Inline help:
python3 submit.py --help
Create the Job configuration using the command:
python3 submit.py -f submit.yaml --create
Submit the jobs to the queues via the command:
python submit.py -f submit.yaml --submit
The conf file uses the YAML
format. You can find several examples in this directory. Most of the parameters should be self explenatory.
Each task corresponds to a sample.
The tool supports:
-
HT-Condor local submission: each task will be mapped to a dedicated condor cluster. Input files can be specified via the dataset name using
input_dataset
or via the directory where the files sit usinginput_directory
. The implemented input splitting modes are:file_based
andlumi_based
(in the sense of LS). For thejob_flavor
and other condor parameters please refer to the CERN Condor documentation -
CRAB remote submission: this is controlled setting the
crab: True
parameter. In this case the input needs to be specified usinginput_dataset
and for the configuration ofsplitting_mode
andsplitting_granularity
you need to refer to the CMS CRAB Sw Guide. The rest of the parameters can be customized via the template filecrab_MODE.py
(see the documentation about the template structure in the following).
If you want to add a new set of templates you need to define a new mode (MODE
) in:
https://github.com/cerminar/submission/blob/master/submit.py#L101
Note the code assumes that all keys starting with TEMPL_
are variables in the templated files that need to be replaced.
And create the corresponding template files in the directory
templates/
You need 1 or 2 templated files:
condorSubmit_MODE.sub
: this defines the structure of the Condor submission file. If you don't need to modify it you can use thecondorSubmit_DEFAULT.sub
one (see CERN Condor documentation);jobCustomization_MODE_cfg.py.
which defines what needs to be changed for each job of the cluster.run_MODE.sh
: this is the script which actually gets executed on the node, setups the config and calls cmsRun. Again arun_DEFAULT.sh
exists and should work for most cases.