Hi Bjørn-Helge, Can't you use directly srun?
Regards, Carles Fenoy On Tue, Mar 5, 2013 at 3:24 PM, Bjørn-Helge Mevik <b.h.me...@usit.uio.no>wrote: > > (This is on slurm 2.4.3.) > > We are trying to get a piece of software (that wants to play queue system) > to run on our cluster, and it needs a command to submit a job and block > until the job has finished. > > SGE's qsub command has an option "-sync y", which makes the qsub command > wait for the job to complete before it exits. PBS's qsub has > "-Wblock=true", which does the same. > > Is it possible to emulate this behaviour in SLURM, either with commands > or the Perl API (or even the C API, if need be)? > > The closest I've got with commands is > > salloc <job specific slurm options> /usr/bin/srun --ntasks=1 --nodes=1 > --preserve-env <jobscript> > > for instance: > > salloc --job-name=t1 --mem-per-cpu=500 --time=5 --account=staff > --nodes=2 --ntasks-per-node=2 /usr/bin/srun --ntasks=1 --nodes=1 > --preserve-env env.sm > > This makes sure that the commands in the jobscript are run on the job > allocation's batch node, and mpirun and multithreaded programs work > well. However, srun commands don't: unless you specify --ntasks less > than the number of tasks for the whole job, you get "srun: error: Unable > to create job step: Requested node configuration is not available". > > There is a simpler version that avoids the srun problem: > > salloc <job specific slurm options> <jobscript> > > However, this runs all commands in the script on the submit node, and we > cannot assume that all commands in the job script use srun or mpirun to > do the work. > > I'd rather not create a command that uses sbatch to submit the job, and > then polls the queue system until the job has finished. > > -- > Regards, > Bjørn-Helge Mevik, dr. scient, > Research Computing Services, University of Oslo -- -- Carles Fenoy
