kokkos-workshop

Useful links

• Kokkos Core GitHub repository
• Kokkos Tutorials
• Kokkos slides and videos
• Ruche documentation
• Cheat sheet

Ruche cluster

Ruche is the mesocentre of Paris-Saclay. It provides a computing environment for research and teaching. Ruche is equipped with both CPU (Intel Xeon Gold 6230 20C @ 2.1GHz) and GPU nodes (Nvidia Tesla V100 and A100).

Connect to Ruche

ssh <login>@ruche.mesocentre.universite-paris-saclay.fr

SSH configuration:

Host ruche
   User <login>
   Hostname ruche.mesocentre.universite-paris-saclay.fr

Editors like vim and nano are available on Ruche. VSCode can be connected to Ruche using the Remote SSH extension.

Environments for Kokkos

CPU environment (OpenMP backend)

GNU compiler

Preferred modules for GNU compilers:

module load gcc/11.2.0/gcc-4.8.5
module load cmake/3.28.3/gcc-11.2.0
module load git/2.39.1/gcc-11.2.0-tcltk

We provide an already installed version of Kokkos 4.3.1 in case:

source /gpfs/workdir/cexaci/installation/env/kokkos4.3.1_gnu11.2.sh
# -> provides $Kokkos_ROOT=...

GPU environment (CUDA backend)

Preferred modules for the NVIDIA environment:

module load gcc/11.2.0/gcc-4.8.5
module load cmake/3.28.3/gcc-11.2.0
module load git/2.39.1/gcc-11.2.0-tcltk
module load cuda/12.2.1/gcc-11.2.0

Installed Kokkos version for V100:

We have installed for you a Kokkos version optimized for V100:

source /gpfs/workdir/cexaci/installation/env/kokkos4.3.1_cuda12_v100.sh
# -> provides $Kokkos_ROOT=...

The Kokkos version for V100 uses the following flags:

cmake -B build -DCMAKE_CXX_COMPILER=g++ \
-DCMAKE_INSTALL_PREFIX=${KOKKOS_HOME} \
-DKokkos_ENABLE_OPENMP=ON \
-DKokkos_ENABLE_CUDA=ON \
-DKokkos_ARCH_VOLTA70=ON \
-DKokkos_ENABLE_CUDA_LAMBDA=ON \
-DKokkos_ENABLE_CUDA_CONSTEXPR=ON

Installed Kokkos version for A100:

We have installed for you a Kokkos version optimized for A100:

source /gpfs/workdir/cexaci/installation/env/kokkos4.3.1_cuda12_a100.sh
# -> provides $Kokkos_ROOT=...

The Kokkos version for A100 uses the following flags:

cmake -B build -DCMAKE_CXX_COMPILER=g++ \
-DCMAKE_INSTALL_PREFIX=${KOKKOS_HOME} \
-DKokkos_ENABLE_OPENMP=ON \
-DKokkos_ENABLE_CUDA=ON \
-DKokkos_ARCH_AMPERE80=ON \
-DKokkos_ENABLE_CUDA_LAMBDA=ON \
-DKokkos_ENABLE_CUDA_CONSTEXPR=ON

SLURM scripts examples

CPU node

Slurm example to compile and execute a tutorial exercise with OpenMP backend on a CPU node:

#!/bin/bash
#SBATCH --job-name=kokkos
#SBATCH --output=output
#SBATCH --error=error
#SBATCH --time=00:10:00
#SBATCH --ntasks=4
#SBATCH --partition=cpu_short

# ____________________________________________________________________
# Environment

# For script debugging
set -x

# Ensure that the job is launched from the directory where the script is located
cd ${SLURM_SUBMIT_DIR}

# Load here your modules and prepare your env

module load ...

# to use an already installed version of Kokkos
export Kokkos_ROOT=...

# ____________________________________________________________________
# Compilation

cmake -B build -DCMAKE_CXX_COMPILER=g++
cmake --build build --parallel 4

# ____________________________________________________________________
# Execution

# OpenMP parameters
export OMP_PROC_BIND=spread
export OMP_PLACES=threads
export OMP_NUM_THREADS=4

# Execution
./build/Begin/$(basename $(pwd))_Exercise

# Execution solution
./build/Solution/$(basename $(pwd))_Solution

V100 GPU node

Slurm example to compile and execute a tutorial exercise with OpenMP and CUDA backend on a GPU node:

#!/bin/bash
#SBATCH --job-name=kokkos
#SBATCH --output=output
#SBATCH --error=error
#SBATCH --time=00:10:00
#SBATCH --cpus-per-task=10
#SBATCH --partition=gpu_test
#SBATCH --exclude=ruche-gpu08  # This node has an older CUDA version
#SBATCH --gres=gpu:1

# ____________________________________________________________________
# Environment

# For script debugging
set -x

# Ensure that the job is launched from the directory where the script is located
cd ${SLURM_SUBMIT_DIR}

# Load here your modules and prepare your env

module load ...

# to use an already installed version of Kokkos
export Kokkos_ROOT=...

# ____________________________________________________________________
# Compilation

cmake -B build_cuda -DCMAKE_CXX_COMPILER=g++ -DKokkos_ENABLE_CUDA=ON
cmake --build build_cuda --parallel 4

# ____________________________________________________________________
# Execution

# OpenMP parameters
export OMP_PROC_BIND=spread
export OMP_PLACES=threads
export OMP_NUM_THREADS=1

# Execution
./build_cuda/Begin/$(basename $(pwd))_Exercise

# Execution solution
./build_cuda/Solution/$(basename $(pwd))_Solution

A100 GPU node

Slurm example to compile and execute a tutorial exercise with OpenMP and CUDA backend on a A100 GPU node:

#!/bin/bash
#SBATCH --job-name=kokkos
#SBATCH --output=output
#SBATCH --error=error
#SBATCH --time=00:10:00
#SBATCH --cpus-per-task=10
#SBATCH --partition=gpua100
#SBATCH --gres=gpu:1

# ____________________________________________________________________
# Environment

# For script debugging
set -x

# Ensure that the job is launched from the directory where the script is located
cd ${SLURM_SUBMIT_DIR}

# Load here your modules and prepare your env

module load ...

# to use an already installed version of Kokkos
export Kokkos_ROOT=...

# ____________________________________________________________________
# Compilation

cmake -B build_cuda -DCMAKE_CXX_COMPILER=g++ -DKokkos_ENABLE_CUDA=ON
cmake --build build_cuda --parallel 4

# ____________________________________________________________________
# Execution

# OpenMP parameters
export OMP_PROC_BIND=spread
export OMP_PLACES=threads
export OMP_NUM_THREADS=1

# Execution
./build_cuda/Begin/$(basename $(pwd))_Exercise

# Execution solution
./build_cuda/Solution/$(basename $(pwd))_Solution