http://bugzilla.mcs.anl.gov/swift/show_bug.cgi?id=1409 Mihael Hategan <[email protected]> changed: What |Removed |Added ---------------------------------------------------------------------------- AssignedTo|[email protected] |[email protected] --- Comment #1 from Mihael Hategan <[email protected]> 2015-03-01 18:24:30 --- I committed code for the local scheduler providers that hopefully fixes various quoting issues in the generated scripts. This is trunk git-rev: 948f986c0f72ce699e2a457abdf912673d3be0e0 heads/master 6285 While the providers are able to run MPI jobs (using mpirun/aprun/srun, etc.), this does not mean that swift is. There are two issues. One is file staging. If you run a swift job through a provider and specify jobType = "MPI", the provider will happily run multiple copies of an app instance wrapped around _swiftwrap. Each instance of _swiftwrap will try to stage in and stage out files, which will result in multiple conflicting stage-ins and stage-outs. The second issue is that there is no guarantee that I can see that mpirun/aprun/srun will be OK with running _swiftwrap instead of an MPI executable. I believe that, in the long run, the right way to run MPI apps is for swift to quietly force jobType = "single", run _swiftwrap on the lead node (to do the staging) and then instructing _swiftwrap to run the app instances through one of mpirun/aprun/srun/ssh. In the mean-time, this can be achieved through an application wrapper as follows: use jobType = "single" count = <number of nodes needed> then, in the application wrapper, start the app through mpirun (or equivalent). I am re-assigning this bug to Yadu. It is critical that we test the providers on various clusters. -- Configure bugmail: http://bugzilla.mcs.anl.gov/swift/userprefs.cgi?tab=email ------- You are receiving this mail because: ------- You are watching all bug changes.