User documentation for the PLEIADES cluster at the University of Wuppertal
There are multiple MPI environments available:
module load 2021a GCC/10.3.0 OpenMPI/4.1.1
module load 2021a iimpi/2021a(oneAPI) or through Parallel Studio with
source /beegfs/Tools/intel/setup.sh(Version from 2020)
/lib64/openmpi3on our worker nodes
All MPI versions were tested on PLEIADES with an MPI benchmark. These tests covered the mpirun and srun approach (see below), as well as ethernet and infiniband communication.
Many problems with MPI are caused by a mismatch between the applications expected MPI version/configuration and the used MPI version in your environment. If you experience problems, try a clean build and investigate MPI related options during your application build and at runtime.
When compiling your own MPI, make sure to send a build job to the worker nodes and provide the
--with-ucx flag (and possibly more!).
Otherwise your MPI version is likely to not utilize our InfiniBand network and rely on ethernet communication instead.
The PMI library acts as an interface between Slurm and various MPI versions.
This way Slurm can manage MPI processes, if the MPI library has been compiled correctly, e.g. by using
If you intend to compile your own MPI versions, you may have to mention the location of PMI libraries:
# find /lib64/ -name libpmi* /lib64/libpmi.so.1.0.1 /lib64/libpmi.la /lib64/libpmi.so /lib64/libpmi2.la /lib64/libpmi2.so /lib64/libpmi.so.1 /lib64/libpmi2.so.1 /lib64/libpmi2.so.1.0.0 /lib64/libpmix.la /lib64/libpmix.so /lib64/libpmix.so.2.2.32 /lib64/libpmix.so.2
These paths may also become relevant if you use Intel MPI.
There are two approaches to use MPI in your Slurm batch scripts:
srun is by default executing
srun --mpi=pmix_v3, which may require your software to be build against PMIx, available as mentioned above.
Alternative PMI options can be listed with
Applications that implicitly ship MPI may need additional configuration, e.g. enabling slurm support or pointing to PMI libraries. This is a case-by-case situation where you should study the corresponding documentation of your application.
Consider reading the Slurm MPI documentation.
Deciding on the number of nodes, processes and cores per process can confusing sometimes.
srun --cpu-bind=help to show available options to bind CPU resources managed by Slurm to your (MPI-)processes and have a look at the Slurm CPU Management guide.
#!/bin/bash #SBATCH -p short #SBATCH -t 60 #SBATCH -N 4 # 4 Nodes #SBATCH -n 4 # 4 processes in total module load 2021a GCC/10.3.0 OpenMPI/4.1.1 # Option one: srun --mpi=pmix_v3 /path/to/mpiapplication <arguments> # Or using mpirun directly: mpirun /path/to/mpiapplication <arguments>
It is possible to pass
--mca options in these commands as well.
#!/bin/bash #SBATCH -p short #SBATCH -t 60 #SBATCH -N 4 # 4 Nodes #SBATCH -n 4 # 4 processes in total # Chose ONE of the following: # Use recent version of oneAPI impi module load 2021a iimpi/2021a # Alternatively use impi shipped with Parallel Studio 2020 source /beegfs/Tools/intel/setup.sh # Tell impi where to pick up the PMI library, paths above. # Maybe try libpmix.so or libpmi2.so, if you have problems export I_MPI_PMI_LIBRARY=/usr/lib64/libpmi.so # Option one: srun --mpi=pmix_v3 /path/to/mpiapplication <arguments> # Or using mpirun directly: mpirun /path/to/mpiapplication <arguments> # Third option through the hydra process manager mpiexec.hydra -bootstrap slurm -n <num_procs> /path/to/mpiapplication <arguments>
If you need more information about the locally installed OpenMPI 3 version, you can look around a worker node by
# Look into local openmpi3 manually on WN $ salloc -N1 -n1 $ ssh wn21123 # replace with correct wn number of your interaction session $ yum info openmpi3 $ ls /lib64/openmpi3 $ <ctrl-d to exit salloc>
Alternatively, if you are just interested in file locations:
$ srun -N1 -n1 tree /lib64/openmpi3 | less