Skip to content

Latest commit

 

History

History
177 lines (133 loc) · 5.56 KB

README-Spack.md

File metadata and controls

177 lines (133 loc) · 5.56 KB
Raw

Building ParFlow with Spack

This document covers building ParFlow with Spack. This feature is experimental and under development. Building with Spack may be useful in some situations since Spack builds a complete dependency tree and thus may have fewer compiling and linking issues. Spack is a multi-platform package manager that builds and installs multiple versions and configurations of software. Spack has support for Linux, macOS, and many supercomputers.

See Spack documentation for additional information on Spack and how to use it.

Obtaining and installing Spack

A special version of Spack is currently needed for a ParFlow build. There are only two small changes to Spack to address a Silo build issue and adding the rules for the ParFlow configuration. Once the ParFlow build is stable we will ask for ParFlow to be added to the standard Spack distribution.

The following will download Spack and setup the environment. The source command needs to be done in every new shell to set setup the environment.

For bash:

git clone https://github.com/parflow/spack.git
export SPACK_ROOT=$(pwd)/spack
source spack/share/spack/setup-env.sh

For tcsh/csh:

git clone https://github.com/parflow/spack.git
setenv SPACK_ROOT `pwd`/spack
source spack/share/spack/setup-env.csh

Building GNU compiler suite

This step is optional for Linux systems but is currently critical on Mac platforms. We have found issues with some releases of Clang producing executables that do not yield correct results. Clang is the compiler used by Apple XCode so running Parflow on MacOS shows this issues. We are currently investigating the issue, a work-around is to compile the GNU compiler suite under Spack and avoid using Clang. Builing GCC will take considerable amount of time.

spack install gcc@7.2.0 languages="fortran,c,c++"

Add the compiler to the set of compilers Spack can use:

spack compiler add $(spack location --install-dir gcc@7.2.0)

Make the Spack built GCC the default for compiling by adding the following to the Spack package config file (~/.spack/packages.yaml):

packages:
  all:
    compiler: [gcc@7.2.0]

Side note: We have found that Clang/LLVM versions in the 4.* release set exhibit large numerical differences in the results of a running Parflow. This occurs on both MacOS and Linux systems. The 3.* and 5.0.0 releases do not show this problem. Note MacOS versioning is different from the standard version of Clang. We are not certain that this is a Clang issue or some odd Parflow bug. We are leaning to a Clang problem since Parflow works under multiple other compilers and the issue seems to confined to a limited set of Clang releases. However that is a bit of a handy-wavy conclusion since we have not narrowed down the problem, we will continue to investigate as time permits.

Building ParFlow

First download, configure, and build ParFlow and all of the libraries ParFlow depends on. Spack does all of this for you! This step will take a considerable amount of time; be prepared to wait. You must be connected to the internet so Spack can download everything. You must have a compiler suite (C/C++/Fortran) installed.

spack install parflow@develop

Hopefully everything built correctly.

Running the Spack build of ParFlow

First setup some environment variables for running out of the Spack directories. The Spack command has a location option to determine the location of installs of TCL, OpenMPI and ParFlow. We use the "spack location" command to set some environment variables and then add some bin directories to the path. This will ensure that the Spack installations will be used to run ParFlow.

For bash:

export PARFLOW_DIR=$(spack location --install-dir parflow)
export PARFLOW_TCL_DIR=$(spack location --install-dir tcl)
export PARFLOW_MPI_DIR=$(spack location --install-dir openmpi)
export PATH=${PARFLOW_MPI_DIR}/bin:${PARFLOW_TCL_DIR}/bin:${PARFLOW_DIR}/bin:$PATH

# Are we using spack versions?
echo "Using PARFLOW_DIR=${PARFLOW_DIR}"
echo "Using tclsh : $(which tclsh)"
echo "Using mpirun : $(which mpirun)"

For tcsh/csh:

setenv PARFLOW_DIR `spack location --install-dir parflow`
setenv PARFLOW_TCL_DIR `spack location --install-dir tcl`
setenv PARFLOW_MPI_DIR `spack location --install-dir openmpi`
set path = ( ${PARFLOW_MPI_DIR}/bin ${PARFLOW_TCL_DIR}/bin ${PARFLOW_DIR}/bin $path )
setenv export PATH 

# Are we using spack versions?
echo "Using PARFLOW_DIR=${PARFLOW_DIR}"
echo "Using tclsh : `which tclsh`"
echo "Using mpirun : `which mpirun`"

We are looking at how to make this installation simpler; possibly using a Spack view.

Testing

Spack is managing the directories for the ParFlow build so we checkout the ParFlow repository to get a copy of the ParFlow regression tests so we can run them.

Clone the ParFlow repository:

git clone https://github.com/parflow/parflow.git

Run a few tests:

cd parflow/test

tclsh default_single.tcl 1 1 1

tclsh default_single.tcl 2 2 2

tclsh default_richards_with_netcdf.tcl 1 1 1

tclsh default_richards_with_netcdf.tcl 2 2 2

Hopefully these all pass!

To run the full test suite:

cd parflow/test
make test

Some of the tests that require a large number of MPI ranks may fail (The error message is something like "There are not enough slots available in the system to satisfy the 27 slots") if you don't have enough cores. The test suite goes up to 27 cores. We are working to fix this issue (for MPI folks, over-committing is not being enabled or working in the Spack OpenMPI setup).