Skip to content

Latest commit

 

History

History
366 lines (277 loc) · 12.3 KB

installation.md

File metadata and controls

366 lines (277 loc) · 12.3 KB
Raw

Installation

PetIBM only officially supports Linux systems. MacOS and Windows may or may not work. We will not provide any assist to issues of using PetIBM on MacOS and Windows.

A. Installation guide for end-users

End-users can install PetIBM with Anaconda or use it with container technologies.

A.1. With conda/mamba

We uploaded our pre-built packages to our Anaconda channels barbagroup. These packages were built with GPU support. Linux users can install them through

$ mamba install -c barbagroup -c nvidia -c conda-forge "petibm=*=cuda114*"

We suggest using mamba. If conda is preferred, just replace mamba with conda. Replace install with create -n <env name> instead to create a new and fresh environment.

If an older CUDA is desired (e.g., for old GPU models), we also have packages for CUDA 10.2, which supports old GPUs like K20/K40. To do so, use "petibm=*=cuda102*" instead.

Note: The CUDA 10.2 variant may also work on newer GPUs and drivers. If you encounter issues using the CUDA 11.4 variant with newer GPUs/drivers, it should be fine to use the CUDA 10.2 variant. AmgX uses somehow different code w.r.t. whether it's CUDA 10 or CUDA 11, and sometimes there are issues in AmgX's CUDA 11 codebase.

A.2. Using container technology

For users familiar with Docker or Singularity Container (now renamed Apptainer), we provide images for the both to save you from compiling, buliding, and installing:

A.3. Optional dependencies for example cases' post-processing

The above two installation methods only provide the solver. Some example cases in the examples folder have Python scripts for post-processing. To be able to execute these Python scripts, install the following Python packages (using mamba/conda, for example):

$ mamba install -c conda-forge pyyaml numpy h5py matplotlib

B. Building from the source

B.1. Using pre-built dependencies

If you just need to set up a development environment to extend/debug PetIBM, we suggest installing dependencies with conda/mamba:

$ mamba create \
    -n <env name> \
    -c barbagroup \
    -c nvidia \
    -c conda-forge \
    "*amgxwrapper=*=cuda114*" \
    "yaml-cpp>=0.7" \
    "symengine>=0.9" \
    "cmake>=3.23" \
    "make" \
    "pkg-config" \
    "git" \
    "gxx_linux-64>=11.2"

The package amgxwrapper pulls in most of the required dependencies (e.g., petsc, cuda, etc.). yaml-cpp and symengine are dependencies not covered by amgxwrapper. cmake, make, git, and gxx_linux-64 are for building PetIBM in case your system does not have them or they are too old.

Once the process is done, jump into the environment through either mamba activate <env name> or source <mamba installation prefix>/bin/activate <env name>.

The above command only installs required packages for the flow solvers. See section A.3 for dependencies for post-processing with example cases.

To proceed, let's assume you would like to download PetIBM's source files to path $HOME/sfw/petibm/PetIBM, and to install PetIBM's final solvers and libraries to $HOME/sfw/petibm/petibm-linux-dbg.

B.1.1. Obtaining the source

Pull the latest development version of PetIBM from GitHub:

$ cd $HOME/sfw
$ mkdir petibm
$ cd petibm
$ git clone https://github.com/barbagroup/PetIBM.git

PetIBM's source files are now at $HOME/sfw/petibm/PetIBM. Our current location is at $HOME/sfw/petibm.

B.1.2. Configuring

Create a folder to host temporary files from building and compilation:

$ mkdir build
$ cd build

Configure the build:

$ cmake \
    -DCMAKE_INSTALL_PREFIX=$HOME/sfw/petibm/petibm-linux-dbg \
    -DCMAKE_CXX_COMPILER=${CXX} \
    -DCMAKE_BUILD_TYPE=Debug \
    -DPETSC_DIR=$CONDA_PREFIX \
    -DPETSC_ARCH="" \
    -DCUDA_DIR=$CONDA_PREFIX \
    -DAMGX_DIR=$CONDA_PREFIX \
    -DAMGXWRAPPER_DIR=$CONDA_PREFIX \
    -DYAMLCPP_DIR=$CONDA_PREFIX \
    -DSYMENGINE_DIR=$CONDA_PREFIX \
    -DPETIBM_ENABLE_TESTS=ON \
    -DPETIBM_USE_AMGX=ON \
    $HOME/sfw/petibm/PetIBM
  • PETSc, CUDA, AmgX, AmgXWrapper, and yaml-cpp were all installed through mamba (or conda), so their paths are $CONDA_PREFIX.
  • -DPETIBM_ENABLE_TESTS controls whether to build unit tests.
  • -DPETIBM_USE_AMGX notifies PetIBM to build against GPU-related code.

Later on when you want to have a production build, just change -DCMAKE_BUILD_TYPE from Debug to Release. (And you may also want to change the installation path in -DCMAKE_INSTALL_PREFIX.)

B.1.3. Building and installing

Start to build:

$ make all -j <number of CPUs>

Specifying the number of CPUs with <number of CPUS> to speed up the compilation procedure.

We can do some unit tests:

$ make check

Then install the final solvers and libraries:

$ make install

The solvers are now available at $HOME/sfw/petibm/petibm-linux-dbg/bin.

B.2. Building dependencies from scratch

If you would like to build everything from scratch, here's the list of dependencies:

Required:

  • GNU C++ compiler g++ (9.4 and 11.2 have been test with the latest PetIBM release)
  • PETSc (3.16+) with parallel HDF5 enabled
  • MPI: OpenMPI, MPICH, or Intel MPI
  • yaml-cpp (0.7.0+)
  • SymEngine (0.9+)

Optional for GPU linear solvers:

Optional for pre- and post-processing Python scripts:

  • Python (3.6+)
  • NumPy (1.12.1+)
  • h5py (2.7.0+)
  • Matplotlib (2.0.2+)
  • pyyaml (6.0+)

Note:

  • yaml-cpp and AmgXWrapper can be automatically installed during PetIBM configuration.
  • MPI can be either installed during PETSc configuration or installed explicitly by users. If you are using HPC clusters, it's better to use the shipped MPI packages on the clusters because they usually have specific optimizations for the clusters' hardware. NVIDIA's HPC-X is another alternative, which includes OpenMPI optimized for NVIDIA's HPC infrastructure.
  • The latest release of PetIBM (0.5.4) has been tested with:
    • Arch Linux with g++-11.2.0 and PETSc v3.16.6
    • Ubuntu 20.04 with g++-9.4 and PETSc v3.16.6
  • Though not suggested, if using older compilers are desired, previous releases of PetIBM (before v0.5.2) have at some point been tested with the following combinations. However, we will not provide any support for using older versions of PetIBM.
    • Ubuntu 16.04 with g++ v5.4, and PETSc v3.11.2
    • MacOS Sierra with g++ v6.0, and PETSc v3.8.2
    • Arch Linux with g++ v7.2, and PETSc v3.8.2

B.2.1. Compilers and configuration tools

Different linux distributions use different package management systems. Here we use Ubuntu 20.04 as an example. Assuming you have permission to use sudo, to install compilers and tools:

$ sudo apt-get install build-essential pkg-config make git curl

You can check the version of G++ installed with g++ --version.

Now we also need cmake. Unfortunately, cmake from Ubuntu 20.04 is too old. We need to install it from upstream directly:

$ curl -LO https://github.com/Kitware/CMake/releases/download/v3.23.1/cmake-3.23.1-linux-x86_64.sh
$ sudo sh ./cmake-3.23.1-linux-x86_64.sh --prefix=/usr --exclude-subdir --skip-license

This will install the cmake v3.23 to /usr as if it's a package from the package manager.

B.2.2. Building SymEngine

SymEngine requires some other third-party dependencies. The building process is not trivial. Hence we are not covering how to build and install SymEngine. Some Linux distributions officially provide pre-built SymEngine. For example, in Arch Linux, the pre-built SymEngine can be installed with sudo pacman -S symengine. Check with your Linux distribution's package manager first.

If building SymEngine from scratch is needed, please refer to SymEngine's documentation here.

B.2.3. Building PETSc

Here, we provide the command-line instructions to install PETSc v3.16.6. PETSc is powerful and has a lot of options to configure and for different levels of optimizations. We only show the bare minimum configuration that at least work for PetIBM.

Get and unpack PETSc:

$ cd $HOME/sfw
$ mkdir -p petsc/3.16.6
$ curl -LO http://ftp.mcs.anl.gov/pub/petsc/release-snapshots/petsc-lite-3.16.6.tar.gz
$ tar -xvf petsc-lite-3.16.6.tar.gz -C petsc/3.16.6 --strip-components=1
$ cd petsc/3.16.6

Configure and build PETSc in debugging mode:

$ export PETSC_DIR=$HOME/sfw/petsc/3.16.6
$ export PETSC_ARCH=linux-dbg
$ ./configure --PETSC_DIR=$PETSC_DIR --PETSC_ARCH=$PETSC_ARCH \
    --with-cc=gcc \
    --with-cxx=g++ \
    --with-fc=0 \
    --with-pic=1 \
    --with-shared-libraries=1 \
    --with-debugging=1 \
    --with-fortran-bindings=0 \
    --download-f2cblaslapack \
    --download-openmpi \
    --download-hdf5 \
    --download-hypre \
    --download-ptscotch \
    --download-metis \
    --download-parmetis \
    --download-superlu_dist
$ make all
$ make test

To configure and build PETSc for production runs, change the value of --with-debugging form 1 to 0, and add --COPTFLAGS="-O3" --CXXOPTFLAGS="-O3" to the list of flags. You may also want to change the name in PETSC_ARCH.

The above configuration also downloads and installs OpenMPI. However, it is not optimized. When running the code on an external cluster, we suggest configure and build PETSc with a proper MPI package for performance.

To understand and use a more advanced configuration of PETSc, please refer to PETSc's instructions. Also, you can run ./configure --help to list all available configure flags.

B.2.3. Building PetIBM

Here we will build a CPU-only PetIBM with yaml-cpp being installed automatically.

Pull the latest development version of PetIBM from GitHub:

$ cd $HOME/sfw
$ mkdir petibm
$ cd petibm
$ git clone https://github.com/barbagroup/PetIBM.git

Configure and build PetIBM:

$ export PETSC_DIR=$HOME/sfw/petsc/3.16.6
$ export PETSC_ARCH=linux-dbg
$ cmake \
    -DCMAKE_INSTALL_PREFIX=$HOME/sfw/petibm/petibm-linux-dbg \
    -DCMAKE_CXX_COMPILER=$PETSC_DIR/$PETSC_ARCH/bin/mpicxx \
    -DCMAKE_BUILD_TYPE=Debug \
    -DPETSC_DIR=$PETSC_DIR \
    -DPETSC_ARCH=$PETSC_ARCH \
    -DSYMENGINE_DIR=<path to SymEngine> \
    -DPETIBM_ENABLE_TESTS=ON \
    -DPETIBM_USE_AMGX=OFF \
    -DPETIBM_BUILD_YAMLCPP=ON \
    $HOME/sfw/petibm/PetIBM
$ make all
$ make check -j <num CPUs>
$ make install

For production runs, change the value of -DCMAKE_BUILD_TYPE to Release and refer to the corresponding PETSc builds in -DPETSC_DIR and -DPETSC_ARCH.

Et voila! Now PetIBM is installed to $HOME/sfw/petibm/petibm-linux-dbg.

B.2.4. Optional: using NVIDIA AmgX to solve linear systems on multiple GPUs

PetIBM's GPU support comes from NVIDIA's AmgX. And the interface between PETSc and AmgX is provided by AmgXWrapper.

Building AmgX is an expensive job. Without a powerful machine, building AmgX may take hours. Hence we would like to refer interested readers to AmgX's manual for how to build and install it.

Once AmgX is available on your machine, AmgXWrapper can be automatically downloaded and installed during installing PetBM:

$ export PETSC_DIR=$HOME/sfw/petsc/3.16.6
$ export PETSC_ARCH=linux-dbg
$ cmake \
    -DCMAKE_INSTALL_PREFIX=$HOME/sfw/petibm/petibm-linux-dbg \
    -DCMAKE_CXX_COMPILER=$PETSC_DIR/$PETSC_ARCH/bin/mpicxx \
    -DCMAKE_BUILD_TYPE=Debug \
    -DPETSC_DIR=$PETSC_DIR \
    -DPETSC_ARCH=$PETSC_ARCH \
    -DCUDA_DIR=<path to cuda> \
    -DAMGX_DIR=<path to amgx> \
    -DPETIBM_ENABLE_TESTS=ON \
    -DPETIBM_USE_AMGX=ON \
    -DPETIBM_BUILD_YAMLCPP=ON \
    -DPETIBM_BUILD_AMGXWRAPPER=ON \
    $HOME/sfw/petibm/PetIBM
$ make all
$ make check -j <num CPUs>
$ make install

C. PetIBM examples

Examples files are installed at <PetIBM installation>/share/petibm/examples. For example, for installation with conda/mamba, the <PetIBM installation> is $CONDA_PREFIX.