The processflow performs post processing and diagnostics jobs automatically, removing many of the difficulties of performing post-run analysis on model data.
Documentation can be found here
Latest stable build from the master branch:
conda create --name <YOUR_NEW_ENVIRONMENT> -c e3sm -c conda-forge -c cdat processflow
Latest (potentially unstable but with the latest bug fixes) build from the nightly branch:
conda create --name <YOUR_NEW_ENVIRONMENT> -c conda-forge -c e3sm/label/nightly -c e3sm -c cdat -c pcmdi processflow
Update your environment to the newest version
conda update -c e3sm -c conda-forge -c cdat processflow
usage: processflow [-h] [-m MAX_JOBS] [-l LOG] [-a] [-r RESOURCE_PATH]
[--debug] [--dryrun] [-v] [-s]
[config]
positional arguments:
config Path to configuration file.
optional arguments:
-h, --help show this help message and exit
-m MAX_JOBS, --max-jobs MAX_JOBS
Maximum number of jobs to run at any given time
-l LOG, --log LOG Path to logging output file, defaults to
project_path/output/processflow.log
-a, --always-copy Always copy diagnostic output, even if the output
already exists in the host directory. This is much
slower but ensures old output will be overwritten
-r RESOURCE_PATH, --resource-path RESOURCE_PATH
Path to custom resource directory
--debug Set log level to debug
--dryrun Do everything up to starting the jobs, but dont start
any jobs
-v, --version Print version information and exit.
-s, --serial Run in serial on systems without a resource manager