitude       = 0.2_r8,

   distributed_state            = .true.
/

Item	Type	Description
single_file_in	logical	True means that all of the restart and inflation information is read from a single NetCDF file. False means that you must specify an input_state_file_list and DART will be expecting input_{priorinf,postinf}_{mean,sd}.nc files for inflation.
single_file_out	logical	True means that all of the restart and inflation information is written to a single NetCDF file. False means that you must specify a output_state_file_list and DART will be output files specified in the list. Inflation files will be written in the form i nput_{priorinf,postinf}_{mean,sd}.nc.
input_restart_files	character array	This is used for single file input for low order models. For multiple domains you can specify a file for each domain. When specifying a list single_file_in, single_file_out must be set to .true.
output_restart_files	character array	This is used for single file input for low order models. For multiple domains you can specify a file for each domain. When specifying a list single_file_in, single_file_out must be set to .true.
input_state_file_list	character array	A list of files containing input model restarts. For multiple domains you can specify a file for each domain. When specifying a list single_file_in, single_file_out must be set to .false.
output_state_file_list	character array	A list of files containing output model restarts. For multiple domains you can specify a file for each domain. When specifying a list single_file_in, single_file_out must be set to .false.
stages_to_write	character array	Controls which stages to write. Currently there are four options: `input` – writes input mean and sd only `preassim` – before assimilation, before prior inflation is applied `postassim` – after assimilation, before posterior inflation is applied `output` – final output for filter which includes clamping and inflation
write_all_stages_at_end	logical	True means output all stages at the end of filter. This is more memory intensive but requires less time. For larger models IO begins to dominate the overall cost of the assimilation, so writting all stages at the end writes more files in parallel, reducing the IO time. `output_state_file_list`.
output_restarts	logical	True means output a restart file(s). Filenames are defined in `output_state_file_list`.
output_mean	logical	True means output a restart file which contains the ensemble mean for the stages that have been turned on in `stages_to_write`. The file name will have the stage with `_mean` appended.
output_sd	logical	True means output a restart file which contains the ensemble standard deviation for the stages that have been turned on in `stages_to_write`. The file name will have the stage with `_sd` appended.
perturb_from_single_instance	logical	Read a single file and perturb this to create an ensemble
perturbation_amplitude	float	Perturbation amplitude
distribute_state	logical	True keeps the state distributed across all tasks throughout the entire execution of filter.

For NetCDF reads and writes

For input file names:

give input_state_file_list a file for each domain, each of which contains a list of restart files.
if no input_state_file_list is provided then default filenames will be used e.g. input_member_000*.nc, input_priorinf_mean.nc, input_priorinf_sd.nc

For output file names:

give output_state_file_list a file for each domain, each of which contains a list of restart files.
if no output_state_file_list is provided then default filenames will be used e.g. output_member_000*.nc, output_priorinf_mean.nc, output_priorinf_sd.nc

For small models you may want to use single_file_in, single_file_out which contains all copies needed to run filter.

Assim_tools_nml¶

&assim_tools_nml
   distribute_mean  = .true.
/

In previous DART releases, each processor gets a copy of the mean (in ens_mean_for_model). In RMA DART, the mean is distributed across all processors. However, a user can choose to have a copy of the mean on each processor by setting distribute_mean = .false. . Note that the mean state is accessed through get_state whether distribute_mean is .true. or .false.

Removed from existing namelists¶

&filter_nml
   input_qc_threshold          = 3,
   outlier_threshold           = 4,
   enable_special_outlier_code = .false.
   start_from_restart          = .false.
   output_inflation            = .true.
   output_restart              = .true.
   /

NOTE : output_restart has been renamed to output_restarts. ``output_inflation`` is no longer supported and only writes inflation files if inf_flavor > 1

&ensemble_manager_nml
   single_restart_file_out = .true.
   perturbation_amplitude  = 0.2,
   /

&assim_manager_nml
   write_binary_restart_files = .true.,
   netCDF_large_file_support  = .false.
   /