Can Jetson Orin support mpi?

[Big_mao]

Hi~

I want to use a jetson orin nx and jetson orin nano to build a distributed environment ,and I want to use mpi to communicate , so I install openmpi as ucx and openmpi official website :

configure as follow:

# configure environment 

export CUDA_HOME="/usr/local/cuda-12.2" && \
export UCX_HOME="/usr/local/ucx" && \
export OMPI_HOME="/usr/local/ompi" && \
export PATH="${CUDA_HOME}/bin:$PATH}" && \
export PATH="{UCX_HOME}/bin:$PATH}" && \
export PATH="{OMPI_HOME}/bin:$PATH}" && \
export LD_LIBRARY_PATH="${CUDA_HOME}/lib64:$LD_LIBRARY_PATH}" && \
export LD_LIBRARY_PATH="${UCX_HOME}/lib64:$LD_LIBRARY_PATH}" && \
export LD_LIBRARY_PATH="${OMPI_HOME}/lib64:$LD_LIBRARY_PATH}"


# install UCX
git clone https://github.com/openucx/ucx.git
cd ucx
git clean -xfd 
./autogen.sh

mkdir build && cd build
../configure --prefix=$UCX_HOME --enable-debug --with-cuda=$CUDA_HOME --enable-mt --disable-cma

make -j$(nproc)
sudo make install

# install openmpi
git clone --recursive https://github.com/open-mpi/ompi.git  
cd ompi
git clean -xfd 
./autogen.pl

mkdir build && cd build
./configure --prefix=$OMPI_HOME --with-cuda=$CUDA_HOME --with-ucx=$UCX_HOME --with-ucx-libdir=$UCX_HOME/lib 
make -j$(nproc)
sudo make install


# configure 
MPI_HOME=/usr/local/ompi && \
export PATH=${MPI_HOME}/bin:$PATH && \
export LD_LIBRARY_PATH=${MPI_HOME}/lib:$LD_LIBRARY_PATH && \
export MANPATH=${MPI_HOME}/share/man:$MANPATH

Then, I use two command to verify:

ompi_info | grep "MPI extensions"
ompi_info --parsable --all | grep mpi_built_with_cuda_support:value

I don’t know why , I build with cuda all the time .

Has anyone succeeded on jetson?

[kayccc]

Please refer to OpenMP in jetson nano - Jetson & Embedded Systems / Jetson Nano - NVIDIA Developer Forums to see if can help.

Thanks

[Big_mao]

Thanks for your quick reply~

I want to use CUDA AWARE MPI, but the solution probably not。

According to my search results, to realize CUDA AWARE MPI, I must build from the source code, but after I build from the source code,use vertify command:

ompi_info | grep "MPI extensions"
ompi_info --parsable --all | grep mpi_built_with_cuda_support:value

It shows mpi_built_with_cuda_support:value:false ; I don’t know why.

I followed the official documents and referred this link: Jetpack with cuda-aware OMPI could be default - Jetson & Embedded Systems / Jetson Xavier NX - NVIDIA Developer Forums

but it always failure

[AastaLLL]

Hi,

It looks like a duplicate issue of the below topic:

Does Jeston support distributed computing of cuda? How to achieve it?

I am a student ,and I have a Jetson Orion NX and a jetson orin nano, the configuration is as follows: Jetpack 6 CUDA 12.2 Pytorch 2.3.0 (from source) Pytorch _ Geometric 2.6.0 (from source) python 3.9 I want to use them to build a distributed environment， when I use CPU + GLOO ,it’s OK. Then, I want to change to GPU and NCCL, I met with failure，fortunately ，I find nccl not support jetson ,and I find that , jetson support mpi, so I have a try to build openmpi from source : export CUDA…

If so, recommended to use one topic so the information will be in the same place.

Thanks.

[Big_mao]

Yes, it’s the same question, I will I will turn off this topic and focus on that topic.

Thank you!