As of June 2022, users are no longer recommended to transfer large output files using stashcp and instead should use HTCondor's transfer_output_remaps feature. More information on this transition can be found below. Please direct any questions about this change to support@osg-htc.org

Overview

For input files >100MB and output files >1GB in size, the default HTCondor file transfer mechanisms run the risk of over-taxing the login nodes and their network capacity. And this is exactly why the OSG Data Federation exists for researchers with larger per-job data!

Users on an OSG Connect login node can handle such files via the OSG Connect data caching origin (mounted and visible as the /public location) and use OSG's caching tools to scalably transfer them between the running jobs and the origin. The OSG caching tools ensure faster delivery to and from execute nodes by taking adantage of regional data caches in the OSG Data Federation, while preserving login node performance.

Important Considerations and Best Practices

  1. As described in OSG Connect's Introduction to Data Management on OSG Connect, the /public location must be used for:

    • Any input data or software larger than 100MB for transfer to jobs using OSG caching tools
    • Any per-job output >1GB and <10GB, which should ONLY be transferred back to the origin by setting transfer_output_remaps in the job's submit file.
  2. User must never submit jobs from the /public location, and should continue to ONLY submit jobs from within their /home directory. All log, error, output files and any other files smaller than the above values should ONLY ever exist within the user's /home directory, unless otherwise directed by an OSG staff member.

    Thus, files within the /public location should only be referenced within the submit file by using the methods described further below.

    The /public location is a mount of the OSG Connect origin filesystem. It is mounted to the OSG Connect login nodes only so that users can appropriately stage large job inputs or retrieve outputs via the login nodes.

  3. Because of impacts to the filesystem of the data origin, files in the data origin (/public) should be organized in one or very few files per job. The filesystem is likely to encounter performance issues if/when files accumulated there are highly numerous and/or small.

  4. The /public location is otherwise unnecessary for smaller files (which can and should be served via the user's /home directory and regular HTCondor file transfer). Smaller files should only be handled via /public with explicit instruction from an OSG staff member.

  5. Files placed within a user's /public directory are publicly accessible, discoverable and readable by anyone, via the web. Data is made public via stash transfer (and, thus, via http addresses), and mirrored to a shared data repository which is available on a large number of systems around the world.

Use a 'stash' URL to Transfer Large Input Files to Jobs

Jobs submitted from the OSG Connect login nodes will transfer data from the origin
when files are indicated with an appropriate stash:/// URL in the transfer_input_files line of the submit file:

  1. Upload your larger input and/or software files to your /public directory which is accessible via your OSG Connect login node at /public/username for which our Using scp To Transfer Files To OSG Connect guide may be helpful.

    Because of the way your files in /public are cached across the Open Science Pool, any changes or modifications that you make after placing a file in /public will not be propagated. This means that if you add a new version of a file to your /public directory, it must first be given a unique name (or directory path) to distinguish it from previous versions of that file. Adding a date or version number to directories or file names is strongly encouraged to manage your files in /public.

  2. Add the necessary details to your HTCondor submit file to tell HTCondor which files to transfer, and that your jobs must run on executes nodes that have access to the Open Science Data Federation.

    # Submit file example of large input/software transfer
    
    log = my_job.$(Cluster).$(Process).log
    error = my_job.$(Cluster).$(Process).err
    output = my_job.$(Cluster).$(Process).out
    
    #Transfer input files
    transfer_input_files = stash:///osgconnect/public/<username>/<dir>/<filename>, <other files>
    
    ...other submit file details...
    

    Note how the /public mount (visible on the login node) corresponds to the /osgconnect/public namespace across the Open Science Federation. For example, if the data file is located at /public/<username>/samples/sample01.dat, then the stash:/// URL to transfer this file to the job's working directory on the execute point would be:

    stash:///osgconnect/public/<username>/samples/sample01.dat
    

Use transfer_output_remaps to Transfer Larger Job Outputs to the Data Origin

For output, users should use the transfer_output_remaps option within their job's submit file, which will transfer the user's specified file to the specific location in the data origin.

By using transfer_output_remaps, it is possible to specify what path to save a file to and what name to save it under. Using this approach, it is possible to save files back to specific locations in /public (as well as your /home directory, if desired).

The syntax for transfer_output_remaps is:

transfer_output_remaps = "Output.txt = path/to/save/file/under/output.txt; Output.txt = path/to/save/file/under/RenamedOutput.txt"

When saving large output files back to /public, the path provided will look like:

transfer_output_remaps = "Output.txt = stash:///osgconnect/public/<username>/Output.txt"

  1. Using transfer_output_remaps, tell HTCondor which output files need to be transferred back to your /public directory and what name you want these files to be saved under.

    # submit file example for large output
    
    log = my_job.$(Cluster).$(Process).log
    error = my_job.$(Cluster).$(Process).err
    output = my_job.$(Cluster).$(Process).out
    
    requirements = (OSGVO_OS_STRING =?= "RHEL 7")
    
    transfer_output_remaps = "Output.txt = stash:///osgconnect/public/<username>/Output.txt"
    
    ...other submit file details...
    
  2. If you have several output files being sent to /public, you may wish to define a new submit file variable to avoid having to re-write the stash:/// path repeatedly. For example,

    # submit file example for large output
    
    log = my_job.$(Cluster).$(Process).log
    error = my_job.$(Cluster).$(Process).err
    output = my_job.$(Cluster).$(Process).out
    
    requirements = (OSGVO_OS_STRING =?= "RHEL 7")
    
    STASH_LOCATION = stash:///osgconnect/public/<username>
    transfer_output_remaps = "file1.txt = $(STASH_LOCATION)/file1.txt; file2.txt = $(STASH_LOCATION)/file2.txt; file3.txt = $(STASH_LOCATION)/file3.txt"
    
    ...other submit file details...
    

Phase out of Stashcp command

Historically, output files could be transferred from a job to a /public location using the stashcp command within the job's executable, however, this mechanism is no longer encouraged for OSPool users. Instead, jobs should use transfer_output_remaps (an HTCondor feature) to transfer output files to /public. By using transfer_output_remaps, HTCondor will manage the output data transfer for your jobs. Data transferred via HTCondor is more likely to be transferred successfully and errors with transfer are more likely to be reported to the user.

Get Help

For assistance or questions, please email the OSG Research Facilitation team at support@osg-htc.org or visit the help desk and community forums.

 

This page was updated on Jul 20, 2022 at 21:24 from start/data/stashcache.md.