Skimming

3.4.12. Skimming#

What is skimming?#

Skims are sets of selections made on data and MC with particular analyses in mind. The purpose of skims is to produce data and MC files that have been reduced from their original size. This is done by applying a list of criteria to the data and MC, such that only events that interested a given analyst will be stored and provided. The analyst can then use the skimmed samples to further fine tune and improve their research. Skimmed samples are usually around 90% smaller than the original data and MC samples they are produced from. These samples are thus more manageable to use for analysis development and reduce the overall CPU and storage usage requirements of each analyst. Belle II is expecting to collect 50 ab-1 of data, which will be almost impossible to run on without skimming.

The criteria for skims varies from analysis to analysis. The general gist is to use a loose selection which can then be optimized by the analyst. For example, an analyst looking for the decay of a \(B\to D \ell \nu\) and \(D^0 \to K^- \pi^+\) will want to examine events where there are at least 3 tracks: two for the \(D\) daughter tracks and one for the lepton. The corresponding skim can include such a criteria where only events with more 3 tracks or more are included. The skim will also include a loose selection for the reconstruction of a \(B\) meson. Tighter selection criteria related to the lepton or D reconstruction are usually not applied at skim level. The analyst applied their optimized selection on the skimmed samples.

Skims are intended into reduce the overall CPU requirements of the collaboration, and to make your life easier. Skims can make your life easier in the following ways:

  • Skimmed files are generally less than 10% the size of the original (unskimmed) files, so your steering file will not need to process as many events, and your jobs will finish quicker.

  • The particles reconstructed during a skim are available when you load in the skimmed uDST, so you can use these in further reconstruction. For example, there are skims which use the FEI, so this computationally expensive reconstruction is performed during the skimming step and does not need to be repeated in later reconstruction.

Mechanics of a skim#

Under the hood, skims operate by running over the mDST, which is the format of the produced Monte Carlo samples and the processed data at Belle II, reconstructing a specific particle list, and writing the reconstructed particle information to a microDST, referred to as uDST. The skim filter removes any events which do not satisfy the criteria of the given skim, and thus do not have any candidates in the skim particle lists. For example, for the decay of \(B\to D \ell \nu\) and \(D^0 \to K^- \pi^+\), all events with less than 3 tracks are not included. Furthermore, in the skim itself, \(B\) meson is reconstructed using very loose criteria on the lepton and \(D\) daughters. Events that do not have a \(B\) candidate satisfying the loose criteria defined by the skim will not be included.

Question

uDST files contain the skimmed particle list information in addition to all the information contained in the mDST. Why is the file size of a uDST skim still smaller than the original size of the mDST?

Solution

Even though we are adding more information to each event by saving the reconstructed particle lists, only a fraction of events are kept by the skim, so the overall file sizes are reduced.

List of available skims#

At Belle II, we already have a list of skims developed by different analysts and skim liaisons. All available skims are listed on Sphinx (although not all of these are produced in skim campaigns). Although we try to keep the docstrings for each skim up-to-date, the best way to find out what selections are in a skim is to read the source code. The most important part of a skim’s source code is the build_lists method, where particles are reconstructed and selections are applied.

Exercise

Find the source code for the electroweak penguin (EWP) skims by navigating the Sphinx documentation.

Solution

Click the [Source] button on any of the skims in the EWP section to be taken to source code for that skim.

Running a skim locally#

As mentioned above, there is a list of developed skims available in the skim package. An analyst starting a new project is strongly encouraged to browse through the list of available skims and find out if there is a skim that meets their needs. Available skims are ready to run on any data and MC sample.

There are two ways to run a skim yourself: including the skim in a steering file, or using the command-line tool b2skim-run.

Including a skim in a steering file#

Skims in the skim package are defined via the BaseSkim class. To add all the required modules for a skim to your steering file, simply run:

from skim.leptonic import LeptonicUntagged
skim = LeptonicUntagged()
skim(path)  # add required skim modules to path

Running the above code will add modules to the path to load so-called standard particle lists, reconstruct the skim particle lists, and write the particle list to an output uDST file. If you would like to disable the uDST output, you can do so via:

skim = LeptonicUntagged(udstOutput=False)
skim(path)

Once the skim modules have been added to the path, you can retrieve a Python list of particle lists:

>>> skim.SkimLists
["B-:LeptonicUntagged_0", "B-:LeptonicUntagged_1"]

You can then use this list of particle list names in further reconstruction or ntuple output.

Using b2skim-run#

The command b2skim-run is a simple tool for applying a skim to a sample.

b2skim-run single SkimName -i MyDataFilename.mdst.root

By default the output filename will simply be the corresponding skim code (more on this in the next part of the lesson), but this can be controlled with the -o flag.

The full documentation of this tool can be found on Sphinx, or by using the -h flag.

Exercise

Use b2skim-run to apply the skim XToD0_D0ToHpJm to the file $BELLE2_VALIDATION_DATA_DIR/mdst16.root.

Solution

The command to run the XToD0_D0ToHpJm skim on this sample is:

b2skim-run single XToD0_D0ToHpJm -i $BELLE2_VALIDATION_DATA_DIR/mdst16.root

By default, this will output a uDST file in the current directory titled 17230100.udst.root.

Exercise

What is the retention rate (fraction of events passing the skim) of the XToD0_D0ToHpJm skim on this sample?

Hint

You can use the tool b2file-metadata-show to print the number of events in an mDST or uDST file.

Solution

b2file-metadata-show $BELLE2_VALIDATION_DATA_DIR/mdst16.root
b2file-metadata-show 17230100.udst.root

We find the unskimmed file has 90000 events, and the skimmed file has 3519 events, so the retention rate on this sample is 3.9%.

Accessing skims on the grid#

Analysts do not have to run the skims themselves on data or generic MC. Each new MC campaign or data collection, a list of skims is requested by the analysts in the Belle II physics working groups. This is done via the skim liaison or via GitLab issues. Once requested, the skim is run on the large MC and/or data samples by the skim production managers. These skims are then announced when ready and made available to the analyst.

Each skim campaign on data or MC samples has a given name. For example, skims of MC13a run-independent MC are listed under the campaign name SkimM13ax1. Skims of data are usually made available for official processing, like Proc11, or for individual buckets like bucket9, bucket10, etc. The corresponding skim campaign names are SkimP11x1 and SkimB9x1-SkimB13x1. The production status of available MC and data samples is continuously updated on the Data Production Status page. Status updates on the readiness of a skim campaign are also posted on the Skim Production Status XWiki page.

To find the list of skim campaign campaigns available on the, simply browse through the app, select Data type: MC or Data and look in the drop-down menu under Campaigns. All skim campaigns start with the not so mysterious name “Skim”.

Skimmed samples are produced and stored on the grid. The output LFNs are documented on the dataset searcher. You can then run your analysis these centrally-produced skims with gbasf2. LFNs on the grid have a maximum length restriction, so we can’t include the plain skim name in the LFN. Instead, we have standardised eight-digit skim codes to identify skims. When searching for skimmed datasets on the grid, use the skim codes. The documentation of each skim on Sphinx contains its corresponding skim code, and the full table of codes can be found in the documentation of skim.registry.SkimRegistryClass.

Note

The details of the numbering scheme are explained on the skimming Expert XWiki page.

Exercise

Use the dataset searcher to get the list of LPNs for the B0toDpi_Kspi skim from the MC skim campaign SkimM13ax1.

Hint

Find the skim code for B0toDpi_Kspi on the skim documentation on Sphinx.

Solution

Visit the DIRAC webapp and navigate to the dataset searcher. The LFNs can be found by selecting MC and BGx1, and passing SkimM13ax1 in the Campaigns field, and 14120601 in the Skim Types field.

Tip

All skims on the grid are produced using some release of the software. If you’re unsure which version was used to produce your skim, check the LFN, as it is recorded in there! You can then directly read source code for that release to find the skim definitions.

Exercise

Run the analysis script in B2A303-MultipleDecays-Reconstruction.py on one of the LPNs for the mixed samples of the B0toDpi_Kspi skim from the MC skim campaign SkimM13ax1 on the grid.

Skimmed data samples are made available in directories on the grid, where each directory corresponds to a given run. This results in an inconvenient number of directories the user has to run on, however, this preserves the run information of a given skim, as inherited from data production.

Warning

Currently the dataset searcher does not list all available directories for a given skim production job. It only lists one directory. In reality, there are usually ~100 directories per production. This is a known bug and will be improved in future developments of the dataset searcher.

For now, a workaround in order is described on the Skim XWiki page to run your analysis script on the full set of skimmed data samples available for a given campaign.

Getting involved#

Each working group has an assigned data production liaison (all listed on XWiki), whose job it is to survey the needs of the group and develop skims. If there is an existing skim that might be useful for your analysis and is not currently being produced, talk to your local skim liaison.

If you would like to get more involved in the writing and testing of skims, then you may find the skim experts section of the Sphinx documentation helpful.

Key points

  • The two sources of documentation on skims are the Sphinx documentation and the skimming XWiki page. The best way to find out how a particular skim is currently defined is to read the source code (either on Sphinx, or in the directory skim/scripts/skim/ in the software repo).

  • You can run a skim by adding a short segment of code to your steering file, or by using the command-line tool b2skim-run.

  • Centrally-produced skims can be accessed on the grid with gbasf2. Use the dataset searcher to locate skimmed data by using the relevant skim code.

  • Running on skimmed data and MC can make your life as an analyst easier. However, skims are only useful if they are developed through communication between analysts and skim liaisons, so don’t hesitate to contact your working group’s liaison.

Stuck? We can help!

If you get stuck or have any questions to the online book material, the #starterkit-workshop channel in our chat is full of nice people who will provide fast help.

Refer to Collaborative Tools. for other places to get help if you have specific or detailed questions about your own analysis.

Improving things!

If you know how to do it, we recommend you to report bugs and other requests with GitLab. Make sure to use the documentation-training label of the basf2 project.

If you just want to give us feedback, please open a GitLab issue and add the label online_book to it.

Please make sure to be as precise as possible to make it easier for us to fix things! So for example:

  • typos (where?)

  • missing bits of information (what?)

  • bugs (what did you do? what goes wrong?)

  • too hard exercises (which one?)

  • etc.

If you are familiar with git and want to create your first merge request for the software, take a look at How to contribute. We’d be happy to have you on the team!

Quick feedback!

Do you want to leave us some feedback? Open a GitLab issue and add the label online_book to it.

Authors of this lesson

Phil Grace, Racha Cheaib

Contents