GenerateFcl
Introduction
This procedure is used to generate a set of fcl files, one for each job to be run in a project. The fcl might drive a simulation project that starts with a generator, or a later stage of simulation starting with the art file output of an earlier stage, or a file concatenation project, or even an analysis project.
You would typically come to this page through the MCProdWorkflow procedure.
Preparation
A fcl file should be developed and verified
with interactive jobs before it is prepared for a grid job.
Several examples of simulation fcl files are under the JobConfig
subdirectory of Offline.
If a fcl file is working interactively, it will also work on the grid, with the following caveat. Using the fcl with the production system imposes one extra requirement: all output files of the job should satisfy the Mu2e naming conventions. You must be familiar with the naming convention and its fields or this documentation will not make sense.
Files under the JobConfig directory, intended for production, for example,
write output files like:
TFileService : { fileName : "nts.owner.cd3-beam-g4s1.configuration.sequencer.root" }
...
fileName : "sim.owner.cd3-beam-g4s1-mubeam.configuration.sequencer.art"
The values of the owner, configuration,
and sequencer fields used in the prepared fcl file are
not important, because they will be overridden later when the fcl is generated. The values of
data_tier, description,
and file_format will be used as-is and must be set
correctly, please see file names for guidelines on how to use these fields.
Use the "art" extension ("file format") for framework
outputs (written by RootOutput modules) and "root"
for TFileService (ntuple and histogram) file output files.
Setup
Setup the utilities
setup mu2e source the appropriate Offline setup.sh file for your project setup mu2etools setup mu2efiletools
The setup of mu2etools above makes the generate_fcl command available in your path.
When you write to the SAM database, usually by mu2eFileDeclare, you will need to have a valid certificate.
Generating fcl
There are two invocation modes that require mutually exclusive sets of parameters
- For jobs with
EmptyEventinput source (events created by a generator) specify--run-numberthat will be used for all the generated fcl files--events-per-job--njobs
- Jobs with
RootInputsource (art files) require --inputsa file containing a list of all input data full filespecs--merge-factorhow many input files should be analyzed by a single job
The number of generated fcl files will be determined from the above inputs.
The following parameters are used to construct the names of fcl
files produced by the generate_fcl invocation.
Please see file names for guidelines on how to use these fields.
--descriptionthe "description" field of fcl file names--dsconfthe "configuration" field of fcl file names--dsownerthe "owner" field of fcl file names. This parameter defaults to the user who executesgenerate_fcl, but can be overridden. For example,--dsowner=mu2eshould be used to generate an official production dataset of fcl files.
One set of fcl files can be used to produce a series of different output datasets by changing the version of Offline used with the fcl. Conceptually then, the description and dsconf should be relevant to the fcl files themselves, and not the output dataset you have in mind today. For example, the same fcl file generating protons from stopped muons can be run with the DS magnet on or off, producing two very different datasets. So the fcl should reflect "protons from stops" not "magnet on/off study".
The --old-seeds parameter can be used
for incremental generation of fcl datasets. For example, one can
generate a test batch of 1000 fcl files and run them through the
grid. If the result is satisfactory, and one wants to increase the
statistics to 10,000 jobs, care should be taken to guarantee the
uniqueness of random seeds across all the 10,000 jobs. Each run
of generate_fcl produces a text file that contains the
values of all random seeds used so far for the current set of jobs.
So when generate_fcl is used the second time to add
9,000 jobs to the dataset, one should use the file with 1000 seeds
from the first run for the --old-seeds parameter to
make sure those seeds are not re-used.
(Also, --first-subrun should be adjusted so that subrun
numbers do not repeat.) The second run will dump a list of 10,000
seeds, which can be used in a subsequent generation if a further
increase in statistics is desired. For the initial run you can
specify --old-seeds=/dev/null.
Note: If you use the auxinput switch of generate_fcl in order to provide auxiliary input files (such as mixing files or stopped muon files) to your job and declare your fcl files to SAM, then the files used in the auxinput switch should already be declared to SAM. The auxinput files will be considered as parents (or precursors) to your fcl file, so the fcl file SAM record should record that fact. For standard mixing files, the upload and declare is already done for you. If you use personal files, then they should be uploaded and declared before using them in the fcl. For a personal, temporary job, you also can chose to not declare your fcl files, which will avoid this issue.
Run generate_fcl --help to see all the options.
Examples
Create a working dir. The data disk is a good place to work since we want fast response and some moderate space.
mkdir -p /mu2e/data/users/$USER/projects/my_project/fcl/job cd /mu2e/data/users/$USER/projects/my_project/fcl/job
Example 1 - generator
A first stage simulation job with no input files.
Prepare a template file. Usually it can be a single line file with just an include directive, in our case
#include "JobConfig/cd3/pions/pions_g4s1.fcl"
but one can also add e.g. geometry file overrides, or even write a
completely new fcl configuration and use it as a template.
In this example
we use template.fcl as the template file name.
Note that the include directive should specify include file pathname
relative to the Offline directory that you setup earlier. (More precisely, relative to
a directory listed in the FHICL_FILE_PATH.) Absolute filenames
do not work in fhicl #include.
Now generate the files:
generate_fcl --description=my-test-s1 \
--dsconf=v0 \
--run=2700 \
--events=1000 \
--njobs=5 \
template.fcl
After the command completes, we will see something like
> ls 000 template.fcl seeds.gandr.my-test-s1.v0.Td6j.txt > ls 000 cnf.gandr.my-test-s1.v0.002700_00000000.fcl cnf.gandr.my-test-s1.v0.002700_00000002.fcl.json cnf.gandr.my-test-s1.v0.002700_00000000.fcl.json cnf.gandr.my-test-s1.v0.002700_00000003.fcl cnf.gandr.my-test-s1.v0.002700_00000001.fcl cnf.gandr.my-test-s1.v0.002700_00000003.fcl.json cnf.gandr.my-test-s1.v0.002700_00000001.fcl.json cnf.gandr.my-test-s1.v0.002700_00000004.fcl cnf.gandr.my-test-s1.v0.002700_00000002.fcl cnf.gandr.my-test-s1.v0.002700_00000004.fcl.json
The generated fcl files and their corresponding json files are written into subdirectories 000, 001, etc, with up to 1000 fcl files per subdirectory. Random number seeds used for all the fcl files have been dumped into the "seeds" file.
Example 2 - secondary stage or concatenation
This is an example for running a later stage of a simulation job, where the the input file comes from an earlier stage. Concatenation can be thought of as a stage of simulation, just a particularly a simple one.
Prepare a template file. Usually it can be a single line file with just an include directive, in our case
#include "JobConfig/cd3/beam/beam_g4s2.fcl"
but one can also add e.g. geometry file overrides, or even write a
completely new fcl configuration and use it as a template.
In this example
we use template.fcl as the template file name.
Note that the include directive should specify include file pathname
relative to the Offline directory that you setup earlier. (More precisely, relative to
a directory listed in the FHICL_FILE_PATH.) Absolute filenames
do not work in fhicl #include.
If the jobs is simply concatenation, the template file will be, with a choice of output file name:
#include "JobConfig/cd3/common/artcat.fcl" outputs.out.fileName: "sim.DSOWNER.cd3-beam-cs3-mothers.DSCONF.SEQ.art"
Please look inside this fcl for directions on how to deal with the output file name. (Since the description is not known when the fcl file is written, you have to specify it in the fcl file.)
In all cases, you will also need a list of input files. If the list comes from a previously-uploaded dataset, it will be in SAM and in tape-backed dCache, so we can use SAM to generate the file list:
mu2eDatasetFileList sim.mu2e.cd3-beam-g4s1-dsregion.0506a.art > inputs.txt
If the input dataset is on disk in dCache because it is the output of a previous stage that just completed, you can make the input file list from that area. If you ran mu2eCheckAndMove, it still be under the "good" subdirectory.
cd /pnfs/mu2e/persistent/users/mu2epro/workflow/beam_g4s1_g4v10_p03_validation_rc1/good mu2eDatasetFileList sim.mu2e.cd3-beam-g4s1-dsregion.0506a.art > \ /mu2e/data/users/mu2epro/projects/geant_val_p03_s2/fcl/job/inputs.txt
Now go to the fcl working area and generate the files. It should be part of your job plan
to figure out how many of the files from the earlier stage should be used in this stage,
which is the --merge factor.
generate_fcl --desc=my-project-s2 \
--dsconf=v1 \
--inputs=inputs.txt \
--merge=30 \
template.fcl
After the command completes, we will see something like
> ls 000 inputs.txt seeds.mu2e.my-project-s2.v1.ZqcU.txt template.fcl > ls 000 cnf.mu2e.my-project-s2.v1.001600_00089044.fcl cnf.mu2e.my-project-s2.v1.001600_00089044.fcl.json cnf.mu2e.my-project-s2.v1.001600_00089046.fcl cnf.mu2e.my-project-s2.v1.001600_00089046.fcl.json ...
The generated fcl files and their corresponding json files are written into subdirectories 000, 001, etc, with up to 1000 fcl files per subdirectory. Random number seeds used for all the fcl files have been dumped into the "seeds" file.
Example 3 - mixing
A digitization+reconstruction job on a conversion electron file, with background mixing.
Prepare a template file. We want to use
JobConfig/cd3/beam/dra_mix_baseline.fcl, but in many
Offline releases this file is, strictly speaking, not a valid fcl
because it references a variable bgHitFiles that is not
defined. (A legacy of the mu2eart way of running grid
jobs.) To fix that, we define the variable before including the
"baseline" file. This example template.fcl also
shows how to set names of output histogram files.
BEGIN_PROLOG bgHitFiles: @nil END_PROLOG #include "JobConfig/cd3/beam/dra_mix_baseline.fcl" services.TFileService.fileName: "nts.owner.my-ce-reco.ver.seq.root"
We also need a list of input files, and a list of background overlay files. We want to run 5 jobs with one input file per job, so we need to shorten the conversion list to have only 5 input files. We only need one background hits file per job, but there is no harm of listing more background overlay files than necessary, so we will use a complete detmix-cut dataset:
setup mu2efiletools mu2eDatasetFileList sim.mu2e.cd3-beam-g4s4-detconversion.v566.art | head -n 5 > inputs.txt mu2eDatasetFileList sim.mu2e.cd3-detmix-cut.v566b.art > backgrounds.txt
We are ready to generate the fcl dataset. Note the '@' symbol in the --aux parameter - it says that bgHitFiles should be devined in a PROLOG, as the included file expects instead of being appended at the end. The leading number is how many background files each job will need.
generate_fcl --desc=my-reco-test \
--dsconf=v0 \
--inputs=inputs.txt \
--merge=1 \
--aux=1:@bgHitFiles:backgrounds.txt \
template.fcl
Take a look:
> ls 000 backgrounds.txt template.fcl inputs.txt seeds.gandr.my-reco-test.v0.CRdP.txt > ls 000 cnf.gandr.my-reco-test.v0.004001_00000000.fcl cnf.gandr.my-reco-test.v0.004001_00000002.fcl.json cnf.gandr.my-reco-test.v0.004001_00000000.fcl.json cnf.gandr.my-reco-test.v0.004001_00000003.fcl cnf.gandr.my-reco-test.v0.004001_00000001.fcl cnf.gandr.my-reco-test.v0.004001_00000003.fcl.json cnf.gandr.my-reco-test.v0.004001_00000001.fcl.json cnf.gandr.my-reco-test.v0.004001_00000004.fcl cnf.gandr.my-reco-test.v0.004001_00000002.fcl cnf.gandr.my-reco-test.v0.004001_00000004.fcl.json
Test fcl
It is highly recommended to test a newly generated fcl file by running a small interactive job. Following up on Example 1 above, one can do
mkdir test cd test /usr/bin/time mu2e -c `ls ../000/*.fcl | head -1`
to run a full size job, or add a -n 10 option to the mu2e command line
to quickly make sure that there are no obvious problems with the configuration.
Save fcl
In this step you will put the fcl in its final position, ready to use. You will have a choice of where to keep the fcl.
- Option 1 - upload to persistent dCache and declare to SAM. Use this for collaboration sponsored jobs, that will be uploaded to tape and will be used for a long time in the future. If any of the descendants of the fcl will be declared to SAM, or uploaded to tape, then you will need to use this option. If you want to use the mu2eMissingJobs tool for job recovery, it also requires this option. If you choose this option and this is the first time you are writing your personal fcl files to persistent dCache, you will need to ask for your personal directory to be created, please send mail to mu2eDataAdmin.
- Option 2 - move fcl files to scratch dCache. Use this for jobs which are temporary or personal, not to be uploaded to tape.
- Option 3 - make a tarball of fcl files, move that to scratch dCache, also for jobs which are not to be uploaded. Keeping the files as one tarball reduces your exposure to dCache rate and reliability issues.
You will also make a list of the fcl files that will be used to submit jobs. The result is several files named fcllist.??, each containing the names of a subset of fcl files. Each fclist file is used to submit a cluster of jobs. For production, we typically use 10K jobs in each submission, but for smaller personal projects, you might this to be smaller. You can also split off a smaller set for tests.
Option 1 -Move and SAM declare
First create the list of fcl to submit:
(for dir in ???; do cd $dir; ls *.fcl; cd ..; done) | mu2eabsname_disk | split -l 10000 -d - fcllist.
This should take a few minutes per 100K files.
A common recovery situation... If for some reason the files were already moved and declared before you made the list, you can make the list with mu2eDatasetList which reads the SAM records. If you moved then, but not declared them, you can make the list from the json files:
(for dir in ???; do cd $dir; ls *.fcl.json | sed 's/\.json//'; cd ..; done) | mu2eabsname_disk | split -l 10000 -d - fcllist.
In the normal flow, move them to the persistent dCache area:
(for dir in ???; do ls $dir/*.fcl; done) | mu2eFileUpload --disk >& upload.log &
This should run at about 100K to a few 100K files per day. If you can't write to the output area in persistent dCache, see the note above about creating your personal directory.
(for dir in ???; do ls $dir/*.fcl.json; done) | mu2eFileDeclare >& declare.log &
This should run at about 100K files per day. If you see errors while declaring files, check that you have a valid certificate.
Option 2 - Move fcl files to scratch
First create the list of fcl to submit:
(for dir in ???; do cd $dir; ls *.fcl; cd ..; done) | mu2eabsname_scratch | split -l 10000 -d - fcllist.
This should take a few minutes per 100K files.
Then move them to the scratch dCache disk, in the default area:
(for dir in ???; do ls $dir/*.fcl; done) | mu2eFileUpload --scratch >& upload.log &
Option 3 - Tar fcl files, move tarballs to scratch
Create a tarball for each set of files to be used in a submission. Typically a submission is 10K fcl files or fewer. This command will put each set of 10K fcl files in a tarball.
ls -1 -d ??? | cut -c 1-2 | sort | uniq | while read NN; do echo $NN; tar -cjf fcllist_${NN}.bz2 ${NN}?; done
Then move them to the scratch dCache disk, in your area:
cp fcllist* /pnfs/mu2e/scratch/users/$USER
When submitting each cluster of jobs, you can point to one of these fcl file tarballs.
Return to workflow
At the end of this procedure, you should have a set of files, each containing a list of fcl files.