LUMI (CSC)
The LUMI cluster is located at CSC (Finland). LUMI has multiple partitions.
On LUMI-G, each node contains four (4) AMD MI250X GPUs, each with 2 Graphics Compute Dies (GCDs) for a total of 8 GCDs per node. You can think of the 8 GCDs as 8 separate GPUs, each having 64 GB of high-bandwidth memory (HBM2E).
On LUMI-C, each node contains two (2) AMD EPYC 7763 CPUs.
Introduction
If you are new to this system, please see the following resources:
Batch system: Slurm
-
$HOME: single user, intended to store user configuration files and personal data (20GB default quota)/project/$proj: shared with all members of a project, purged at the end of a project (50 GB default quota)/scratch/$proj: temporary storage, main storage to be used for disk I/O needs when running simulations on LUMI, purged every 90 days (50TB default quota)
Preparation
Use the following commands to download the ImpactX source code:
git clone https://github.com/BLAST-ImpactX/impactx.git $HOME/src/impactx
On LUMI, you can run either with fast MI250X GPUs (LUMI-G, recommended) or CPU nodes (LUMI-C).
We use system software modules, add environment hints and further dependencies via the file $HOME/lumi_gpu_impactx.profile.
Create it now:
cp $HOME/src/impactx/docs/source/install/hpc/lumi-csc/lumi_gpu_impactx.profile.example $HOME/lumi_gpu_impactx.profile
Edit the 2nd line of this script, which sets the export proj="project_..." variable using a text editor
such as nano, emacs, or vim (all available by default on LUMI login nodes).
You can find out your project name by running lumi-ldap-userinfo on LUMI.
For example, if you are member of the project project_465000962, then run nano $HOME/lumi_gpu_impactx.profile and edit line 2 to read:
export proj="project_465000962"
Exit the nano editor with Ctrl + O (save) and then Ctrl + X (exit).
Important
Now, and as the first step on future logins to LUMI, activate these environment settings:
source $HOME/lumi_gpu_impactx.profile
Finally, since LUMI does not yet provide software modules for some of our dependencies, install them once:
bash $HOME/src/impactx/docs/source/install/hpc/lumi-csc/install_gpu_dependencies.sh
source $HOME/sw/lumi/gpu/venvs/impactx-gpu-lumi/bin/activate
We use system software modules, add environment hints and further dependencies via the file $HOME/lumi_cpu_impactx.profile.
Create it now:
cp $HOME/src/impactx/docs/source/install/hpc/lumi-csc/lumi_cpu_impactx.profile.example $HOME/lumi_cpu_impactx.profile
Edit the 2nd line of this script, which sets the export proj="project_..." variable using a text editor
such as nano, emacs, or vim (all available by default on LUMI login nodes).
You can find out your project name by running lumi-ldap-userinfo on LUMI.
For example, if you are member of the project project_465000962, then run nano $HOME/lumi_cpu_impactx.profile and edit line 2 to read:
export proj="project_465000962"
Exit the nano editor with Ctrl + O (save) and then Ctrl + X (exit).
Important
Now, and as the first step on future logins to LUMI, activate these environment settings:
source $HOME/lumi_cpu_impactx.profile
Finally, since LUMI does not yet provide software modules for some of our dependencies, install them once:
bash $HOME/src/impactx/docs/source/install/hpc/lumi-csc/install_cpu_dependencies.sh
source $HOME/sw/lumi/cpu/venvs/impactx-cpu-lumi/bin/activate
Compilation
Use the following cmake commands to compile the application executable:
cd $HOME/src/impactx
cmake --fresh -S . -B build_lumi_gpu -DImpactX_COMPUTE=HIP -DImpactX_FFT=ON
cmake --build build_lumi_gpu -j 16
The ImpactX application executables are now in $HOME/src/impactx/build_lumi_gpu/bin/.
Additionally, the following commands will install ImpactX as a Python module:
cmake --fresh -S . -B build_lumi_gpu_py -DImpactX_COMPUTE=HIP -DImpactX_FFT=ON -DImpactX_PYTHON=ON
cmake --build build_lumi_gpu_py -j 16 --target pip_install
cd $HOME/src/impactx
cmake --fresh -S . -B build_lumi_cpu -DImpactX_FFT=ON
cmake --build build_lumi_cpu -j 16
The ImpactX application executables are now in $HOME/src/impactx/build_lumi_cpu/bin/.
Additionally, the following commands will install ImpactX as a Python module:
cmake --fresh -S . -B build_lumi_cpu_py -DImpactX_FFT=ON -DImpactX_PYTHON=ON
cmake --build build_lumi_cpu_py -j 16 --target pip_install
Now, you can submit LUMI compute jobs for ImpactX Python scripts (example scripts).
Or, you can use the ImpactX executables to submit LUMI jobs (example inputs).
For executables, you can reference their location in your job script or copy them to a location in /project/$proj or /scratch/$proj.
Update ImpactX & Dependencies
If you already installed ImpactX in the past and want to update it, start by getting the latest source code:
cd $HOME/src/impactx
# read the output of this command - does it look ok?
git status
# get the latest ImpactX source code
git fetch
git pull
# read the output of these commands - do they look ok?
git status
git log # press q to exit
And, if needed,
update the lumi_gpu_impactx.profile and lumi_cpu_impactx.profile files,
log out and into the system, activate the now updated environment profile as usual,
As the last step, recompile ImpactX.
Running
On LUMI, compute jobs are run via the Slurm resource manager. There are various Slurm priority queues besides the defaults documented below.
For interactive runs on a single node, simply use the aliases getNode or runNode ... that are defined by the lumi_*_impactx.profile.
For runs on multiple nodes, use and adjust the templates below.
The LUMI-G (GPU) partition on the supercomputer LUMI at CSC has up to 2978 nodes, each with 8 Graphics Compute Dies (GCDs). ImpactX runs one MPI rank per Graphics Compute Die.
The batch script below can be used to run an ImpactX simulation on multiple nodes (change -N accordingly).
Replace descriptions between chevrons <> by relevant values, for instance <project id> or the concete inputs file.
Copy the executable or point to it via EXE and adjust the path for the INPUTS variable accordingly.
$HOME/src/impactx/docs/source/install/hpc/lumi-csc/lumi_gpu.sbatch.#!/bin/bash -l
#SBATCH -A <project id>
#SBATCH --nodes=1
#SBATCH --time=00:10:00
#SBATCH --job-name=impactx
#SBATCH --output=ImpactX.o%j
#SBATCH --error=ImpactX.e%j
#SBATCH --partition=standard-g
#SBATCH --ntasks-per-node=8
#SBATCH --gpus-per-node=8
# executable & inputs file or ...
EXE=./impactx
INPUTS=inputs
# ... python interpreter & PICMI script here
#EXE=python3
#INPUTS=run_fodo.py
date
# note (12-12-22)
# this environment setting is currently needed on LUMI to work-around a
# known issue with Libfabric
#export FI_MR_CACHE_MAX_COUNT=0 # libfabric disable caching
# or, less invasive:
export FI_MR_CACHE_MONITOR=memhooks # alternative cache monitor
# Seen since August 2023 seen on OLCF (not yet seen on LUMI?)
# OLCFDEV-1597: OFI Poll Failed UNDELIVERABLE Errors
# https://docs.olcf.ornl.gov/systems/frontier_user_guide.html#olcfdev-1597-ofi-poll-failed-undeliverable-errors
#export MPICH_SMP_SINGLE_COPY_MODE=NONE
#export FI_CXI_RX_MATCH_MODE=software
# note (9-2-22, OLCFDEV-1079)
# this environment setting is needed to avoid that rocFFT writes a cache in
# the home directory, which does not scale.
export ROCFFT_RTC_CACHE_PATH=/dev/null
# Seen since August 2023
# OLCFDEV-1597: OFI Poll Failed UNDELIVERABLE Errors
# https://docs.olcf.ornl.gov/systems/frontier_user_guide.html#olcfdev-1597-ofi-poll-failed-undeliverable-errors
export MPICH_SMP_SINGLE_COPY_MODE=NONE
export FI_CXI_RX_MATCH_MODE=software
# LUMI documentation suggests using the following wrapper script
# to set the ROCR_VISIBLE_DEVICES to the value of SLURM_LOCALID
# see https://docs.lumi-supercomputer.eu/runjobs/scheduled-jobs/lumig-job/
cat << EOF > select_gpu
#!/bin/bash
export ROCR_VISIBLE_DEVICES=\$SLURM_LOCALID
exec \$*
EOF
chmod +x ./select_gpu
sleep 1
# LUMI documentation suggests using the following CPU bind
# in order to have 6 threads per GPU (blosc compression in adios2 uses threads)
# see https://docs.lumi-supercomputer.eu/runjobs/scheduled-jobs/lumig-job/
#
# WARNING: the following CPU_BIND options don't work on the dev-g partition.
# If you want to run your simulation on dev-g, please comment them
# out and replace them with CPU_BIND="map_cpu:49,57,17,25,1,9,33,41"
#
CPU_BIND="mask_cpu:7e000000000000,7e00000000000000"
CPU_BIND="${CPU_BIND},7e0000,7e000000"
CPU_BIND="${CPU_BIND},7e,7e00"
CPU_BIND="${CPU_BIND},7e00000000,7e0000000000"
export OMP_NUM_THREADS=6
export MPICH_GPU_SUPPORT_ENABLED=1
srun --cpu-bind=${CPU_BIND} ./select_gpu \
${EXE} ${INPUTS} \
| tee outputs.txt
rm -rf ./select_gpu
To run a simulation, copy the lines above to a file lumi_gpu.sbatch and run
sbatch lumi_gpu.sbatch
to submit the job.
The LUMI-C (CPU) partition on the supercomputer LUMI at CSC has up to 2048 nodes. ImpactX runs 16 MPI ranks per node, each with 8 OpenMP threads.
The batch script below can be used to run an ImpactX simulation on multiple nodes (change -N accordingly).
Replace descriptions between chevrons <> by relevant values, for instance <project id> or the concete inputs file.
Copy the executable or point to it via EXE and adjust the path for the INPUTS variable accordingly.
$HOME/src/impactx/docs/source/install/hpc/lumi-csc/lumi_cpu.sbatch.#!/bin/bash -l
#SBATCH -A <project id>
#SBATCH --nodes=1
#SBATCH --time=00:10:00
#SBATCH --job-name=impactx
#SBATCH --output=ImpactX.o%j
#SBATCH --error=ImpactX.e%j
#SBATCH --partition=standard
#SBATCH --ntasks-per-node=16
#SBATCH --cpus-per-task=8
# executable & inputs file or ...
EXE=./impactx
INPUTS=inputs
# ... python interpreter & PICMI script here
#EXE=python3
#INPUTS=run_fodo.py
date
# note (12-12-22)
# this environment setting is currently needed on LUMI to work-around a
# known issue with Libfabric
#export FI_MR_CACHE_MAX_COUNT=0 # libfabric disable caching
# or, less invasive:
export FI_MR_CACHE_MONITOR=memhooks # alternative cache monitor
# Seen since August 2023 seen on OLCF (not yet seen on LUMI?)
# OLCFDEV-1597: OFI Poll Failed UNDELIVERABLE Errors
# https://docs.olcf.ornl.gov/systems/frontier_user_guide.html#olcfdev-1597-ofi-poll-failed-undeliverable-errors
#export MPICH_SMP_SINGLE_COPY_MODE=NONE
#export FI_CXI_RX_MATCH_MODE=software
# Seen since August 2023
# OLCFDEV-1597: OFI Poll Failed UNDELIVERABLE Errors
# https://docs.olcf.ornl.gov/systems/frontier_user_guide.html#olcfdev-1597-ofi-poll-failed-undeliverable-errors
export MPICH_SMP_SINGLE_COPY_MODE=NONE
export FI_CXI_RX_MATCH_MODE=software
# The LUMI-C compute nodes are configured with 4 NUMA zones ("quadrant mode") with 2 CCDs per quadrant.
# https://docs.lumi-supercomputer.eu/hardware/lumic/
#
# LUMI documentation suggests using the following CPU bind
# see https://docs.lumi-supercomputer.eu/runjobs/scheduled-jobs/lumic-job/
export OMP_NUM_THREADS=${SLURM_CPUS_PER_TASK}
srun \
${EXE} ${INPUTS} \
| tee outputs.txt
To run a simulation, copy the lines above to a file lumi_cpu.sbatch and run
sbatch lumi_cpu.sbatch
to submit the job.
Post-Processing
LUMI provides a Jupyter Lab service that can be used for interactive post-processing.
You can reuse your CPU ImpactX virtual environment for post-processing in Jupyter Lab:
* Python: Cray-python (3.10.10)
* Virtual environment path: ${HOME}/sw/lumi/cpu/venvs/impactx-cpu-lumi/ (replace ${HOME} with the output of echo ${HOME})
Known System Issues
Warning
December 12th, 2022: There is a caching bug in libFabric that causes ImpactX simulations to occasionally hang on LUMI on more than 1 node.
As a work-around, please export the following environment variable in your job scripts until the issue is fixed:
#export FI_MR_CACHE_MAX_COUNT=0 # libfabric disable caching
# or, less invasive:
export FI_MR_CACHE_MONITOR=memhooks # alternative cache monitor
Warning
January, 2023: We discovered a regression in AMD ROCm, leading to 2x slower current deposition (and other slowdowns) in ROCm 5.3 and 5.4.
June, 2023: Although a fix was planned for ROCm 5.5, we still see the same issue in this release and continue to exchange with AMD and HPE on the issue.
Stay with the ROCm 5.2 module to avoid a 2x slowdown.
Warning
May 2023: rocFFT in ROCm 5.1-5.3 tries to write to a cache in the home area by default. This does not scale, disable it via:
export ROCFFT_RTC_CACHE_PATH=/dev/null