Running the analysis#

CLI interface#

Once installed in your software environemnt, either by using an apptainer image or a custom python environment, PocketCoffea exposes different scripts and utilities with a command-line-interface (CLI)

$> pocket-coffea 

    ____             __        __  ______      ________
   / __ \____  _____/ /_____  / /_/ ____/___  / __/ __/__  ____ _
  / /_/ / __ \/ ___/ //_/ _ \/ __/ /   / __ \/ /_/ /_/ _ \/ __ `/
 / ____/ /_/ / /__/ ,< /  __/ /_/ /___/ /_/ / __/ __/  __/ /_/ /
/_/    \____/\___/_/|_|\___/\__/\____/\____/_/ /_/  \___/\__,_/


Running PocketCoffea version 0.8.0
- Documentation page:  https://pocketcoffea.readthedocs.io/
- Repository:          https://github.com/PocketCoffea/PocketCoffea

Run with --help option for the list of available commands

The commands and their options can be explored directly with the CLI:

$> pocket-coffea  --help
Usage: pocket-coffea [OPTIONS] COMMAND [ARGS]...

Options:
  -v, --version BOOLEAN  Print PocketCoffea package version
  --help                 Show this message and exit.

Commands:
  build-datasets      Build dataset fileset in json format
  hadd-skimmed-files  Regroup skimmed datasets by joining different files...
  make-plots          Plot histograms produced by PocketCoffea processors
  run                 Run an analysis on NanoAOD files using PocketCoffea...

The run command is the one used to execute the analysis workflow on the datasets. This is replacing the previous pocket-coffea run script, but it has the same user interface.

Executors#

The PocketCoffea analysis can be runned in different ways, locally or sending out jobs to a cluster throught the Dask scheduling system.

  • local iterative processing The processor works on one chunk at a time. Useful for debugging and local testing.

  • local multiprocessing The processor works on the chunks with multiple processes in parallel. Fast way to analyze small datasets locally.

  • Dask scale-up Scale the processor to hundreds of workers in HTCondor through the Dask scheduler. Automatic handling of jobs submissions and results aggregation.

Assuming that PocketCoffea is installed (for example inside the singularity machine), to run the analysis just use the pocket-coffea run command:

$> pocket-coffea run --help

    ____             __        __  ______      ________
   / __ \____  _____/ /_____  / /_/ ____/___  / __/ __/__  ____ _
  / /_/ / __ \/ ___/ //_/ _ \/ __/ /   / __ \/ /_/ /_/ _ \/ __ `/
 / ____/ /_/ / /__/ ,< /  __/ /_/ /___/ /_/ / __/ __/  __/ /_/ /
/_/    \____/\___/_/|_|\___/\__/\____/\____/_/ /_/  \___/\__,_/


Usage: pocket-coffea run [OPTIONS]

  Run an analysis on NanoAOD files using PocketCoffea processors

Options:
  --cfg TEXT                      Config file with parameters specific to the
                                  current run  [required]
  -ro, --custom-run-options TEXT  User provided run options .yaml file
  -o, --outputdir TEXT            Output folder  [required]
  -t, --test                      Run with limit 1 interactively
  -lf, --limit-files INTEGER      Limit number of files
  -lc, --limit-chunks INTEGER     Limit number of chunks
  -e, --executor TEXT             Overwrite executor from config (to be used
                                  only with the --test options)
  -s, --scaleout INTEGER          Overwrite scalout config
  -c, --chunksize INTEGER         Overwrite chunksize config
  -q, --queue TEXT                Overwrite queue config
  -ll, --loglevel TEXT            Console logging level
  -f, --full                      Process all datasets at the same time
  --executor-custom-setup TEXT    Python module to be loaded as custom
                                  executor setup
  --help                          Show this message and exit.

To run with a predefined executor just pass the --executor option string when calling the the run command

$> pocket-coffea run --cfg analysis_config.py -o output --executor dask@lxplus

Have a look below for more details about the available executor setups.

Executors availability#

The iterative and futures executors are available everywhere as they run locally (single thread and multi-processing respectively).

Site Supported executor Executor string
lxplus dask dask@lxplus
T3_CH_PSI dask dask@T3_CH_PSI
T2_RWTH_Aachen parsl parsl-condor@RWTH
Purdue Analysis Facility dask dask@purdue-af

Executors setup#

The analysis processors is handled by executors. The setup of the executor can vary between sites. A set of predefined executors has been prepared and configured with default options for tested analysis facilties (lxplus, T3_CH_PSI). More sites are being included, please send us a PR when you have successfully run PocketCoffea at your facility!

The preconfigured executors are located in the pocket_coffea/executors module. The default options for the running options and different type of executors are stores in pocket_coffea/parameters/executor_options_defaults.yaml.

For example:

general: 
  scalout: 1
  chunksize: 100000
  limit-files: null
  limit-chunks: null
  retries: 20
  tree-reduction: 20
  skip-bad-files: false
  voms-proxy: null

dask@lxplus:
  scaleout: 10
  cores-per-worker: 1
  mem-per-worker: "2GB"
  disk-per-worker: "2GB"
  worker-image: /cvmfs/unpacked.cern.ch/gitlab-registry.cern.ch/cms-analysis/general/pocketcoffea:lxplus-cc7-stable
  death-timeout: "3600"
  queue: "microcentury"
  adaptive: false
  performance-report: true
  custom-setup-commands: null
  conda-env: false
  local-virtualenv: false

Executor options#

The dataset splitting (chunksize), the number of workers, and the other options, which may be executor-specific, must be configured by the user passing to pocket-coffea run a .yaml file containing options. These options are overwritten over the default options for the requested executor.

For example:

$> cat my_run_options.yaml

scaleout: 400
chunksize: 50000
queue: "espresso"
mem-per-worker: 6GB


$> pocket-coffea run  --cfg analysis_config.py -o output --executor dask@lxplus  --custom-run-options my_run_options.yaml

The user can also modify on the fly some run options using arguments of the pocket-coffea run script. For example by limiting the number of files or number of chunks to analyse (for testing purposes)

$> pocket-coffea run  --cfg analysis_config.py -o output --executor dask@lxplus \
              --limit-files 10  --limit-chunks 10 \
              --chunksize 150000 --queue espresso

Customize the executor software environment#

The software environment where the executor runs the analysis is defined by the python environment where the analysis is launched but also by the executor options.

In particular if the user is using a virtual environment or conda to develop the core PocketCoffea code inside a singularity image, there is an option to make the remote executors pickup the correct python env.

Just specify local-virtualenv: true in the custom run options for virtualenv inside the singularity or conda-env: true for using the conda (or mamba/micromamba) env activated where the pocket-coffea run script is run.

Local environment support

The local environment propagation to the remote executors has been implemented at the moment for lxplus and some other sites. It is dependent of the presence of a shared filesystem to propagate the environment to the workers and activate it before executing the dask worker jobs.

Moreover the user can add a list of completely custom setup commands that are run inside a worker job before executing the analysis processor. Just specify them in the run options:

$> cat my_run_options.yaml

custom-setup-commands:
  - echo $HOME
  - source /etc/profile.d/conda.sh
  - export CUSTOM_VARIABLE=1

Dask scheduler on lxplus#

The dask scheduler started by the pocket-coffea run script needs to stay alive in the user interactive session. This means that if you start a runner process directly in the lxplus machine (in a singularity session) you cannot logout from the session.

The solution is using the tmux program to keep your analysis session in the background. tmux allows you to create a session, detach from it, exit lxplus, and at the next login reattch to the running session.

This service needs to be activate, only once, for your user with systemctl --user enable --now tmux.service. The full documentation about this (new) feature is available on the Service Portal.

Once setup you can start a tmux session as:

tmux new -s your-session-name
# start an apptainer image and launch your analysis

# press `Ctrl+b d` to detach from the session

your running session are visible with tmux ls. To reconnect do tmux a -t your-session-name. Look here for more info about tmux.

Easy debugging#

The easiest way to debug a new processor is to run locally on a single process. The run command has the --test options which enables the iterative processor independently from the running configuration specified in the configuration file. The processor is run on a file of each input dataset.

$ pocket-coffea run --cfg config.py --test

If you want to run locally with multiple processes for a fixed number of chunks just use the options:

$ pocket-coffea run --cfg config.py --test -e futures -s 4 --limit-files 10 --limit-chunks 10

Adding support for a new executor/site#

If you want to run PocketCoffea in a analysis environment that is still not centrally implemented you can implement a custom ExecutorFactory and pass it to the pocket-coffea run script on the fly. In practice, this means that the user is free to define from scratch the configuration of its cluster for running with Dask for example.

Have a look at pocket_coffea/executors/executors_lxplus.py for a full-fledged example.

The user factory must implement a class deriving from ExecutorFactoryABC. The factory returns an instance of a Executor that is passed to the coffea Runner.

 1## custom executor defined in my_custom_executor.py by the user
 2
 3from pocket_coffea.executors.executors_base import ExecutorFactoryABD
 4from .executors_base import IterativeExecutorFactory, FuturesExecutorFactory
 5
 6from coffea import processor as coffea_processor
 7
 8class ExecutorFactoryCustom(ExecutorFactorABC):
 9
10    def get(self):
11        return coffea_processor.dask_executor(**self.customized_args())
12
13    def setup(self):
14        '''This function is called by the base class constructor'''
15        # do all the needed setup
16        self.start_custom_dask_cluster()
17        
18    def start_custom_dask_cluster(self):
19        self.dask_cluster = .......... #custom configuration
20
21    def customized_args(self):
22        '''This function customized the args that coffea uses to instantiate 
23        the executor class passed by the get() method'''
24        args = super().customized_args()
25        args["custom-arg"] = "..."
26        return args
27
28    def close(self):
29        # cleanup
30        self.dask_cluster.close()
31
32
33def get_executor_factory(executor_name, **kwargs):
34    if executor_name == "iterative":
35        return IterativeExecutorFactory(**kwargs)
36    elif executor_name == "futures":
37        return FuturesExecutorFactory(**kwargs)
38    elif  executor_name == "dask":
39        return ExecutorFactoryCustom(**kwargs)
40

The user’s module must implement a get_executor_factory(string, run_options) method which returns the instantiated Executor.

The module is then used like this:

$> pocket-coffea run --cfg analysis_config.py -o output --executor dask  --executor-custom-setup my_custom_executor.py
--run-options my_run_options.py

Tip

When the setup is working fine we would highly appreciate a PR to add the executor to the list of centrally supported sites with default options!