CUDA with OpenMPI on Ubuntu 8.04

Accelerated Computing CUDA CUDA Programming and Performance
1

This is a pretty random question, but here it goes…

I am attempting to integrate mpi code into my CUDA enabled program. The goal is to offload background processes to a separate box that I have (2x Tesla cards) rather than trying to fit more GPU cores into my current box (3x 280). The internal gigabit network is set up and I can execute simple openmpi programs (hello world -ish) across the network without any problems. When I try to execute my full CUDA utilizing program, I get the following error.

program: error while loading shared libraries: libcudart.so.2: cannot open shared object file: No such file or directory

The program itself can run on each machine without any problems (i.e. mpirun -np 1 program). I only run into trouble with (mpirun -np 2 -machinefile hostnames program) Any thoughts or suggestions?

My setup

Computer 1
3x 280

Computer 2
9500 + 2x Tesla

Both computers have:
quad core AMD
8GB RAM
Ubuntu 8.04

2

Undoubtedly environment. You will need to pass LD_LIBRARY_PATH settings to the process on the remote machine. The mpirun/mpiexec manpage should hold all the answers you need.

3

avidday… thanks for your help. For anyone who may have this problem in the future, here is the command line that runs everything flawlessly

mpirun -np 2 -machinefile hostfile -x LD_LIBRARY_PATH program_name