User Tools

Site Tools


clusters:hypatia:openmpi

OpenMPI on Hypatia

MPI tasks may be easier to handle in a Slurm setup. Use the BilbyFarm for that.

Caution, experimental, not extremely well tested as of now

(Instructions provided by Patrick Brem for Vulcan, thank you!)

NEW: You may use /usr/local/bin/openmpiscript which has been adapted, and will be maintained.

The submit file submit_mpi.sub should look something like this:

universe = parallel
executable = /usr/local/bin/openmpiscript
initialdir = /home/user/workingdir
arguments = /home/user/workingdir/executable arg1 arg2 etc
machine_count = 8
log = log
output = out.$(NODE)
error = err.$(NODE)
getenv = true
should_transfer_files = yes
when_to_transfer_output = on_exit
queue

(Patrick Brem:) So far I have had bad experiences using relative paths in the executable, so to be safe use absolute paths there, since the child processes find themselves in $HOME initially even though setting an explicit initialdir above. Feel free to experiment there.

(Steffen Grunewald:) It's good practice also for other reasons to use full paths, also (and especially) for initialdir. The arguments (of the openmpiscript executable) will be used to complete a mpirun invocation. The machine_count is the number of cores (like Datura, unlike Minerva) MPI nodes get mapped on. Not always will you get full machines. (Kyohei, I was wrong.)

If you have suggestions how to get 4 cores per machine please add them here!

  • Request 4 cores per machine, quadruple the resulting machinefile
condor_submit submit_mpi.sub

(last modified by Patrick Brem on 16 Jul 2014 / Steffen Grunewald on 17 Oct 2018)

clusters/hypatia/openmpi.txt · Last modified: 2021/08/03 12:17 by Steffen Grunewald