body p { margin-bottom: 0cm; margin-top: 0pt; }
Hi all,
I'm used to working with openmpi, and thus had to refer to the
documentation when I had to implement a solution based on Intel mpi.
In the slurm docs (
http://www.schedmd.com/slurmdocs/mpi_guide.html#intel_mpi ) 3
methods (not including mpd ) are discussed, in the Intel reference
docs (
http://software.intel.com/sites/products/documentation/hpc/mpi/linux/reference_manual.pdf
p. 23 s. 2.3.1.1) a JMI method is also mentioned (with the
possibility of setting the jmi library through
I_MPI_HYDRA_JMI_LIBRARY).
So my main question is - which method should I use inside an sbatch
script, and what are the differences?
The candidates being:
1. mpirun -bootstrap slurm -n <num_procs> a.out
2. export
I_MPI_HYDRA_JMI_LIBRARY=/opt/intel/impi/4.0.3/lib64/libjmi_slurm.so.1.0
mpirun -bootstrap jmi -n <num_procs> a.out
3. export I_MPI_PMI_LIBRARY=/usr/lib64/libpmi.so
srun -n <num_procs> a.out
I have not mentioned mpiexec.hydra, as I believe this is just a
pseudonym for mpirun in ver 4.0u3. Please correct me if I'm wrong.
I have two more questions, of lesser importance to me:
1. What is the difference between mpirun -bootstrap, and mpirun
-bootstrap-exec ? are they complementary? should I use one, or the
other, or both?
2. Should I have includded mpd in the candidates list?
Thanks in advance
