collection of hands on exercises
Download the Intel MPI Benchmark from here: http://software.intel.com/en-us/articles/intel-mpi-benchmarks
-bash-3.2$ wget http://software.intel.com/sites/default/files/article/157859/imb-3.2.4-updated.tgz
Extract it:
-bash-3.2$ tar -xvf imb-3.2.4-updated.tgz
Edit file imb/3.2.4/src/make_ict: change compiler from from mpiicc to cc
Compile MPI1: you should get an executable named IMB-MPI1*
-bash-3.2$ make MPI1
Copy executable to work directory
cp ./IMB-MPI1 $WORKDIR
Ready to run! Let's start an interactive job:
qsub -I -A TRN001 -l nodes=1,walltime=00:30:00
Now we send the executable to a compute node
aprun ./IMB-MPI1
Alternatively try 2 mpi processes:
aprun -n2 ./IMB-MPI1
Now it's your turn to play. Here's some simple commands you'll need to know:
module list
module avail
module del
module swap PrgEnv-pgi PrgEnv-gnu
Here's your HW:
[ ]Run MPI, IO benchmarks
[ ]Try 1,2,4,8,16 mpi processes on up to 2 nodes.
-
Variations include same numa node, defferent, odd numbered
-
see https://www.olcf.ornl.gov/support/system-user-guides/titan-user-guide/#340
[ ]Play with some flags found here: https://www.olcf.ornl.gov/kb_articles/using-the-aprun-command/
[ ]Find out what the hardware looks like
If you have any questions, see our guide here https://www.olcf.ornl.gov/support/system-user-guides/titan-user-guide/