Skip to content

Latest commit

 

History

History
249 lines (162 loc) · 16.3 KB

PARALLEL.md

File metadata and controls

249 lines (162 loc) · 16.3 KB

Developing in parallel MODFLOW 6

This document describes how to set up your build environment for developing and testing the parallel version of MODFLOW. It further builds on the instructions given in DEVELOPER.md so make sure to read those first.


DISCLAIMER

Expectations on platform compatibility

The serial version of the MODFLOW 6 program has had no external dependencies and is traditionally available for a variety of platforms (Windows, GNU/linux, macOS) and compatible with the mainstream Fortran compilers (gfortran, ifort). The parallel version comes with dependencies on third party components, most notably the MPI and PETSc libraries. While the goal is a continued support of the above mentioned configurations, this has become more challenging and can generally not be guaranteed. To assist developers as well as end users who are planning to compile the code themselves, a list of successfully tested build configurations will be included in this document.


Introduction

The design philosophy has been to maintain MODFLOW as a single codebase and have it compile to either a serial or a parallel program. The former continues to be a lightweight, highly compatible code which does not require external dependencies other than those provided by the standard compiler libraries. The latter has two distinct dependencies on 3rd party libraries: MPI and PETSc, as described below. Note that the parallel capability is a true extension and the executable will in all cases be capable of serial execution with equivalent results.

Prerequisites

The parallel version of MODFLOW 6 requires the Message Passing Interface (MPI) and the Portable, Extensible Toolkit for Scientific Computation (PETSc - pronounced PET-see (/ˈpɛt-siː/)) libraries.

MPI

The parallel version of MODFLOW 6 uses MPI to synchronize data between processes. There are a couple of implementations of the MPI standard available. Their applicability usually depends on the platform that is used:

On Linux and macOS, if you haven't installed a MPI framework yet, your best bet is to automatically download the MPI implementation if building PETSc from source using the option flag --download-openmpi or --download-mpich and the configure approach described below. Alternatively, you can install OpenMPI or MPICH using a package manager for your OS. Tested configurations are listed in the table below.

In addition to compiling, the MPI toolset is also required to run a parallel simulation. The implementations above all come with mpiexec (or mpiexec.exe) to start an executable in parallel.

PETSc

The PETSc library is a suite of data structures and routines for the scalable (parallel) solution of scientific applications modeled by partial differential equations:

https://petsc.org/release/

The PETSc library (version 3.16 or higher) is used by MODFLOW for its parallel linear solver capabilities and the distributed data formats (vectors and matrices) that go along with it. Parallel PETSc uses MPI internally as well, so setting up this library should typically be coordinated with the installation of the MPI library. A lot of obscure things can happen if the binaries are not compatible, so in general it is a good strategy to compile MPI, PETSc, and MODFLOW with the same compiler toolchain or install MPI and PETSc using a package manager and built with the same compiler being used to compile MODFLOW.

Compiling MPI and PETSC from source

The PETSc website gives details on a large number of configurations, depending on the target platform/OS, and many different ways to configure/make/install the library: https://petsc.org/release/install/. Building on Windows is notoriously challenging and discouraged by the PETSc development team. On Linux, however, PETSc can be installed (configure/make/install) by executing the following command

$ ./configure --download-openmpi --download-fblaslapack
$ make all

in a terminal open in the root directory of your PETSc download

Using a package manager to install MPI and PETSc

Use of a package manager can simplify the process of building the parallel version of MODFLOW 6.

MacOS

OpenMPI and PETSc are available on Homebrew for Intel and Apple Silicon (M1). Both of these depend on gcc 13.1.0. pkg-config should also be installed from Homebrew, if not already installed, so that Meson will be able to resolve the installation location of MPI and PETSc.

Ubuntu

OpenMPI and PETSc are available for a variety of Ubuntu versions using the Advanced Packaging Tool (apt).

Windows

Under evaluation.

Using pkg-config to check your PETSc installation

Eventually, the MODFLOW build process has to resolve the installation location of all external dependencies. The pkg-config tool (https://en.wikipedia.org/wiki/Pkg-config) can be used to take care of that.

pkg-config --libs petsc

If PETSc was build from source, you can check the contents of the folder

$PETSC_DIR/$PETSC_ARCH/lib/pkgconfig/

and confirm that there are one or more *.pc files in there. A similar pkgconfig/ folder has to present for the MPI installation that was used. For example, for Open MPI on WSL2 this folder is /lib/x86_64-linux-gnu/pkgconfig/.

To connect everything, both of these folder paths have to be added to the PKG_CONFIG_PATH variable so that the pkg-config executable can resolve the installed libraries.

Building the parallel version of MODFLOW 6

The primary build system for MODFLOW is Meson (https://mesonbuild.com/). The meson.build script takes an additional argument to activate a parallel build of the software. E.g for building and installing a parallel release version:

meson setup builddir -Ddebug=false -Dparallel=true \
                     --prefix=$(pwd) --libdir=bin
meson install -C builddir
meson test --verbose --no-rebuild -C builddir

Note that changing the option flags in the meson setup command requires the flag --reconfigure to reconfigure the build directory. If the PKG_CONFIG_PATH was set as described above, the linking to PETSc and MPI is done automatically.

It's always a good idea to check your parallel MODFLOW executable to confirm that it is successfully linked against the external dependencies. You can use the command line tools ldd (Linux), otool (macOS), or Dependencies.exe (Windows, https://github.com/lucasg/Dependencies) to do that. In the list of dependencies, you should be able to identify libpetsc and libmpi for parallel builds.

The other build systems in the MODFLOW project (MS Visual Studio, pymake, Makefile) continue to be supported for serial builds only. pymake uses the excludefiles.txt to ignore those files that can only be build when MPI and PETSc are present on the system. In MS Visual Studio these same files are included in the solution but not in the build process.


IMPORTANT

Don't use MPI and PETSc directly in your code

Parallel MODFLOW was designed to have all third party functionality (MPI and PETSc currently) made available through the framework. Developers of models and packages should not directly call these libraries and change the set of excluded files described above. If you feel you need to include MPI or PETSc functionality in your code (e.g. you want to use mpi in your source file), contact the MODFLOW development team on how to best proceed.


Testing the parallel of MODFLOW 6

Parallel MODFLOW can be tested using the same test framework as the serial program, with just a few modifications. To run a test inside the autotest folder in parallel mode, make sure to add a marker @pytest.mark.parallel so that the test is only executed in the Continuous Integration when running a configuration with a parallel build of MODFLOW.

The TestSimulation object that is being run from the framework should be configured for parallel run mode with the flag parallel=True and the number of processes ncpus=.... As an example, see the test case in autotest/test_par_gwf01.py, which can be run with

$ pytest -s --parallel test_par_gwf01.py

Running without the --parallel flag will simply skip the test.

Debugging

The most straightforward way to debug a parallel simulation is to start a run and have it pause to attach the debugger(s). Make sure that the MODFLOW executable was compiled with -Ddebug=true. In parallel mode the program uses the PETSc solver and a configuration file .petscrc should be present in the same folder as the simulation's mfsim.nam. In that PETSc resource file, you should add the following option:

-wait_dbg

telling MODFLOW to pause immediately after startup. This will give you time to attach one or multiple debuggers to the processes. Then start the parallel program, for example on two cores:

mpiexec -np 2 mf6 -p

In the process explorer you should now see 2 processes called mf6 or mf6.exe. On the prompt where the command was executed, MODFLOW waits for input:

$ Hit enter to continue...

In order to truly debug in parallel, i.e. step through the instructions side-by-side, you will need to start two instances of the debugger and attach them. The following section describes how to do that with VSCode.

Debugging with VSCode

In VSCode parallel debugging is easiest done by duplicating the development environment. First, make sure that you have set up your environment to build a parallel version in debug mode. The VSCode launch.json should contain an entry to attach to a running process:

  {
    "name": "Attach to ...",
    "type": "cppdbg",
    "request": "attach",
    "processId": "${command:pickProcess}",
    "program": "<path_to>/mf6",
    "MIMode": "gdb",
    "miDebuggerPath": "<path_to>/gdb",
    "setupCommands": [
      {
        "description": "Enable pretty-printing for gdb",
        "text": "-enable-pretty-printing",
        "ignoreFailures": true
      }
    ]
  }

After building parallel MODFLOW, press Ctrl+Shift+p to execute Workspaces: Duplicate As Workspace in New Window. This will open a second VSCode window, identical to the first. Starting the debug process and selecting "Attach to ..." opens a process selection window with the processes started from the mpiexec command described above. Select both, each from their own instance of the VSCode program. Now you can put breakpoints in the code, "Hit enter to continue" on the command prompt, and step through the parallel processes side-by-side.


TIP

Make sure that you work with gdb versions >= 10. We have found that earlier versions are only partially compatible with the VSCode debugging interface and crash when inspecting the data of Fortran derived types


Compatibility

Parallel MODFLOW has been built successfully with the following configurations:

Operating System Toolchain MPI PETSc Package Manager
MS Windows ? ? ? NA
WSL2 (Ubuntu 20.04.5) gcc 9.4.0 OpenMPI 4.0.3 3.18.2 NA
macOS 12.6.3 gcc 9.5.0 OpenMPI 4.1.4 3.18.5 NA
macOS 12.6.6 gcc 13.1.0 OpenMPI 4.1.5 3.19.1 Homebrew
Ubuntu 22.04 gcc 9.5.0 OpenMPI 4.1.4 3.18.5 NA
Ubuntu 22.04 ARM64 gcc 11.4.0 OpenMPI 4.1.5 3.19.3 apt
Ubuntu 22.04 ARM64 gcc 9.5.0 MPICH 3.4.1 3.15.5 NA
Ubuntu 22.04 ARM64 gcc 12.3.0 MPICH 4.1.1 3.19.6 NA
Ubuntu 22.04 ARM64 gcc 12.3.0 MPICH 4.1.1 3.20.0 NA
SUSE Linux Enterprise Server 15 SP2 intel 19.1.0.166 20191121 CRAY-MPICH 7.7.19 CRAY-PETSC 3.14.5.0 NA
Red Hat Enterprise Linux 8.7 intel 2021.10.0 20230609 CRAY-MPICH 8.1.26 3.15.5 NA

The most up-to-date configurations are available in the GitHub CI script: .github/workflows/ci.yml under the task parallel_test. These are being tested upon every change to the develop branch of MODFLOW.

To improve support, we kindly ask you to share your experience with building and running parallel MODFLOW and report back if you have a successful setup that is not in this table.

Known issues

Building PETSc on Ubuntu 22.04 with MPICH and GNU compilers

Versions of PETSc that use MPICH 3.4 (v3.14, v3.15, v3.16)) must be built with gcc-9 or earlier. Versions of PETSc that use MPICH 4.1 (v3.17 or newer) can be built with newer versions of the gcc compiler (gcc-11, gcc-12, etc.).

Meson does not correctly load the Fortran compiler flags from the mpich.pc package configuration file in the $PETSC_DIR/$PETSC_ARCH/lib/pkgconfig directory. To overcome this issue, make a copy of mpich.pc and name it mpichfort.pc. Then determine the appropriate Fortran flags using

$PETSC_DIR/$PETSC_ARCH/bin/mpifort -show

which will return something like

$ linux-real-gcc12.3.0-3.20.0/bin/mpifort -show
gfortran -fPIC -ffree-line-length-none -ffree-line-length-0 -Wno-lto-type-mismatch -O2 -fallow-argument-mismatch -I/media/psf/Development/petsc/linux-real-gcc12.3.0-3.20.0/include -I/media/psf/Development/petsc/linux-real-gcc12.3.0-3.20.0/include -L/media/psf/Development/petsc/linux-real-gcc12.3.0-3.20.0/lib -lmpifort -Wl,-rpath -Wl,/media/psf/Development/petsc/linux-real-gcc12.3.0-3.20.0/lib -Wl,--enable-new-dtags -lmpi

Copy the returned Fortran flags and replace the Libs: and Cflags: attributes in the mpichfort.pc file. Also modify the Name: attribute to mpichfort. The modified mpichfort.pc file should look something like

# this gives access to the mpich header files
prefix=/media/psf/Development/petsc/linux-real-gcc9.5.0-3.20.0
exec_prefix=${prefix}
libdir=/media/psf/Development/petsc/linux-real-gcc9.5.0-3.20.0/lib
includedir=${prefix}/include

Name: mpichfort
Description: High Performance and portable MPI
Version: 4.1.2
URL: http://www.mcs.anl.gov/research/projects/mpich
Requires:
Libs: -fPIC -ffree-line-length-none -ffree-line-length-0 -Wno-lto-type-mismatch -O2 -fallow-argument-mismatch -I/media/psf/Development/petsc/linux-real-gcc12.3.0-3.20.0/include -I/media/psf/Development/petsc/linux-real-gcc12.3.0-3.20.0/include -L/media/psf/Development/petsc/linux-real-gcc12.3.0-3.20.0/lib -lmpifort -Wl,-rpath -Wl,/media/psf/Development/petsc/linux-real-gcc12.3.0-3.20.0/lib -Wl,--enable-new-dtags -lmpi 
Cflags:  -fPIC -ffree-line-length-none -ffree-line-length-0 -Wno-lto-type-mismatch -O2 -fallow-argument-mismatch -I/media/psf/Development/petsc/linux-real-gcc12.3.0-3.20.0/include -I/media/psf/Development/petsc/linux-real-gcc12.3.0-3.20.0/include -L/media/psf/Development/petsc/linux-real-gcc12.3.0-3.20.0/lib -lmpifort -Wl,-rpath -Wl,/media/psf/Development/petsc/linux-real-gcc12.3.0-3.20.0/lib -Wl,--enable-new-dtags -lmpi

# pkg-config does not understand Cxxflags, etc. So we allow users to
# query them using the --variable option

cxxflags= -Wno-lto-type-mismatch -Wno-psabi -O2 -std=gnu++17 -fPIC -I${includedir}
fflags=-fPIC -ffree-line-length-none -ffree-line-length-0 -Wno-lto-type-mismatch -O2 -I${includedir}
fcflags=-fPIC -ffree-line-length-none -ffree-line-length-0 -Wno-lto-type-mismatch -O2 -I${includedir}

The /media/psf/Development/petsc/linux-real-gcc12.3.0-3.20.0/include and /media/psf/Development/petsc/linux-real-gcc12.3.0-3.20.0/lib entries in the Libs: and Cflags: attributes can be replaced with ${includedir} and ${libdir}, respectively, to simplify mpichfort.pc.