Hi,

I am just wondering if there are any srun flags to

allocate the last MPI-rank to an exclusive node, for example,


#SBATCH -N 3 # 3 nodes

#SBATCH --tasks-per-node=10

#SBATCH --cpu-per-task=1     # 10 MPI-rank per node


srun -n 21 ...   ./application    # MPI-rank 0-9 on node 0;  MPI-rank 10-19 on 
node 1; MPI-rank 20 on node 2 !?


For my case, the last MPI-rank is consider as master and it takes much memory. 
So it would be better if the last MPI-rank can be distributed on an exclusive 
node . I have tried all distribution patterns introduced in


https://doc.zih.tu-dresden.de/hpc-wiki/bin/view/Compendium/BindingAndDistributionOfTasks


but it seems to me that all patterns allocate MPI-rank equivalently into nodes.


Thanks.


Regards, Jing Gong

Reply via email to