Hello SLURM Users! I'm having a bit of trouble working with job arrays and I'm hoping you folks can give me a hand.
I have a node with 2 sockets and 16 cores per socket (no hyper threading). I would like to run 320 single threaded instances of an application confined to this node with 32 instances of the application active at any one time and bound to an individual core. Currently, my submission script is: #SBATCH --nodelist=broadwell01 #SBATCH --ntasks=1 #SBATCH --oversubscribe #SBATCH --array=0-320%32 srun ....... I would like to be able to run this on any node without specifying a name but If I replace the --nodelist directive with --nodes, the job spreads to all the available nodes. Right now I'm using: SelectType=select/linear NodeName=broadwell[01-32] Sockets=2 CoeresPerSocket=16 ThreadsPerCore=1 PartitionName=broadwell_compute Nodes=broadwell[01-32] Default=YES MaxTime=INFINITE State=UP OverSubscribe=YES:32 In an ideal world, I would like SLURM to default to allocating a whole node to a user and only allow over-subscription upon request. Any suggestions would be greatly appreciated! -- Neale Petrillo Knolls Atomic Power Laboratory
