ipcluster about ipyparallel

[minrk]

I've had the best experience on clusters submitting a single job with a given node/cpu allocation, and then starting the engines with mpiexec ipengine in the job file.

[josh-gree]

So something along the lines of this?

#PBS -N ipython
#PBS -j oe
#PBS -l walltime=00:10:00
#PBS -l nodes=2:ppn=16

module load openmpi/1.6.4/intel

ipcontroller --log-to-file
mpiexec -np 31 ipengine --log-to-file

[minrk]

Just about, though I often put the Controller in a different job or on a head node, so I don't pass -np to mpiexec. That script as-is won't quite work to start the controller and engines both, because you'll need a condition for the controller, but otherwise 👍:

if [[ $rank = 0 ]]; then # whatever the PBS rank env variable is
 ipcontroller;
else
  mpiexec ...

[josh-gree]

Thanks very much for your help, I have it all up and running now!

[minrk]

@josh-gree thanks! I need to do some writeups for clusters, because often ipcluster is not the easiest tool for the job.