Christoph Niethammer
2017-08-21 15:10:02 UTC
Hello
I try to place executables on different sockets on different nodes with Open MPI 2.1.1.
Therefore I use something like the following command:
mpirun --map-by ppr:1:node -np 1 numactl -N 0 /bin/hostname : -np 1 numactl -N 1 /bin/hostname
But from the output I see, that all processes are placed on the same node.
Trying the same with Intel MPI using the following command
export I_MPI_JOB_RESPECT_PROCESS_PLACEMENT=no
mpirun -ppn 1 -np 1 /bin/hostname : -np 1 /bin/hostname
processes go to different nodes as desired.
So I am wondering if "--map-by" is a "local option" according to the mpirun
man page and how I can achieve the desired behaviour with Open MPI.
Thanks for your help.
Best
Christoph Niethammer
I try to place executables on different sockets on different nodes with Open MPI 2.1.1.
Therefore I use something like the following command:
mpirun --map-by ppr:1:node -np 1 numactl -N 0 /bin/hostname : -np 1 numactl -N 1 /bin/hostname
But from the output I see, that all processes are placed on the same node.
Trying the same with Intel MPI using the following command
export I_MPI_JOB_RESPECT_PROCESS_PLACEMENT=no
mpirun -ppn 1 -np 1 /bin/hostname : -np 1 /bin/hostname
processes go to different nodes as desired.
So I am wondering if "--map-by" is a "local option" according to the mpirun
man page and how I can achieve the desired behaviour with Open MPI.
Thanks for your help.
Best
Christoph Niethammer