Cori (NERSC)

The Cori cluster is located at NERSC.

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

Installation

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

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

KNL

We use the following modules and environments on the system ($HOME/knl_impactx.profile).

module swap craype-haswell craype-mic-knl
module swap PrgEnv-intel PrgEnv-gnu
module load cmake/3.22.1
module load cray-hdf5-parallel/1.10.5.2
module load cray-fftw/3.3.8.10
module load cray-python/3.9.7.1

export PKG_CONFIG_PATH=$FFTW_DIR/pkgconfig:$PKG_CONFIG_PATH
export CMAKE_PREFIX_PATH=$HOME/sw/knl/adios2-2.7.1-install:$CMAKE_PREFIX_PATH

if [ -d "$HOME/sw/knl/venvs/impactx" ]
then
  source $HOME/sw/knl/venvs/impactx/bin/activate
fi

export CXXFLAGS="-march=knl"
export CFLAGS="-march=knl"

For PICMI and Python workflows, also install a virtual environment:

# establish Python dependencies
python3 -m pip install --user --upgrade pip
python3 -m pip install --user virtualenv

python3 -m venv $HOME/sw/knl/venvs/impactx
source $HOME/sw/knl/venvs/impactx/bin/activate

python3 -m pip install --upgrade pip
MPICC="cc -shared" python3 -m pip install -U --no-cache-dir -v mpi4py
python3 -m pip install -r $HOME/src/impactx/requirements.txt

Haswell

We use the following modules and environments on the system ($HOME/haswell_impactx.profile).

module swap PrgEnv-intel PrgEnv-gnu
module load cmake/3.22.1
module load cray-hdf5-parallel/1.10.5.2
module load cray-fftw/3.3.8.10
module load cray-python/3.9.7.1

export PKG_CONFIG_PATH=$FFTW_DIR/pkgconfig:$PKG_CONFIG_PATH
export CMAKE_PREFIX_PATH=$HOME/sw/haswell/adios2-2.7.1-install:$CMAKE_PREFIX_PATH

if [ -d "$HOME/sw/haswell/venvs/impactx" ]
then
  source $HOME/sw/haswell/venvs/impactx/bin/activate
fi

For PICMI and Python workflows, also install a virtual environment:

# establish Python dependencies
python3 -m pip install --user --upgrade pip
python3 -m pip install --user virtualenv

python3 -m venv $HOME/sw/haswell/venvs/impactx
source $HOME/sw/haswell/venvs/impactx/bin/activate

python3 -m pip install --upgrade pip
MPICC="cc -shared" python3 -m pip install -U --no-cache-dir -v mpi4py
python3 -m pip install -r $HOME/src/impactx/requirements.txt

GPU (V100)

Cori provides a partition with 18 nodes that include V100 (16 GB) GPUs. We use the following modules and environments on the system ($HOME/gpu_impactx.profile).

export proj="m1759"

module purge
module load modules
module load cgpu
module load esslurm
module load gcc/8.3.0 cuda/11.4.0 cmake/3.22.1
module load openmpi

export CMAKE_PREFIX_PATH=$HOME/sw/cori_gpu/adios2-2.7.1-install:$CMAKE_PREFIX_PATH

if [ -d "$HOME/sw/cori_gpu/venvs/impactx" ]
then
  source $HOME/sw/cori_gpu/venvs/impactx/bin/activate
fi

# compiler environment hints
export CC=$(which gcc)
export CXX=$(which g++)
export FC=$(which gfortran)
export CUDACXX=$(which nvcc)
export CUDAHOSTCXX=$(which g++)

# optimize CUDA compilation for V100
export AMREX_CUDA_ARCH=7.0

# allocate a GPU, e.g. to compile on
#   10 logical cores (5 physical), 1 GPU
function getNode() {
    salloc -C gpu -N 1 -t 30 -c 10 --gres=gpu:1 -A $proj
}

For PICMI and Python workflows, also install a virtual environment:

# establish Python dependencies
python3 -m pip install --user --upgrade pip
python3 -m pip install --user virtualenv

python3 -m venv $HOME/sw/cori_gpu/venvs/impactx
source $HOME/sw/cori_gpu/venvs/impactx/bin/activate

python3 -m pip install --upgrade pip
python3 -m pip install -U --no-cache-dir -v mpi4py
python3 -m pip install -r $HOME/src/impactx/requirements.txt

Building ImpactX

We recommend to store the above lines in individual impactx.profile files, as suggested above. If you want to run on either of the three partitions of Cori, open a new terminal, log into Cori and source the environment you want to work with:

# KNL:
source $HOME/knl_impactx.profile

# Haswell:
#source $HOME/haswell_impactx.profile

# GPU:
#source $HOME/gpu_impactx.profile

Warning

Consider that all three Cori partitions are incompatible.

Do not source multiple ...impactx.profile files in the same terminal session. Open a new terminal and log into Cori again, if you want to switch the targeted Cori partition.

If you re-submit an already compiled simulation that you ran on another day or in another session, make sure to source the corresponding ...impactx.profile again after login!

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

cd $HOME/src/impactx
rm -rf build

#                       append if you target GPUs:    -DImpactX_COMPUTE=CUDA
cmake -S . -B build -DImpactX_OPENPMD=ON -DImpactX_DIMS=3
cmake --build build -j 16

Testing

To run all tests (here on KNL), do:

srun -C knl -N 1 -t 30 -q debug ctest --test-dir build --output-on-failure

Running

Navigate (i.e. cd) into one of the production directories (e.g. $SCRATCH) before executing the instructions below.

KNL

The batch script below can be used to run a ImpactX simulation on 2 KNL nodes on the supercomputer Cori at NERSC. Replace descriptions between chevrons <> by relevant values, for instance <job name> could be laserWakefield.

Do not forget to first source $HOME/knl_impactx.profile if you have not done so already for this terminal session.

For PICMI Python runs, the <path/to/executable> has to read python3 and the <input file> is the path to your PICMI input script.

#!/bin/bash -l

# Copyright 2019 Maxence Thevenet
#
# This file is part of ImpactX.
#
# License: BSD-3-Clause-LBNL


#SBATCH -N 2
#SBATCH -t 01:00:00
#SBATCH -q regular
#SBATCH -C knl
#SBATCH -S 4
#SBATCH -J <job name>
#SBATCH -A <allocation ID>
#SBATCH -e ImpactX.e%j
#SBATCH -o ImpactX.o%j

export OMP_PLACES=threads
export OMP_PROC_BIND=spread

# KNLs have 4 hyperthreads max
export CORI_MAX_HYPETHREAD_LEVEL=4
# We use 64 cores out of the 68 available on Cori KNL,
# and leave 4 to the system (see "#SBATCH -S 4" above).
export CORI_NCORES_PER_NODE=64

# Typically use 8 MPI ranks per node without hyperthreading,
# i.e., OMP_NUM_THREADS=8
export IMPACTX_NMPI_PER_NODE=8
export IMPACTX_HYPERTHREAD_LEVEL=1

# Compute OMP_NUM_THREADS and the thread count (-c option)
export CORI_NHYPERTHREADS_MAX=$(( ${CORI_MAX_HYPETHREAD_LEVEL} * ${CORI_NCORES_PER_NODE} ))
export IMPACTX_NTHREADS_PER_NODE=$(( ${IMPACTX_HYPERTHREAD_LEVEL} * ${CORI_NCORES_PER_NODE} ))
export OMP_NUM_THREADS=$(( ${IMPACTX_NTHREADS_PER_NODE} / ${IMPACTX_NMPI_PER_NODE} ))
export IMPACTX_THREAD_COUNT=$(( ${CORI_NHYPERTHREADS_MAX} / ${IMPACTX_NMPI_PER_NODE} ))

# for async_io support: (optional)
export MPICH_MAX_THREAD_SAFETY=multiple

srun --cpu_bind=cores -n $(( ${SLURM_JOB_NUM_NODES} * ${IMPACTX_NMPI_PER_NODE} )) -c ${IMPACTX_THREAD_COUNT} \
  <path/to/executable> <input file> \
  > output.txt

To run a simulation, copy the lines above to a file batch_cori.sh and run

sbatch batch_cori.sh

to submit the job.

For a 3D simulation with a few (1-4) particles per cell using FDTD Maxwell solver on Cori KNL for a well load-balanced problem (in our case laser wakefield acceleration simulation in a boosted frame in the quasi-linear regime), the following set of parameters provided good performance:

  • amr.max_grid_size=64 and amr.blocking_factor=64 so that the size of each grid is fixed to 64**3 (we are not using load-balancing here).

  • 8 MPI ranks per KNL node, with OMP_NUM_THREADS=8 (that is 64 threads per KNL node, i.e. 1 thread per physical core, and 4 cores left to the system).

  • 2 grids per MPI, i.e., 16 grids per KNL node.

Haswell

The batch script below can be used to run a ImpactX simulation on 1 Haswell node on the supercomputer Cori at NERSC.

Do not forget to first source $HOME/haswell_impactx.profile if you have not done so already for this terminal session.

#!/bin/bash -l

# Just increase this number of you need more nodes.
#SBATCH -N 1
#SBATCH -t 03:00:00
#SBATCH -q regular
#SBATCH -C haswell
#SBATCH -J <job name>
#SBATCH -A <allocation ID>
#SBATCH -e ImpactX.e%j
#SBATCH -o ImpactX.o%j
# one MPI rank per half-socket (see below)
#SBATCH --tasks-per-node=4
# request all logical (virtual) cores per half-socket
#SBATCH --cpus-per-task=16


# each Cori Haswell node has 2 sockets of Intel Xeon E5-2698 v3
# each Xeon CPU is divided into 2 bus rings that each have direct L3 access
export IMPACTX_NMPI_PER_NODE=4

# each MPI rank per half-socket has 8 physical cores
#   or 16 logical (virtual) cores
# over-subscribing each physical core with 2x
#   hyperthreading leads to a slight (3.5%) speedup
# the settings below make sure threads are close to the
#   controlling MPI rank (process) per half socket and
#   distribute equally over close-by physical cores and,
#   for N>8, also equally over close-by logical cores
export OMP_PROC_BIND=spread
export OMP_PLACES=threads
export OMP_NUM_THREADS=16

# for async_io support: (optional)
export MPICH_MAX_THREAD_SAFETY=multiple

EXE="<path/to/executable>"

srun --cpu_bind=cores -n $(( ${SLURM_JOB_NUM_NODES} * ${IMPACTX_NMPI_PER_NODE} )) \
  ${EXE} <input file> \
  > output.txt

To run a simulation, copy the lines above to a file batch_cori_haswell.sh and run

sbatch batch_cori_haswell.sh

to submit the job.

For a 3D simulation with a few (1-4) particles per cell using FDTD Maxwell solver on Cori Haswell for a well load-balanced problem (in our case laser wakefield acceleration simulation in a boosted frame in the quasi-linear regime), the following set of parameters provided good performance:

GPU (V100)

Do not forget to first source $HOME/gpu_impactx.profile if you have not done so already for this terminal session.

Due to the limited amount of GPU development nodes, just request a single node with the above defined getNode function. For single-node runs, try to run one grid per GPU.

A multi-node batch script template can be found below:

#!/bin/bash -l

# Copyright 2021 Axel Huebl
# This file is part of ImpactX.
# License: BSD-3-Clause-LBNL
#
# Ref:
# - https://docs-dev.nersc.gov/cgpu/hardware/
# - https://docs-dev.nersc.gov/cgpu/access/
# - https://docs-dev.nersc.gov/cgpu/usage/#controlling-task-and-gpu-binding

# Just increase this number of you need more nodes.
#SBATCH -N 2
#SBATCH -t 03:00:00
#SBATCH -J <job name>
#SBATCH -A m1759
#SBATCH -q regular
#SBATCH -C gpu
# 8 V100 GPUs (16 GB) per node
#SBATCH --gres=gpu:8
#SBATCH --exclusive
# one MPI rank per GPU (a quarter-socket)
#SBATCH --tasks-per-node=8
# request all logical (virtual) cores per quarter-socket
#SBATCH --cpus-per-task=10
#SBATCH -e ImpactX.e%j
#SBATCH -o ImpactX.o%j


# each Cori GPU node has 2 sockets of Intel Xeon Gold 6148 ('Skylake') @ 2.40 GHz
export IMPACTX_NMPI_PER_NODE=8

# each MPI rank per half-socket has 10 physical cores
#   or 20 logical (virtual) cores
# we split half-sockets again by 2 to have one MPI rank per GPU
# over-subscribing each physical core with 2x
#   hyperthreading leads to often to slight speedup on Intel
# the settings below make sure threads are close to the
#   controlling MPI rank (process) per half socket and
#   distribute equally over close-by physical cores and,
#   for N>20, also equally over close-by logical cores
export OMP_PROC_BIND=spread
export OMP_PLACES=threads
export OMP_NUM_THREADS=10

# for async_io support: (optional)
export MPICH_MAX_THREAD_SAFETY=multiple

EXE="<path/to/executable>"

srun --cpu_bind=cores --gpus-per-task=1 --gpu-bind=map_gpu:0,1,2,3,4,5,6,7 \
  -n $(( ${SLURM_JOB_NUM_NODES} * ${IMPACTX_NMPI_PER_NODE} )) \
  ${EXE} <input file> \
  > output.txt

Post-Processing

For post-processing, most users use Python via NERSC’s Jupyter service (Docs).

As a one-time preparatory setup, create your own Conda environment as described in NERSC docs. In this manual, we often use this conda create line over the officially documented one:

conda create -n myenv -c conda-forge python mamba ipykernel ipympl matplotlib numpy pandas yt openpmd-viewer openpmd-api h5py fast-histogram

We then follow the Customizing Kernels with a Helper Shell Script section to finalize the setup of using this conda-environment as a custom Jupyter kernel.

When opening a Jupyter notebook, just select the name you picked for your custom kernel on the top right of the notebook.

Additional software can be installed later on, e.g., in a Jupyter cell using !mamba install -c conda-forge .... Software that is not available via conda can be installed via !python -m pip install ....