Hi, On 2012-11-19 20:42, Mario Kadastik wrote: > > Hi, > > we're setting up a test instance of slurm. So far I've gotten as far > as slurm control node and one worker node configured. I'm able to use > the slurm-torque interface commands to send a job that actually does > things on the workernode, but the output is kept at the workernode as > we do not use shared home directories. For some reason I cannot run > srun directly:
We are using Lustre on DDN storage and it works as a shared filesystem. I see mostly our local users are more happy with shared files system, because it is easier to use. > [mario@slurm-1 ~]$ srun /bin/hostname srun: error: Couldn't find the > specified plugin name for launch/slurm looking at all files srun: > error: cannot find launch plugin for launch/slurm srun: error: cannot > create launch context for launch/slurm Segmentation fault > [mario@slurm-1 ~]$ I'm running for some time SLURM in production at VU-MIF-LCG2. Most jobs are from local users. Recently installed ARC-CE and it works at least at ARC UI level. To run SLURM you have properly configure slurm.conf (and make it identical on WN and SLURM server). Also munge security should be setup (by distributing the same munge.key on WN and SLURM server). I'm still running stable version (2.4.3) of SLURM on Debian 6.0 so I cannot comment on packages layout. I made debian package directly from my own SLURM build. > the slurm I built is built from slurm-2.5.0-0.rc1.tar.bz2 using > rpmbuild -ta slurm-2.5.0-0.rc1.tar.bz2. So far on the headnode I > installed: [root@slurm-1 ~]# rpm -qa|grep slurm > slurm-plugins-2.5.0-0.rc1.el6.x86_64 > slurm-torque-2.5.0-0.rc1.el6.x86_64 > slurm-sjstat-2.5.0-0.rc1.el6.x86_64 slurm-2.5.0-0.rc1.el6.x86_64 > slurm-sjobexit-2.5.0-0.rc1.el6.x86_64 > slurm-munge-2.5.0-0.rc1.el6.x86_64 > slurm-perlapi-2.5.0-0.rc1.el6.x86_64 > > haven't set up accounting yet so those packages weren't installed. > > The srun peculiarity somewhat disturbs me, but as qrun worked i'm not > 100% sure what's going on. However I was curious if slurm supports > input / output stageing using scp or similar? Our current cluster is > based on torque, which uses scp to stage in input files and stage out > output files. Most of the tools used for submission by users as well > as through grid assume the PBS staging code so my interest is wether > this is possible to be set up as well? Should this be done through > prologue/epilogue scripts or is there some native way to do it? I didn't investigate such possibilities. They are not needed on shared file system layout, but in theory it is possible to run non shared setup, but in this case every script should have some prolog/epilog section to copy data to/from WN. SLURM probably doesn't provide that in Torque way, but I can be wrong. Regards, Rolandas Naujikas > Thanks, > > Mario Kadastik, PhD Researcher > > --- "Physics is like sex, sure it may have practical reasons, but > that's not why we do it" -- Richard P. Feynman >
