Juwels (JSC)

Note

For the moment, WarpX doesn’t run on Juwels with MPI_THREAD_MULTIPLE. Please compile with this compilation flag: MPI_THREAD_MULTIPLE=FALSE.

The Juwels supercomputer is located at JSC.

Introduction

If you are new to this system, please see the following resources:

See this page for a quick introduction. (Full user guide).

  • Batch system: Slurm

  • Production directories:

    • $SCRATCH: Scratch filesystem for temporary data (90 day purge)

    • $FASTDATA/: Storage location for large data (backed up)

    • Note that the $HOME directory is not designed for simulation runs and producing output there will impact performance.

Installation

Use the following commands to download the WarpX source code and switch to the correct branch:

git clone https://github.com/ECP-WarpX/WarpX.git $HOME/src/warpx

We use the following modules and environments on the system.

Listing 21 You can copy this file from Tools/machines/juwels-jsc/juwels_warpx.profile.example.
# please set your project account
#export proj=<yourProject>

# required dependencies
module load ccache
module load CMake
module load GCC
module load CUDA/11.3
module load OpenMPI
module load FFTW
module load HDF5
module load Python

# JUWELS' job scheduler may not map ranks to GPUs,
# so we give a hint to AMReX about the node layout.
# This is usually done in Make.<supercomputing center> files in AMReX
# but there is no such file for JSC yet.
export GPUS_PER_SOCKET=2
export GPUS_PER_NODE=4

# optimize CUDA compilation for V100 (7.0) or for A100 (8.0)
export AMREX_CUDA_ARCH=8.0

Note that for now WarpX must rely on OpenMPI instead of the recommended MPI implementation on this platform MVAPICH2.

We recommend to store the above lines in a file, such as $HOME/juwels_warpx.profile, and load it into your shell after a login:

source $HOME/juwels_warpx.profile

Then, cd into the directory $HOME/src/warpx and use the following commands to compile:

cd $HOME/src/warpx
rm -rf build

cmake -S . -B build -DWarpX_COMPUTE=CUDA -DWarpX_PSATD=ON -DWarpX_MPI_THREAD_MULTIPLE=OFF
cmake --build build -j 16

The other general compile-time options apply as usual.

That’s it! A 3D WarpX executable is now in build/bin/ and can be run with a 3D example inputs file. Most people execute the binary directly or copy it out to a location in $SCRATCH.

Note

Currently, if you want to use HDF5 output with openPMD, you need to add

export OMPI_MCA_io=romio321

in your job scripts, before running the srun command.

Running

Queue: gpus (4 x Nvidia V100 GPUs)

The Juwels GPUs are V100 (16GB) and A100 (40GB).

An example submission script reads

Listing 22 You can copy this file from Tools/machines/juwels-jsc/juwels.sbatch.
#!/bin/bash -l

#SBATCH -A $proj
#SBATCH --partition=booster
#SBATCH --nodes=2
#SBATCH --ntasks=8
#SBATCH --ntasks-per-node=4
#SBATCH --gres=gpu:4
#SBATCH --time=00:05:00
#SBATCH --job-name=warpx
#SBATCH --output=warpx-%j-%N.txt
#SBATCH --error=warpx-%j-%N.err

export OMP_NUM_THREADS=1
export OMPI_MCA_io=romio321  # for HDF5 support in openPMD

# you can comment this out if you sourced the warpx.profile
# files before running sbatch:
module load GCC
module load OpenMPI
module load CUDA/11.3
module load HDF5
module load Python

srun -n 8 --cpu_bind=sockets $HOME/src/warpx/build/bin/warpx.3d.MPI.CUDA.DP.OPMD.QED inputs

Queue: batch (2 x Intel Xeon Platinum 8168 CPUs, 24 Cores + 24 Hyperthreads/CPU)

todo

See the data analysis section for more information on how to visualize the simulation results.