MPI Problem

User Support
1

Hi, I have weird problem running mpi.

When I use main computer only, mpi works perfectly.
However, when using two or more nodes, it works independently. I’m not sure what the problem is.

Thank you.

2

I’m using OpenMPI 1.6.4 (2013 version), by the way.

2013년 7월 15일 월요일 오후 8시 28분 15초 UTC+9, Kyungkwan Noh 님의 말:

3

Let’s say your MPI is installed in /opt/openmpi/1.6. You would need to do the following to get it to compile and run correctly:

  1. Change the MPI variable in the Makefile to yes
  2. Change the MPI_DIR variable in the Makefile to /opt/openmpi/1.6
  3. Execute ‘make’ to compile
  4. To run, execute ‘/opt/openmpi/1.6/bin/mpiexec -n NPROCS openmc’ where NPROCS is the number of processors you want

If you have problems compiling, try running ‘make distclean’ and then re-compiling. If you get the following error when compiling:

…/xml-fortran/xmlreader: error while loading shared libraries: libmpi_f90.so.1: cannot open shared object file: No such file or directory

Then add /opt/openmpi/1.6/lib to the LD_LIBRARY_PATH environment variable.

Best regards,
Paul

4

FYI, I have had the same problem with OpenMPI. I had to switch to MPICH. If you like, I can try to recreate the problem, and give you more details.

5

It should work fine if everything is set up properly. I just confirmed that I am able to compile/run successfully using OpenMPI on my own machine. Typically you will see the multiple processes running independently if you are using the wrong mpiexec/mpirun executable. A few more questions to narrow this down (for either of you):

  • What OS are you on?
  • Did you install OpenMPI from source, or through a package manager? (or a binary if on Windows?)
  • If you installed from source, what compiler did you use to compile OpenMPI?
  • If you installed from source, what did you put for --with-mpi-f90-size?
  • What version (or commit) of OpenMC are you trying to compile/run?
6
  1. I’m using Linux Red hat
  2. I guess from source.
  3. I used intel 13.1.1
  4. I put small for --with-mpi-f90-size
  5. I think the newest version since I installed through git.

Well, if I use the command ‘/opt/openmpi/1.6/bin/mpiexec -n NPROCS openmc’, it operates only on main computer, not on different nodes.
So, I try to use mpiexec -hostfile [hostfile name] openmc.
The host file may look like this:
node01 slots=8
node02 slots=8

I guess my settings are right, but it still doesn’t work.

Thank you.

2013년 7월 15일 월요일 오후 8시 28분 15초 UTC+9, Kyungkwan Noh 님의 말: