On Apr 11, 2011, at 11:33 PM, kevin.buck...@ecs.vuw.ac.nz wrote: > >>> #!/bin/bash >>> #$ -cwd >>> #$ -j y >>> #$ -S /bin/bash >>> #$ -q all.q >>> #$ -pe orte 18 >>> MPI_DIR=/home/jason/openmpi-1.4.3-install/bin >>> /home/jason/openmpi-1.4.3-install/bin/mpirun -np $NSLOTS myprog > > >> If you have SGE integration, you should not specify the number >> of slots requested on the command-line. Open MPI will speak >> directly to SGE (or vice versa, to get this information. > > I've always thought that the tight integration merely removed the > need for a hostfile, or equivalent, to be specified?
That is correct. -np specifies the number of processes to be run, not the number of slots in the allocation. The latter is determined by the SGE integration. In this case, I believe the author's intent was that the two be the same, which is why I let it go without comment. However, if that was the intent, he could have left the "-np" argument off entirely - with no number of processes specified, mpirun will simply assign a process to each allocated slot. > > I've been basing this belief on the examples here: > > http://www.open-mpi.org/faq/?category=running#qsub-notify > > where, in all three examples, both the SGE submission script's > #$ directives and the mpirun commands invoked, specify the > number of processors but there's no hostfile/list. To be precise, the mpirun commands specify the number of -processes- to be run, not the number of -processors- to be used. Sounds technical, but one needs to remember that you can oversubscribe processes in MPI, so this distinction is critical. > > > -- > Kevin M. Buckley Room: CO327 > School of Engineering and Phone: +64 4 463 5971 > Computer Science > Victoria University of Wellington > New Zealand > > > > _______________________________________________ > users mailing list > us...@open-mpi.org > http://www.open-mpi.org/mailman/listinfo.cgi/users
