EventLoop Batch System Drivers

Last update: 11 Dec 2019 [History] [Edit]

Using a (supported) batch system works essentially like using the other drivers, all you have to do is to create the proper driver:

    CondorDriver driver;
    TorqueDriver driver;
    LSFDriver driver;
    LLDriver driver; // load leveler
    GEDriver driver; // grid engine (this is separate from actual grid submission)

You will normally have to execute some commands on the worker nodes to set up root. You have to add those commands to the driver:

    driver.shellInit = "some commands for setting up root on the nodes";

If you need extra flags for submitting to your batch system you will also need to specify them (you can either do this through job.options or driver.options):

    job.options()->setString (EL::Job::optSubmitFlags, "-x -y -z");

For condor, you can specify extra lines to place into the condor configuration:

job.options()->setString (EL::Job::optCondorConf, "parameter =

The LSFDriver contains a special hack to make it work on inexplicably, that may cause problems at other sites. You can disable it via:

job.options()->setBool (EL::Job::optResetShell, false);

There are a couple of limitations at the moment. These limitations are not fundamental, but mostly exist because nobody entered a feature request to fix them. In particular most drivers assume that your build directory is on a shared filesystem available on all worker nodes. Furthermore I assume that the output directory is located in a directory that can directly be written to by all the worker nodes. On most clusters this can be achieved by placing them in the home directory. Note that the Condor driver has some special options to work around that.

If you find that you need a driver for a different batch system or need extensions to the driver for your batch system, let us know and we will try to help. Adding support for new batch systems turned out fairly easy and straightforward, we just added the support for torque in under 3 hours. Depending on what you ask for, we may need a login to your cluster so that I can test the code.

Changing the number of files per worker

Sometimes when you run on a large number of files, you don’t want to have a separate worker process running for each file. This is mostly to avoid overhead:

  • Each job started on the batch system has a certain overhead. If jobs are too short a large fraction of time is spend on starting and stopping jobs.
  • All the output histograms have to be merged on the submission node. The time this takes is directly proportional to the number of jobs submitted. If you have a many small jobs on a large batch system this step can take up more time then actually running in batch.
  • If you produce output n-tuples, each job will produce a separate output n-tuple file. Processing these output n-tuples afterwards can be inefficient if the individual n-tuple files themselves are very small. It should be noted that this can also be alleviated by merging the output afterwards.

Whether this affects you is hard to say in general, and there is a tradeoff in that if you have too many small jobs you may be unable to utilize your entire batch system. However, if you run a lot of short jobs and suspect that this affects you, you can try changing the number of files per job and see whether this improves things.

You can do this for the whole job by calling:

job.options()->setDouble (EL::Job::optFilesPerWorker, 5);

And if you decide that there is a sample that needs a different number of files per job, you can use:

sample->setMetaDouble (EL::Job::optFilesPerWorker, 10);

Splitting Jobs Based on the Number of Events

If you run = inexplicably = on your sample handler it will scan the number of events in each root file and store them in the sample handler.

SH::scanNEvents (sh);

If you then submit the sample handler to the batch driver it will split jobs so that all jobs have approximately equal number of events.

If you want, you can also configure your job to take a particular number of events per job:

sh.setMetaDouble (EL::Job::optEventsPerWorker, 5000);

This will cause the batch driver to spawn just enough jobs so that no job has more than 5000 events. If you want to, you can also set this separately for each sample:

sh.setMetaDouble ("sampleName", EL::Job::optEventsPerWorker, 10000);

This can be helpful if you have different processing speeds or acceptance rates for different samples.

Special Options for Special Locations

For running CondorDriver on NAF at several German institutes, the following option is reported to make it work:

sh.setMetaString (EL::Job::optCondorConf, "+MyProject = \"af-atlas\"\nshould_transfer_files = NO");

The should_transfer_files indicates that the user job files do not need to be copied over, but instead gets picked up via the shared file system.