I'm using AWS to run gmx simulations. The compute node does not have access to the root. I am guessing gmx does not have access to openmpi as well, but I added a full path to the script to use the version installed on the shared folder, which is accessible with compute node.
Please note the openMPI versions are different on the root and shared folder.
Well, I can't be of much help, but this is how I compiled gmx_mpi on ubuntu to use gpu's. Of course, this compiling requires that I use gmx_mpi instead of gmx. You can compile it so only mdrun uses mpi.
tar -zxf gromacs-version.tgz
cd gromacs-folder
mkdir build-gromacs cd build-gromacs
CC=mpicc CXX=mpicxx cmake .. -DGMX_GPU=ON -DCUDA_TOOLKIT_ROOT_DIR=/usr/local/cuda -DGPU_DEPLOYMENT_KIT_ROOT_DIR=/usr/share/nvidia_gdk/usr/include/nvidia/gdk/ -DGMX_MPI=ON -DGMX_BUILD_OWN_FFTW=ON -DGMX_PREFER_STATIC_LIBS=ON -DCMAKE_INSTALL_PREFIX=/usr/share/ -DNVML_LIBRARY=/usr/share/nvidia_gdk/usr/src/gdk/nvml/lib/libnvidia-ml.so
make
sudo make install
If you're having trouble with the path to openmpi, then I imagine adding a flag like this when compiling might help, but it sounds like you might've already done that.
-DCMAKE_CXX_LINK_FLAGS="-Wl,-rpath,/path/to/gcc/lib64 -L/path/to/gcc/lib64
- Badges
- Science topic