Support application kernels in the CU description

[mturilli]

Here a brief description of the use case that prompted this ticket. For an experiment I need to run the executable mdrun on four XSEDE resources. Several solutions come to mind:

1. User-intensive: preload/compile mdrun on each target resource; save the executable on $HOME/bin/mdrun; specify cu.executable = "$HOME/bin/mdrun for each cu - note that I have no idea on what pilot/resource the CU will run at the moment of specifying cu.executable.
2. User-less-intensive: stage mdrun for each CU (or shared across CU) before CU execution. This is what we do already with CU input files. The issue here is assuming that a 'medun' compiled on one XSEDE resource can be executed on each XSEDE resource available to me - likely an incorrect assumption.
3. User-not-intensive (?): use a 'kernel' abstraction similar to the one used for mdemtk (did I get it right?!). This still seems to leave open the issue of having to compile a mdrun for each resource, depending on architecture and shared library availability.

Please feel free to add to the discussion, indicating new solutions or expanding/clarifying/overriding the descriptions of the three solutions provided.

[andre-merzky]

This is now implemented in the 'experimental' branch. the mdutils need to be installed manually on client side for this to work (otherwise the UM will raise an exception).

[andre-merzky]

closing as it is implemented, as stated above.