How to install Open MPI with Cuda-Aware MPI in Linux Debian

Question

I am installing Open MPI v1.8.8 with CUDA v7.5 on my Linux Debian. I have tested CUDA and it works, tested OpenMPI and it works too. But when i try to combine them into a program, i meet an error: cannot find cuda.h file . This is my scenario:

My program source code include these .h file

include "cuda.h" include "mpi.h"

I run command: mpicc <filePath> -o test And error appear: cuda.h: No such file or directory #include "cuda.h"`

omp_info give me : mca:mpi:base:param:mpi_built_with_cuda_support:value:false

I have googled , and i followed some methods i found: ./configure --with-cuda ./configure --with-cuda=/usr/local/cuda-7.5 ( source link : http://mirror.its.dal.ca/openmpi/faq/?category=buildcuda ) After that, i remake all , remake install Open Mpi. I run: mpicc or mpirun, the compiler give me error : mpirun error mca: base: component find: unable to open /usr/local/lib/openmpi/mca_mpool_sm

I set up soft link : ln -s /usr/local/cuda/include /usr/include ( describe in link : Building CUDA-aware openMPI on Ubuntu 12.04 cannot find cuda.h ). But it cannot fix my issue.

Does anyone successfully install it? Please help me or share your experience. Thanks.

Answer 1

I think you are confusing installation problems with incorrect compiler options. It will be necessary to explicity specify the include paths, library paths, and libraries for CUDA when compiling and linking host code with your mpi wrapped host compiler.

Something like:

mpicc -I/usr/local/cuda-7.5/include -L/usr/local/cuda-7.5/lib -o test <filePath> -lcuda

would be the normal way to build a simple MPI program which call the cuda driver APIs. You will need to add nvcc compilation for device code and host code which uses the runtime API.

The apparent lack of CUDA support in your MPI flavour is a separate question and one you should probably take up in another forum (like the user mailing list of the MPI flavour you use).