cudart error when running the code

[hnakahara79]

Hi @ax3l ,

I build the code as: cmake -S . -B build -DWarpX_DIMS=3

The executable generated in the build/bin/ folder was warpx.3d.MPI.OMP.DP.PDP.OPMD.QED

I copied the executable in the PWD and run example file of the code to test

mpirun -np 80 ./warpx.3d.MPI.OMP.DP.PDP.OPMD.QED inputs_3d

I received the following error:

########
**No component were able to opened in the pml framework.**
**PML ucx cannot be selected**
**PMIX ERROR: UNREACHABLE in file server/pmix_server.c**
########

I used openmpi/4.1.4

########

During building of executable, the messages that appear, can be seen in the attachment file.

Thank you and Best regards.

[hnakahara79]

Hi @ax3l ,

do you have any suggestions how to fix the above problem.

[ax3l]

Hi @hnakahara79 , since your machine is GPU-enabled, use this to build WarPX for CUDA GPUs via -DWarpX_COMPUTE=CUDA:

rm -rf build

cmake -S . -B build -DWarpX_DIMS=3 -DWarpX_COMPUTE=CUDA
cmake --build build -j 8

The default build is for OpenMP (CPU only).

Changing the compute backend to GPU should solve the runtime error of WarpX as well as the linker warning you see for ADIOS2.

Ref.:
https://warpx.readthedocs.io/en/latest/install/cmake.html#build-options

[ax3l]

With regards to the HDF5 warning in your PDF: Did you load HDF5 via a module or did you build it yourself?

By default, HDF5 is non-parallel (has no MPI support). If you like to use HDF5 as an additional openPMD backend (you already have ADIOS2, which is totally sufficient as well), when you just need a HDF5 dependency that enabled MPI "parallel" support.
Maybe your system has another, MPI-parallel HDF5 module that you can load (or ask support to build one for your compilers & MPI)?

Otherwise this is how you build HDF5 from source: (load your compilers and MPI modules first!)

# install once
wget https://support.hdfgroup.org/ftp/HDF5/releases/hdf5-1.12/hdf5-1.12.2/src/hdf5-1.12.2.tar.gz
tar -xvf hdf5-1.12.2.tar.gz

cmake -S hdf5-1.12.2 -B hdf5-build -DHDF5_ENABLE_PARALLEL=ON -DCMAKE_INSTALL_PREFIX=$HOME/sw/hdf5-parallel-1.12.2
cmake --build hdf5-build -j 8
cmake --build hdf5-build --target install

then

# add permanently to your warpx.profile file
export CMAKE_PREFIX_PATH=$HOME/sw/hdf5-parallel-1.12.2:$CMAKE_PREFIX_PATH
export LD_LIBRARY_PATH=$HOME/sw/hdf5-parallel-1.12.2/lib:$HOME/sw/hdf5-parallel-1.12.2/lib64:$LD_LIBRARY_PATH

[ax3l]

With regards to your MPI error on startup:

########
**No component were able to opened in the pml framework.**
**PML ucx cannot be selected**
**PMIX ERROR: UNREACHABLE in file server/pmix_server.c**
########

this is an MPI component that was not properly configured/matched to your system.
You can disable the component at runtime (*) or ask your system support if you used mpirun/srun correctly, if they can disable the broken component in their MPI module, etc.

(*) The runtime flags for mpirun are -mca ... that we need to look up in the OpenMPI manual and there are equivalent environment variables that one can set. But ideally, system support should advise us here.