...
CP2K is an MPI-parallel application. You can use either mpirun or srun as the job starter for CP2K. If you opt for mpirun, then, apart from loading the corresponding impi or openmpi modules, CPU and/or GPU pinning should be carefully carried out.
| CP2K Version | Modulefile | Requirement | Compute Partitions | Support | CPU/GPU |
|---|---|---|---|---|---|
| 7.1 | cp2k/7.1 | impi/2021.13 | CPU CLX | omp libint fftw3 libxc elpa parallel mpi3 scalapack xsmm spglib mkl |  /  |
|
|
| / | ||
|
|
| / | ||
| 2023.1 | cp2k/2023.1 | openmpi/gcc.11/4.1.4cuda/11.8 | GPU A100 | libint, fftw3, libxc, elpa, elpa_nvidia_gpu, scalapack, cosma, xsmm, dbcsr_acc, spglib, mkl, sirius, offload_cuda, spla_gemm, m_offloading, libvdwxc |
|
|
|
|
| ||
| 2023.2 | cp2k/2023.2 | openmpi/gcc.11/4.1.4cuda/11.8 | GPU A100 | libint, fftw3, libxc, elpa, elpa_nvidia_gpu, scalapack, cosma, xsmm, dbcsr_acc, spglib, mkl, sirius, offload_cuda, spla_gemm, m_offloading, libvdwxc |
|
| 2024.1 | cp2k/2024.1 | impi/2021.13 | CPU CLX | omp,libint,fftw3,fftw3_mkl,libxc,elpa,parallel,mpi_f08,scalapack,xsmm,spglib,mkl,sirius,hdf5 | / |
| 2024.1 | cp2k/2024.1 | openmpi/gcc/5.0.3 | CPU Genoa | omp,fftw3,libxc,elpa,parallel,mpi_f08,scalapack,cosma,xsmm,spglib,sirius,hdf5 | / |
| 2025.1 | cp2k/2025.1 | openmpi/gcc/5.0.3 | CPU Genoa | omp,libint,fftw3,libxc,elpa,parallel,mpi_f08,scalapack,cosma,xsmm,spglib,sirius,hdf5 | / |
Remark: cp2k needs special attention when running on GPUs.
...
| Codeblock | ||||
|---|---|---|---|---|
| ||||
#!/bin/bash
#SBATCH --time=12:00:00
#SBATCH --partition=cpu-clx
#SBATCH --nodes=1
#SBATCH --ntasks-per-node=24
#SBATCH --cpus-per-task=4
#SBATCH --job-name=cp2k
export SLURM_CPU_BIND=none
export OMP_NUM_THREADS=${SLURM_CPUS_PER_TASK}
# Binding OpenMP threads
export OMP_PLACES=cores
export OMP_PROC_BIND=close
# Binding MPI tasks
export I_MPI_PIN=yes
export I_MPI_PIN_DOMAIN=omp
export I_MPI_PIN_CELL=core
# Our tests have shown that CP2K has better performance with psm2 as libfabric provider
# Check if this also apply to your system
# To stick to the default provider, comment out the following line
export FI_PROVIDER=psm2
module load impi/2021.13
# Select the appropriate version
module load cp2k/2024.1
mpirun cp2k.psmp input > output |
| Codeblock | ||||
|---|---|---|---|---|
| ||||
#!/bin/bash #SBATCH --time=12:00:00 #SBATCH --partition=cpu-genoa #SBATCH --nodes=1 #SBATCH --ntasks-per-node=48 #SBATCH --cpus-per-task=4 #SBATCH --job-name=cp2k export SLURM_CPU_BIND=none export OMP_NUM_THREADS=${SLURM_CPUS_PER_TASK} # Binding OpenMP threads export OMP_PLACES=cores export OMP_PROC_BIND=close module load openmpi/gcc/5.0.3 # Select the appropriate version module load cp2k/2024.1 # Do not use srun combined with export SLURM_CPU_BIND=none # Important: here we are using mpirun to start the MPI process. The pinning is performed according to the following line mpirun --bind-to core --map-by ppr:${SLURM_NTASKS_PER_NODE}:node:pe=${OMP_NUM_THREADS} cp2k.psmp input > output |
| Codeblock | ||||
|---|---|---|---|---|
| ||||
#!/bin/bash
#SBATCH --partition=gpu-a100
#SBATCH --time=12:00:00
#SBATCH --nodes=1
#SBATCH --ntasks-per-node=4
#SBATCH --cpus-per-task=18
#SBATCH --job-name=cp2k
export SLURM_CPU_BIND=none
export OMP_NUM_THREADS=${SLURM_CPUS_PER_TASK}
export OMP_PLACES=cores
export OMP_PROC_BIND=close
module load gcc/11.3.0 openmpi/gcc.11/4.1.4 cuda/11.8 cp2k/2023.2
# gpu_bind.sh (see the following script) should be placed inside the same directory where cp2k will be executed
# Don't forget to make gpu_bind.sh executable by running: chmod +x gpu_bind.sh
mpirun --bind-to core --map-by numa:PE=${SLURM_CPUS_PER_TASK} ./gpu_bind.sh cp2k.psmp input > output |
...