By default Slurm allocates node in exclusive mode. You have to use Consumable Resources to achieve what you want.
http://schedmd.com/slurmdocs/cons_res.html /David On Wed, Feb 6, 2013 at 10:47 AM, Niels Rothermel < [email protected]> wrote: > Hello, > > I think my problem is pretty easy to solve but I'm a total noob when it > comes to HPC. In our lab we have a single machine for doing some DFT > calculations. It's running with Ubuntu 12.10 and I installed SLURM via the > Ubuntu repositorys. The Installation was no problem, the deamons are > running, the example commands in the SLURM documentation all give the > expected results. > But when it comes to running the software (orca, which is fully > parallelized using openmpi 1.4.5 by itself) SLURM isn't doing what I want. > In principle I don't want to make any restrictions on maximum resources per > job and always want to allocate resources for the individual job. My first > attempt was to do the following: > > "srun orca input.file >& output.file &" > > The number of parallel processes for the job is specified in the > input.file by the line > > "%pal nprocs 8 end" > > if the number of processes shall be eight. If I start the job like this > orca is executed with eight parallel processes each using 100% CPU. When I > start a second job in the same way it won't be executed as the one before > is still running, but its pending in the squeue and waiting for resources. > How can that be? The machine has 64 Procs. > I asked the same question in the orca board and one user adviced me to run > it like this: > > "salloc -n8 orca input.file >& output.file &" > > with the -n8 of course beeing identical witht the number of parallel > processes in the input.file. > > I would be very grateful if somebody could help me on this issue. The > output of "scontrol show config" is attached to the email. > > Thanks a lot > Regards > Niels > >
