Arepo
Useful links
Licensing
Arepo is released under GPLv3 (see the Arepo licence).
Version
Latest version on Arepo
Building
We should load the Cray hdf5 module
module load cray-hdf5
So on XCI system you can:
Install GSL with:
wget https://ftp.gnu.org/gnu/gsl/gsl-latest.tar.gz
tar xxvf gsl-latest.tar.gz
cd gsl-*
./configure --prefix=$HOME/gsl CC=cc
make -j
make install
Then in Arepo specify in Makefile.systype
SYSTYPE="XCI"
And in Makefile
# Isambard XCI
ifeq ($(SYSTYPE),"XCI")
CC=cc
# compiler and its optimization options
OPTIMIZE =
# overwrite default:
MPICH_INCL=
MPICH_LIB =
GSL_INCL = -I$(HOME)/gsl/include
GSL_LIB = -L$(HOME)/gsl/lib -lgsl -lgslcblas
HWLOC_LIB =
# libraries that are included on demand, depending on Config.sh options
FFTW_INCL =
FFTW_LIBS =
HDF5_INCL = -DH5_USE_16_API
HDF5_LIB =
HWLOC_INCL=
endif
Then:
module load cray-hdf5
module load cray-fftw
make
Then it should compile an Arepo executable.
To use and build examples a venv is required:
module load cray-python
python -m venv venv
. venv/bin/activate
pip install “numpy<1.20”
pip install h5py
(numpy<1.20 needed due to broken code using np.int)
Then followed docs on using examples.
Running
Example Job Submission Script
Arm
#!/bin/bash
# Select 1 nodes (maximum of 64 cores)
#PBS -l select=1
# Set maximum job walltime
#PBS -l walltime=00:30:00
# Change to the directory the script was submitted from
cd $PBS_O_WORKDIR
# Load required GSL
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$HOME/gsl/lib
# Launch 8192 MPI ranks (64 per node, 1 per core)
# Launch 1 OpenMP threads per rank (NOT using 4-way hyperthreading)
export OMP_NUM_THREADS=1
aprun -n 64 -d 1 -j 1 Arepo
Notes
For maximum performance, you may wish to tune the MPI/OpenMP decomposition. At large scale, you may find that reducing the number of MPI ranks and increasing the number of OpenMP threads yields better performance.