Submitting DaVinci jobs to GRID

Rationale

The main limitation for local DaVinci docker is: Raw data files (.dst) need to be downloaded locally. Given that the size of .dst files is on the order of TBs, this method is only used for developing DaVinci scripts and first-order validation

Note

It looks like our ganga jobs now only work on lxplus7, so make sure to log on to that.

On lxplus7, several official DaVinci versions are provided. However, there are some drawbacks:
1. .dst files still need to be downloaded to some lxplus7 user directory
2. While DaVinci is running, the connection to lxplus7 must be kept alive

LHCb collaboration provides a solution: Submitting and running DaVinci jobs on a GRID. The advantages are:

GRID know how to access .dst files directly, thus there's no need to manually download them. Instead, users need to specify the links (LFNs) to desired .dst files
While the GRID DaVinci jobs are running, there's no need to keep a connection to lxplus.

The rest of this doc is divided in two parts:

GRID job preparation and submission on lxplus
Offline slimming and annexing of the produced ntuples on a local machine, most possibly on glacier.

GRID job preparation and submission on `lxplus7`

Setup LHCb GRID certificate

Following this twiki to:

Apply for a GRID certificate
Setup the certificate on lxplus

Note

The twiki claimed that a new user must physically go to CERN's user office to be able to apply for a new cert via ca.cern.ch. But I didn't have to do that.

Compile a local `DaVinci` on `lxplus7`

We are using some non-official TupleTool, so we need to compile DaVinci on lxplus7.

First, we need to figure out the runtime dependency for our version of DaVinci: v45r6.

lb-run -L DaVinci/v45r6

For v45r6, we pick the following runtime: x86_64-centos7-gcc9-opt.

We should set that runtime as the default for lxplus7. Add to your login shell config:

export CMTCONFIG=x86_64-centos7-gcc9-opt
export BINARY_TAG=$CMTCONFIG

Now, run this script anywhere on lxplus7 to build a DaVinci with our customizations. The DaVinci will be available at $HOME/build/DaVinciDev_v45r6.

Configure `ganga`

ganga is a command-line interface for LHCb GRID. We need to configure ganga job output directory. On lxplus7:

Run ganga once. This should create a .gangarc in $HOME.
Locate gangadir option, point it to your larger AFS storage, for example:
```
gangadir = /afs/cern.ch/user/s/suny/work/gangadir
```
Go into your gangadir, create a workspace symblink pointing to somewhere in your EOS. For example:
```
workspace -> /eos/home-s/suny/gangadir-workspace
```
Copy ganga/ganga.sample.py to $HOME/.ganga.py on lxplus7.

Submitting a job with `ganga`

Submitting jobs is very simple, you will simply copy a previous script that has similar features (eg. data, MC tracker-only), modify it appropriately, and run it. The detailed steps are

Log on to lxplus7 and get a proxy with lhcb-proxy-init
Clone lhcb-ntuples-gen and go to the appropriate jobs folder, for instance lhcb-ntuples-gen/run2-rdx/jobs/
Copy a script with similar characteristics and rename it with the current date and some description, for instance
```
cp 22_02_03-tracker_only_Bs.sh 22_02_03-tracker_only_ddx_22to25.sh
```
Modify the new script appropriately, for instance adding the MC IDs that you want to run. The script automatically sends the MagDown and MagUp for each sample.
Run the script, eg.
```
./22_02_03-tracker_only_ddx_22to25.sh
```

If there are no errors in the submission, commit the script to git

Note

An error such as

GangaDiracError: All the files are only available on
archive SEs. It is likely the data set has been archived. Contact data
management to request that it be staged (consider --debug option for
more information)

simply means that the files that you are trying to run on do not exist, perhaps because you have a typo.

Monitor the status of your jobs by entering interactive ganga (simply type ganga in lxplus after getting a proxy) and typing jobs

Usage of ganga_jobs.py

See this appendix.

Manage ganga job output

You need to keep ganga running in a lxplus session to get up-to-date info about your jobs and download output of completed (sub)jobs.

Update subjob status and force status to be "failed" when necessary

Sometimes ganga would stuck at updating job status. To reset the status for "completing" and "failed" subjobs, do:

jobs[63].backend.reset(True)

If that still doesn't bring a job to a stable state (i.e. "finished" or "failed"), force the job to fail:

jobs[63].force_status("failed", force=True)

Handling failing subjobs

The GRID job are split into subjobs, enabling parallel execution. Some subjobs may fail. Considering the following jobs output in ganga shell:

[01:30:22]
Ganga In [2]: jobs
Ganga Out [2]:
Registry Slice: jobs (30 objects)
--------------
    fqid |    status |      name | subjobs |    application |        backend |                             backend.actualCE |                       comment |  subjob status
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
      62 | completed |Dst_D0--20 |      26 |      GaudiExec |          Dirac |                                         None |                               |       0/0/0/26
      63 |    failed |Dst_D0--20 |    1007 |      GaudiExec |          Dirac |                                         None |                               |     0/1/0/1006
      66 |    failed |Dst_D0--20 |       1 |      GaudiExec |          Dirac |                                         None |                               |        0/1/0/0
      73 | completed |Dst_D0--20 |     772 |      GaudiExec |          Dirac |                                         None |                               |      0/0/0/772

Jobs 63 and 66 are marked as failed because all of their subjobs are either completed or failed.

To resubmit failed subjobs for, say, jobs[63]:

jobs[63].resubmit()

However, above won't work unless all sub-jobs are either completed or failed. To resubmit the failed sub-jobs ASAP:

jobs[63].subjobs.select(status="failed").resubmit()

If you forced a "failed" status, some sub-jobs may be in "killed" state. A simple job[63].resubmit() won't resubmit these killed jobs. To resubmit them:

jobs[63].subjobs.select(status="killed").resubmit()

There's a helper function to print failed subjobs:

show_subjobs(63)
show_subjobs(63, status='running')  # show subjobs of 'running' status

If for some jobs keep failing, consider the backend bad and ban it in:

ban_site_for_job helper function with interactive ganga:

ban_site_for_job(63, 'LCG.site.a')
ban_site_for_job(63, ['LCG.site.a', 'LCG.site.b'])  # you can ban multiple sites at once

Then resubmit the job using methods described above

ganga/ganga_jobs.py in this repo: In BannedSites
$HOME/.ganga.py on lxplus: In remake_uncompleted_job function parameter

Sometimes it's needed to remake new jobs (say we want to change number of files per subjob). To do so:

remake_uncompleted_job(63)

Info

The remake_uncompleted_job creates a new job for each failed subjob, and add the subjob index to the comment property of the new job.

After the remade job has finished successfuly, merge all of its output .root files and place it in the correct directory of the failing job.

Ntuples can be merged with:

hadd -fk <name>.root */output/*.root

Slimming and annexing of GRID ntuples

Offline slimming

Before you proceed

Clone lhcb-ntuples-gen on glacier and setup git-annex normally.

Also don't forget to do make install-dep in the nix shell!

We prefer to not merge .root files at all. Still, we need to give output files sane names and slim them to remove unneeded branches.

Yipeng has prepared a ganga function that generates a bash script to aid the process. To use it:

Before you proceed

The whole procedure takes a long time. It's recommeded to do it inside a tmux session.

Use scp to copy finished jobs to a folder <glacierntuples> under in your $HOME on glacier

Example

Say your gangadir is at ~/eos/gangadir-workspace/suny/LocalXML, and you want to proceed job index 63, then copy the folder 63 to glacier.
In ganga shell, type in hadd_completed_job_output(63), where 63 is some job index.

Info

This will generated a batch_hadd.sh in $HOME. The generated script contains commands to merge output .root files for all completed jobs with a index that is greater or equal to the specified index.

A script named batch_skim.sh will be generated in $HOME on lxplus.

A sample batch_skim.sh is displayed here.
Copy batch_skim.sh to your lhcb-ntuples-gen root folder on glacier, then give it execution permission with chmod +x batch_skim.sh.

Set the input folder and postprocessing rules as needed in batch_skim.sh:

INPUT_DIR=<glacierntuples>
SKIM_CONFIG=./postprocess/skims/rdx_mc.yml  # NOTE: Make sure you pick the right one!!!

Info

If you don't want to remove any branches, then replace:

./ganga/ganga_skim_job_output.py ${OUTPUT_DIR}/$2 ${INPUT_DIR}/$1 ${SKIM_CONFIG}

with

./ganga/ganga_skim_job_output.py ${OUTPUT_DIR}/$2 ${INPUT_DIR}/$1 ${SKIM_CONFIG} --copy

Noting the use of --copy flag.

Go into a nix shell in your lhcb-ntuples-gen with nix develop, then in the project root, run:
```
./batch_skim.sh ntuples/<folder_to_ntuple_output>
```
For example, <folder_to_ntuple_output> can be:
```
0.9.6-2016-production/Dst_D0-mc-tracker-only
```

Annex ntuples

Info

We decide to use a pull-request-based workflow for ntuple annexation to minimize errors.

Create a new branch on your lhcb-ntuples-gen project on glacier, then checkout that branch:

git branch <branch_name>
git checkout <branch_name>

Example

git branch yipeng-ddx_part1
git checkout yipeng-ddx_part1

Go to the folder that contains your newly slimmed ntuples, then do git annex add . and commit changes with git commit . -m <comment>.
Example
```
cd ntuples/0.9.6-2016-production/Dst_D0-mc-tracker-only
git annex add .
git commit . -m "Part 1 of DDX MC"
```
Push this branch to Github with git push origin <branch_name>, then create a pull-request.

Info

If the git push origin <branch_name> failed, check your git history and make sure you didn't added ntuples (large files) directly with git add.

Github will refuse files larger than 100 MB.

Once everything checks out, one of Manuel or Yipeng will merge the request.

Once the request is merged, do the following:

git checkout master
git pull origin master

verify that your annexed ntuples are there, then do the following in the lhcb-ntuples-gen project root:

git annex sync
git annex copy . --to glacier
git annex sync  # YES, you need to do it twice, once before copying, and once after!

Once you finish all these, inform Manuel or Yipeng.

Info

For more info on git-annex usage, review the git-annex entry.

Appendix

Sample `batch_skim.sh`

#!/usr/bin/env bash
INPUT_DIR=~/ntuples  # NOTE: Configure this before proceed!!!
SKIM_CONFIG=./postprocess/skims/rdx_mc.yml  # NOTE: Make sure you pick the right one!!!
OUTPUT_DIR=$1
MIN_NTUPLE_SIZE=500  # in KiB

RED='[0;31m'
NC='[0m' # No Color


function check_job () {
  local error=0
  local job_dir=${INPUT_DIR}/$1
  local num_of_subjobs=$2

  echo "Verifying output for Job $1, which has $2 subjobs..."
  # might require GNU find
  local folders=$(find $job_dir -mindepth 1 -maxdepth 1 -type d -regex ".*/[0-9]*" | wc -l)

  if [[ $folders -ne $num_of_subjobs ]]; then
      echo -e "${RED}Found $folders subjobs, which =/= $num_of_subjobs. Terminate now.${NC}"
      exit 1
  fi

  for sj in $(ls $job_dir | grep -E "^[0-9].*$"); do
    local file=$(find $job_dir/$sj/output -name '*.root')

    if [[ -z $file ]]; then
      let "error++"
      echo -e "${RED}subjob $sj: ntuple missing!${NC}"
    else
      local size=$(du -b $file | awk '{print int($1 / 1024)}')  # in KiB
      if [ $size -lt ${MIN_NTUPLE_SIZE} ]; then
        let "error++"
        echo -e "${RED}subjob $sj: ntuple has a size of $size KiB, which is too small!${NC}"
      fi
    fi
  done

  if [ $error -gt 0 ]; then
    echo -e "${RED}Job $1 output verification failed with $error error(s).${NC}"
    exit $error  # exit early to make errors easy to see
  fi

  return $error
}

function concat_job () {
  check_job $1 $2

  if [ $? -eq 0 ]; then
    ./ganga/ganga_skim_job_output.py ${OUTPUT_DIR}/$3 ${INPUT_DIR}/$1 ${SKIM_CONFIG}
  fi
}

concat_job 180 111 Dst_D0--22_02_07--mc--tracker_only--MC_2016_Beam6500GeV-2016-MagDown-TrackerOnly-Nu1.6-25ns-Pythia8_Sim09k_Reco16_Filtered_11894600_D0TAUNU.SAFESTRIPTRIG.DST.root

General usage of `ganga_jobs.py`

For this repo, there is a central ganga job submitter that should handle all job submissions for all analyses with all reconstruction scripts in all folders. The script is located at: ganga/ganga_jobs.py. For it

The general syntax is:

ganga_jobs.py <reco_script> <cond_files>

For instance, for run 1 $R(D^{*})$ signal Monte Carlo:

ganga_jobs.py run1-rdx/reco_Dst_D0.py run1-rdx/cond/cond-mc-2012-md-sim08a.py -p mu -P Pythia6 -d 11574020

Note

The -p and -P are optional. They have default values.
The -d flag has a dummy default of 00000000. For actual values, consult data sources.

Info

The usage of ganga_jobs.py is described by:

ganga_jobs.py --help

Most of submissions are wrapped in shell scripts. Some of them can be found here.

Submitting DaVinci jobs to GRID

Rationale

GRID job preparation and submission on lxplus7

Setup LHCb GRID certificate

Compile a local DaVinci on lxplus7

Configure ganga

Submitting a job with ganga

Update subjob status and force status to be "failed" when necessary

Handling failing subjobs

Slimming and annexing of GRID ntuples

Offline slimming

Annex ntuples

Appendix

Sample batch_skim.sh

General usage of ganga_jobs.py

GRID job preparation and submission on `lxplus7`

Compile a local `DaVinci` on `lxplus7`

Configure `ganga`

Submitting a job with `ganga`

Sample `batch_skim.sh`

General usage of `ganga_jobs.py`