Hello All,

  I just issued a pull request that augments Galaxy to allow defining job runners dynamically at runtime (https://bitbucket.org/galaxy/galaxy-central/pull-request/12/dynamic-job-runners). Whether it makes the cut or not, I thought I would describe enhancements here in case anyone else would find it useful.

  There a couple use cases we hope this will help us address for our institution - one is dynamically switching queues based on user (we have a very nice shared memory resource that can only be used by researchers with NIH funding) and the other is inspecting input sizes to give more accurate max walltimes to pbs (a small number of cufflinks jobs for instance take over three days on our cluster but defining max walltimes in excess of that for all jobs could result in our queue sitting idle around our monthly downtimes). You might also imagine using this to dynamically switch queues entirely based on input sizes or parameters, or alter queue priorities based on the submitting user or input sizes/parameters.

  There are two steps to use this - you must add a line in universe.ini and define a function to compute the true job runner string in the new file lib/galaxy/jobs/rules.py.

  This first step is similar to what you would do to statically assign a tool to a particular job runner. If you would like to dynamically assign a job runner for cufflinks you would start by adding a line like one of the following to universe.ini

cufflinks = dynamic:///python
-or-
cufflinks = dynamic:///python/compute_runner

If you use the first form, a function called cufflinks must be defined in rules.py, adding the extra argument after python/  lets you specify a particular function by name (compute_runner in this example). This second option could let you assign job runners with the same function for multiple tools.

The only other step is to define a python function in rules.py that produces a string corresponding to a valid job runner such as "local:///" or "pbs:///queue/-l walltime=48:00:00/".

If the functions defined in this file take in arguments, these arguments should have names from the follow list: job_wrapper, user_email, app, job, tool, tool_id, job_id, user. The plumbing will map these arguments to the implied galaxy object. For instance, job_wrapper is the JobWrapper instance for the job that gets passed to the job runner, user_email is the user's email address or None, app is the main application configuration object used throughout the code base that can be used for instance to get values defined in universe.ini, job, tool, and user are model objects, and job_id and tool_id the relevant ids.

If you are writing a function that routes a certain list of users to a particular queue or increases their priority, you will probably only need to take in one argument - user_email. However, if you are going to look at input file sizes you may want to take in an argument called job and use the following piece of code to find the input size for input named "input1" in the tool xml.

    inp_data = dict( [ ( da.name, da.dataset ) for da in job.input_datasets ] )
    inp_data.update( [ ( da.name, da.dataset ) for da in job.input_library_datasets ] )
    input1_file = inp_data[ "input1" ].file_name
    input1_size = os.path.getsize( input1_file )

This whole concept works for a couple of small tests on my local machine, but there are certain aspects of the job runner code that makes me feel there may be corner cases I am not seeing where this approach may not work - so your millage may vary.

-John

------------------------------------------------
John Chilton
Software Developer
University of Minnesota Supercomputing Institute
Office: 612-625-0917
Cell: 612-226-9223
E-Mail: chilton@msi.umn.edu