Hi,
Just for later reference, it turns out I have only to add thing like
this in the submission script:
###SBATCH --ntasks-per-socket=4
export OMP_NUM_THREADS=${SLURM_CPUS_PER_TASK}
export OMP_PROC_BIND="spread" ## make cores in order inside a task
srun --cpu-bind=v,cores -m block:block ./c
Hi all,
I have a cpu allocation problem on the 2 socket 18-core machine,
with CPUSpecList=17,35 reserved.
Ideally, if I submit a job with two 17-core tasks in a node, it allocates
each task on a single NUMA node.
# *srun -N1 -n2 --cpus-per-task=**17** numactl -s *physcpubind: *0 1 2 3
