Perlmutter

If this is your first time building MFIX-Exa on Perlmutter, please review the general notes below and Basics section first.

• The documentation for this system can be found here.

• Perlmutter can be accessed from a system with ssh client installed. These connections also work from NETL’s SciLAN and Joule:

  ssh <username>@perlmutter-p1.nersc.gov
  ssh <username>@saul-p1.nersc.gov
  

  Login with your Iris password + (Google) Authenticator passcode.

• These instructions build MFIX-Exa on the login nodes using -j8 CPUs. You may have to decrease this value if there is high traffic or you may want to increase this value if you are on a compute node interactively.

• The cmake instructions compile to a build directory. The gmake instructions compile to a exec directory.

• For the dependencies, it is assumed that you have set the following environment variables:

  export HYPRE_INSTALL_DIR=$HOME/<path/to/my/hypre-install-dir>
  export CSG_INSTALL_DIR=$HOME/<path/to/my/csg-dep-install-dir>
  export CSG_LIB_DIR=$HOME/<path/to/my/csg-lib-install-dir>
  export ASCENT_INSTALL_DIR=$HOME/<path/to/my/ascent-install-dir>
  

  to a path that you have read/write access to, such as inside the $SCRATCH space corresponding to your account. You will need to recall these paths later if you want to build MFIX-Exa with the optional dependencies.

• After building the mfix executable (with cmake), you can build the PIC-to-DEM restarter app by executing the following command in the build directory

  cmake --build . --target pic2dem
  

Warning

Currently, the cmake GPU build with hypre support links but does not run. We are still investigating, but recommend using gmake on Perlmutter at this time.

Basics

Source code

Before building, first obtain a copy of the source code following the instructions on the MFIX-Exa website.

Modules

All of the build instructions below have been tested with the following modules and environment helpers

module load PrgEnv-gnu/8.3.3
module load cmake/3.22.0

export CC=cc
export CXX=CC
export FC=ftn

The GPU-enabled builds additionally require

module load cudatoolkit/11.5

export CUDACXX=$(which nvcc)
export CUDAHOSTCXX=CC

export MPICH_GPU_SUPPORT_ENABLED=1
export CRAY_ACCEL_TARGET=nvidia80

Full builds that utilize external dependencies, also require setting certain environment variables as discussed below.

Building MFIX-Exa

The commands below are the superbuild instructions, i.e., AMReX is built as part of the MFIX-Exa build process. To build MFIX-Exa with hypre, csg and/or ascent dependencies, you first need to build and install these libraries and their dependencies. Instructions on building the necessary dependencies are below and should be successfully installed first. There are two primary methods of building the code cmake and gmake which are provided separately below.

cmake

CPUGPUCPU-fullGPU-full

cmake -DMFIX_MPI=yes \
      -DMFIX_OMP=no \
      -DMFIX_GPU_BACKEND=NONE \
      -DAMReX_TINY_PROFILE=no \
      -DMFIX_CSG=no \
      -DMFIX_HYPRE=no \
      -DCMAKE_BUILD_TYPE=Release \
      ../
make -j8

gmake

CPUGPUCPU-fullGPU-full

make -C exec -j8 \
     COMP=gnu \
     USE_MPI=TRUE \
     USE_OMP=FALSE \
     USE_CUDA=FALSE \
     USE_TINY_PROFILE=FALSE \
     USE_CSG=FALSE \
     USE_HYPRE=FALSE \
     DEBUG=FALSE

Optional build dependencies

The following dependencies need to be built and installed prior to following any of the full build instructions above.

1. HYPRE

   CPUGPU

   git clone https://github.com/hypre-space/hypre.git
   pushd hypre/src/
   git checkout v2.26.0
   ./configure --prefix=$HYPRE_INSTALL_DIR --with-MPI
   make -j8 install
   popd
   

2. Catch2

   git clone --depth 1 --branch v2.13.7 https://github.com/catchorg/Catch2
   pushd Catch2/
   cmake -S . -B build -DCMAKE_INSTALL_PREFIX=$CSG_INSTALL_DIR
   cd build/
   make -j8 install
   popd
   

3. GMP

   wget --no-check-certificate https://ftp.gnu.org/gnu/gmp/gmp-6.2.1.tar.xz
   tar -xf gmp-6.2.1.tar.xz
   pushd gmp-6.2.1
   ./configure --prefix=$CSG_INSTALL_DIR
   make -j8 install
   popd
   

4. MPFR

   wget --no-check-certificate https://ftp.gnu.org/gnu/mpfr/mpfr-4.1.0.tar.xz
   tar -xf mpfr-4.1.0.tar.xz
   pushd mpfr-4.1.0/
   ./configure --with-gmp=$CSG_INSTALL_DIR --prefix=$CSG_INSTALL_DIR
   make -j8 install
   popd
   

5. Boost

   wget https://boostorg.jfrog.io/artifactory/main/release/1.81.0/source/boost_1_81_0.tar.gz
   tar -zxvf boost_1_81_0.tar.gz
   pushd boost_1_81_0/
   ./bootstrap.sh
   ./b2 install --prefix=$CSG_INSTALL_DIR
   popd
   

6. CGAL

   git clone --depth 1 --branch v5.3 https://github.com/CGAL/cgal
   pushd cgal/
   cmake -S . -B build -DCMAKE_INSTALL_PREFIX=$CSG_INSTALL_DIR \
                       -DCMAKE_C_COMPILER=$(which cc) \
                       -DCMAKE_CXX_COMPILER=$(which CC)
   cd build/
   make -j8 install
   popd
   

7. PEGTL

   git clone --branch 3.2.2 https://github.com/taocpp/PEGTL
   pushd PEGTL/
   cmake -S . -B build -DCMAKE_INSTALL_PREFIX=$CSG_INSTALL_DIR
   cd build/
   make -j8 install
   popd
   

8. CSG EB library (gmake)

   For the gmake install instructions, you need to install libcsgeb to $CSG_LIB_DIR using either cmake or gmake:

   cmakegmake

   cd subprojects/csg-eb
   
   export CMAKE_PREFIX_PATH=$CMAKE_PREFIX_PATH:$CSG_INSTALL_DIR
   
   cmake -S . -B build -DCMAKE_INSTALL_PREFIX=$CSG_LIB_DIR \
                       -DCMAKE_BUILD_TYPE=Release
   
   cd build
   make -j8 install
   

9. Conduit

   git clone --recursive https://github.com/LLNL/conduit.git
   pushd conduit/
   git checkout v0.8.4
   mkdir build && cd build
   cmake -S ../src -DCMAKE_INSTALL_PREFIX=$ASCENT_INSTALL_DIR \
         -DENABLE_OPENMP=OFF \
         -DENABLE_MPI=ON \
         -DENABLE_CUDA=OFF \
         -DCMAKE_BUILD_TYPE=Release
   make -j8 install
   popd
   

10. Vtk-m

    git clone --branch master https://gitlab.kitware.com/vtk/vtk-m.git
    pushd vtk-m/
    git checkout v1.9.0
    mkdir build && cd build/
    cmake -S ../ -DCMAKE_INSTALL_PREFIX=$ASCENT_INSTALL_DIR \
          -DVTKm_ENABLE_OPENMP=OFF \
          -DVTKm_ENABLE_MPI=ON \
          -DVTKm_ENABLE_CUDA=OFF \
          -DVTKm_USE_64BIT_IDS=OFF \
          -DVTKm_USE_DOUBLE_PRECISION=ON \
          -DVTKm_USE_DEFAULT_TYPES_FOR_ASCENT=ON \
          -DVTKm_NO_DEPRECATED_VIRTUAL=ON \
          -DCMAKE_BUILD_TYPE=Release
    make -j8 install
    popd
    

11. Ascent

    git clone --recursive https://github.com/Alpine-DAV/ascent.git
    pushd ascent
    git checkout v0.9.0
    mkdir build && cd build/
    cmake -S ../src -DCMAKE_INSTALL_PREFIX=$ASCENT_INSTALL_DIR \
          -DCONDUIT_DIR=$ASCENT_INSTALL_DIR \
          -DVTKM_DIR=$ASCENT_INSTALL_DIR \
          -DENABLE_VTKH=ON \
          -DENABLE_FORTRAN=OFF \
          -DENABLE_PYTHON=OFF \
          -DENABLE_DOCS=OFF \
          -DBUILD_SHARED_LIBS=ON \
          -DCMAKE_BUILD_TYPE=Release \
          -DENABLE_GTEST=OFF \
          -DENABLE_TESTS=OFF
    make -j8 install
    popd
    

Running Jobs

Common Slurm commands:

• sinfo see available/allocated resources

• sbatch runit_cpu.sh submit a cpu job to the queue

• squeue -u USER check job status of user USER

• squeue -p PARTITION check job status of partition PARTITION

• scancel JOBID kill a job with id JOBID

• salloc –nodes 1 –qos interactive –time 01:00:00 –constraint gpu –gpus 4 –account=<ACCOUNT> grab 1 GPU node (for up to 1 hrs)

Example run scripts:

CPUGPU

#!/bin/bash
#SBATCH --job-name run-mfix-exa
#SBATCH -o stdout.%x-%j
#SBATCH -e stderr.%x-%j
#SBATCH --account=<ACCOUNT>
#SBATCH --qos=<QOS>
#SBATCH --constraint=cpu
#SBATCH --time=00:02:00
#SBATCH --nodes=2
#SBATCH --ntasks-per-node=16

# load modules used for build
module load PrgEnv-gnu/8.3.3
module load cmake/3.22.0

srun -n 32 --cpu-bind=cores -c 16 ./mfix inputs > screen.txt