xuejingkai/dart-wrf

Observing system simulation experiments with WRF and DART

DART-WRF

This code runs an Ensemble Data Assimilation system with the software packages DART and WRF. All workflow steps are submitted to the cluster manager SLURM, which takes care of the dependencies (the order in which tasks are done).

• Why should I use it?

  • It's pythonic: see what it does at first glance, modular, flexible
  • It handles dependencies using SLURM without 'sleep loops in the bash script'. Functions return a SLURM ID which can be used to trigger the start of another function (interface by brentp/slurmpy).

• How does it work?

  • There are two kinds of code/scripts: One that is run immediately and one that is run later on.
  • The first one instructs the cluster (SLURM) to do the things you tell it - in the right order - using the right input.
  • The latter kind of code actually does the things. Its code is not executed right away, but when there are resources assigned to you by SLURM.

• Can I use it for real weather?

  • Yes, but you need to convert your observations into DART format.

Workflow

Configure your experiment

Define simulation specific variables in config/cfg.py. Define paths for python, ncks, etc. in config/clusters.py. Dependencies are numpy, pandas, scipy, xarray, netCDF4. Install non-standard packages with pip install docopt slurmpy --user. Workflow is defined using meta-routines (functions) like run_ENS which are defined in scheduler.py.

Prepare initial conditions (from input_sounding)

1. Define starting time: begin = dt.datetime(2008, 7, 30, 6)
2. WRF needs directories with certain files: id = prepare_WRFrundir(begin)
3. Create 3D initial conditions from input_sounding etc.: id = run_ideal(depends_on=id)

Run free forecast

Let's say you want to run a free forecast starting at 6z, which you want to use as prior for an assimilation at 9z. Then you need can use the above defined 3 steps to create initial conditions. Then you can run an ensemble forecast using:

id = run_ENS(begin=begin,  # start integration from here
             end=end,      # integrate until here
             input_is_restart=False,
             output_restart_interval=(end-begin).total_seconds()/60,
             depends_on=id)

where begin & end are dt.datetime objects.

Assimilation experiment

Assimilate

To assimilate observations at dt.datetime time use this command:

id = assimilate(time, prior_init_time, prior_valid_time, prior_path_exp, depends_on=id)

Update initial conditions from Data Assimilation

In order to continue after assimilation you need the posterior = prior (1) + increments (2)

1. Set prior with this function:

id = prepare_IC_from_prior(prior_path_exp, prior_init_time, prior_valid_time, depends_on=id)

where path is str, times are dt.datetime.

2. To update the model state with assimilation increments, you need to update the WRF restart files by running

id = update_IC_from_DA(time, depends_on=id)

After this, the wrfrst files are updated with assimilation increments (filter_restart) and copied to the WRF's run directories so you can continue to run the ENS after assimilation using

id = run_ENS(begin=time,  # start integration from here
             end=time + timedelta_integrate,  # integrate until here
             restart_path=cluster.archivedir+prior_init_time.strftime('/%Y-%m-%d_%H:%M/'),
             output_restart_interval=timedelta_btw_assim.total_seconds()/60,
             depends_on=id)

where times are dt.datetime; timedelta variables are dt.timedelta.

Examples

scheduler.py generate_free.py

Finally

SLURM submissions

scheduler.py submits jobs into the SLURM queue with dependencies, so that SLURM starts the jobs itself as soon as resources are available. Most jobs need only one node, but model integration is done in a SLURM job array across e.g. 10 nodes:

$ squeue -u `whoami` --sort=i
             JOBID PARTITION     NAME     USER ST       TIME  NODES NODELIST(REASON)
           1710274  mem_0384 prepwrfr  lkugler PD       0:00      1 (Priority)
           1710275  mem_0384 IC-prior  lkugler PD       0:00      1 (Dependency)
           1710276  mem_0384 Assim-42  lkugler PD       0:00      1 (Dependency)
           1710277  mem_0384 IC-prior  lkugler PD       0:00      1 (Dependency)
           1710278  mem_0384 IC-updat  lkugler PD       0:00      1 (Dependency)
           1710279  mem_0384 preWRF2-  lkugler PD       0:00      1 (Dependency)
    1710280_[1-10]  mem_0384 runWRF2-  lkugler PD       0:00      1 (Dependency)
           1710281  mem_0384 pRTTOV-6  lkugler PD       0:00      1 (Dependency)
           1710282  mem_0384 Assim-3a  lkugler PD       0:00      1 (Dependency)
           1710283  mem_0384 IC-prior  lkugler PD       0:00      1 (Dependency)
           1710284  mem_0384 IC-updat  lkugler PD       0:00      1 (Dependency)
           1710285  mem_0384 preWRF2-  lkugler PD       0:00      1 (Dependency)
    1710286_[1-10]  mem_0384 runWRF2-  lkugler PD       0:00      1 (Dependency)
           1710287  mem_0384 pRTTOV-7  lkugler PD       0:00      1 (Dependency)

Easily switch between clusters

Define cluster specific variables in config/clusters.py :

clusterA = ClusterConfig()
clusterA.name = 'vsc'
clusterA.userdir = '/home/pathA/myuser/'
...
clusterB = ClusterConfig()
clusterB.name = 'jet'
clusterB.userdir = '/home/pathB/myuser/'

References

This workflow was created following the DART-WRF Tutorial. Read the DART documentation: docs.dart.ucar.edu. DART is available at github: @NCAR/DART.

License

This repo is licensed under Apache License 2.0

@NCAR/DART is licensed under the Apache License, Version 2.0 Copyright 2019 University Corporation for Atmospheric Research