Machines

Polaris attempts to be aware of the capabilities of the machine it is running on. This is a particular advantage for so-called “supported” machines with a config file defined for them in the polaris package. But even for “unknown” machines, it is not difficult to set a few config options in your user config file to describe your machine. Then, polaris can use this data to make sure test cases are configured in a way that is appropriate for your machine.

Supported Machines

If you follow the procedure in polaris conda environment, spack environment, compilers and system modules, you will have an activation script for activating the development conda environment, setting loading system modules and setting environment variables so you can build MPAS and work with polaris. Just source the script that should appear in the base of your polaris branch, e.g.:

source load_dev_polaris_0.1.0-alpha.1_anvil_intel_impi.sh

After loading this environment, you can set up tasks or suites, and a link load_polaris_env.sh will be included in each suite or task work directory. This is a link to the activation script that you sourced when you were setting things up. You can can source this file on a compute node (e.g. in a job script) to get the right polaris conda environment, compilers, MPI libraries and environment variables for running polaris tasks and the MPAS model.

Note

Albany (and therefore most of the functionality in MALI) is currently only supported for those configurations with gnu compilers.

Machine

Compiler

MPI lib.

MPAS make target

anvil

intel

impi

intel-mpi

openmpi

ifort

gnu

openmpi

gfortran

chicoma-cpu

gnu

mpich

gnu-cray

chrysalis

intel

openmpi

ifort

gnu

openmpi

gfortran

compy

intel

impi

intel-mpi

frontier

gnu

mpich

gnu-cray

crayclang

mpich

cray-cray

pm-cpu

gnu

mpich

gnu-cray

intel

mpich

intel-cray

Below are specifics for each supported machine

Other Machines

If you are working on an “unknown” machine, the procedure is pretty similar to what was described in polaris conda environment, spack environment, compilers and system modules. The main difference is that we will use mpich or openmpi and the gnu compilers from conda-forge rather than system compilers. To create a development conda environment and an activation script for it, on Linux, run:

./configure_polaris_envs.py --conda <conda_path> -c gnu -i mpich

and on OSX run:

./configure_polaris_envs.py --conda <conda_path> -c clang -i mpich

You may use openmpi instead of mpich but we have had better experiences with the latter.

The result should be an activation script load_dev_polaris_0.1.0-alpha.1_<mpi>.sh. Source this script to get the appropriate conda environment and environment variables.

Under Linux, you can build the MPAS model with

make gfortran

Under OSX, you can build the MPAS model with

make gfortran-clang

Adding a New Supported Machine

If you want to add a new supported machine, you

Adding a Machine Config File

The first step in adding a new supported machine to add a config file in polaris/machines. The config file needs to describe the parallel environment and some paths where shared Spack environments will be installed and shared data will be downloaded. The easiest place to start is one of the examples provided (machine morpheus for now, but more will be added soon.)

# The parallel section describes options related to running jobs in parallel
[parallel]

# parallel system of execution: slurm, cobalt or single_node
system = single_node

# whether to use mpirun or srun to run a task
parallel_executable = mpirun

# cores per node on the machine
cores_per_node = 8


# Config options related to spack environments
[spack]

# whether to load modules from the spack yaml file before loading the spack
# environment
modules_before = False

# whether to load modules from the spack yaml file after loading the spack
# environment
modules_after = False


# The paths section describes paths that are used within the ocean core test
# cases.
[paths]

# A shared root directory where MPAS standalone data can be found
database_root = /home/xylar/data/mpas/mpas_standalonedata

# the path to the base conda environment where polaris environments have
# been created
polaris_envs = /home/xylar/data/mpas/polaris_envs


# Options related to deploying a polaris conda environment on supported
# machines
[deploy]

# the compiler set to use for system libraries and MPAS builds
compiler = gnu

# the system MPI library to use for gnu compiler
mpi_gnu = openmpi

# the base path for spack environments used by polaris
spack = /home/xylar/data/mpas/spack

# whether to use the same modules for hdf5, netcdf-c, netcdf-fortran and
# pnetcdf as E3SM (spack modules are used otherwise)
use_e3sm_hdf5_netcdf = False


# Options related to machine discovery
[discovery]

# a substring used to identify this machine from its hostname
hostname_contains = morpheus

The [parallel] section should describe the type of parallel queuing system (currently only slurm or single_node are supported), the number of cores per node and the command for running an MPI executable (typically srun for Slurm and mpirun for a “single node” machine like a laptop or workstation.

The [spack] section has some config options to do with loading system modules before or after loading a Spack environment. On a “single node” machine, you typically don’t have modules so both modules_before and modules_after can be set to False. On a high-performance computing (HPC) machine, you may find it is safest to load modules after the Spack environment to ensure that certain paths and environment variables are set the way the modules have them, rather than the way that Spack would have them. The recommended starting point would be modules_before = False and modules_after = True, but could be adjusted as needed if the right shared libraries aren’t being found when you try to build an MPAS component.

In the [paths] section, you will first give a path where you would like to store shared data files used in polaris tasks in database_root. Polaris will create this directory if it doesn’t exist. Then, you can specify polaris_envs as a path where shared conda environments will be installed for polaris releases. If developers always create their own conda environments, this path will never be used.

In [deploy], you will specify config options used in setting up conda and Spack environments for developers. The compiler is the default compiler to use for your system. You must supply a corresponding mpi_<compiler> for each supported compiler (not just the default compiler) that specifies the default MPI library for that compiler. If you only support one compiler and MPI library, that’s pretty simple: compiler is the name of the compiler (e.g. intel or gnu) and mpi_<compiler> is the MPI library (e.g. compiler_gnu = mpich or compiler_intel = openmpi). The spack option specifies a path where Spack environment will be created. The option use_e3sm_hdf5_netcdf = False indicates that you will not use the E3SM default modules for HDF5 and NetCDF libraries (which are not available for machines installed in the way described here).

Finally, [discovery] allows you to add a hostname_contains that is used to automatically identify your machine based on its hostname. If your machine has multiple login nodes with different hostnames, hopefully, a string common to all login nodes can be used here. If your machine has a unique hostname, simply give that. This option saves developers from having to specify --machine <machine> each time they setup polaris environments or test cases.

Describing a Spack Environment

The next step is to create a template YAML file that can be used to create Spack environments for your machine. Polaris uses Spack environments to build packages that need MPI support or which should be build for some other reason with system compilers rather than coming from pre-built conda packages. Using a Spack environment allows these packages to be built together in a consistent way that is not guaranteed if you try to install dependencies one-by-one. In Spack parlance, this is known as unified concretization.

To do this, you will create a file deploy/spack/<machine>_<compiler>_<mpi>.yaml similar to the following example for an Ubuntu laptop:

spack:
  specs:
  - gcc
  - openmpi
{{ specs }}
  concretizer:
    unify: true
  packages:
    all:
      compiler: [gcc@11.3.0]
      providers:
        mpi: [openmpi]
    curl:
      externals:
      - spec:  curl@7.81.0
        prefix: /usr
      buildable: false
    gcc:
      externals:
      - spec: gcc@11.3.0
        prefix: /usr
      buildable: false
  config:
    install_missing_compilers: false
  compilers:
  - compiler:
      spec: gcc@11.3.0
      paths:
        cc: /usr/bin/gcc
        cxx: /usr/bin/g++
        f77: /usr/bin/gfortran
        fc: /usr/bin/gfortran
      flags: {}
      operating_system: ubuntu22.04
      target: x86_64
      modules: []
      environment: {}
      extra_rpaths: []

Typically your system will already have compilers if nothing else, and this is what we assume here. Give the appropriate path (replace /usr with the appropriate path on your system). We have had better luck with gcc than other compilers like Intel so far so for new supported machines so that’s our recommendation. Use gcc --version to determine the version and replace 11.3.0 with this number.

Finally, you might need to update the target and operating_system. This is a bit of a “catch 22” in that you can use Spack to find this out but polaris is designed to clone and set up Spack for you so we assume you don’t have it yet. For now, make your best guess using the info on this page and correct it later if necessary.

You may need to load a system module to get the compilers and potentially other libraries such as MPI, HDF5, and NetCDF-C if you prefer to use system modules rather than having Spack build them. If this is the case, the best way to do this is to add a file conda/spack/<machine>_<compiler>_<mpi>.sh along these lines:

module purge
module load perl/5.32.0-bsnc6lt
module load gcc/9.2.0-ugetvbp
module load openmpi/4.1.3-sxfyy4k
module load intel-mkl/2020.4.304-n3b5fye
module load hdf5/1.10.7-j3zxncu
module load netcdf-c/4.4.1-7ohuiwq
module load netcdf-fortran/4.4.4-k2zu3y5
module load parallel-netcdf/1.11.0-mirrcz7

These modules will be loaded either before or after the spack environment, depending on the modules_before and modules_after config options above. You can also add modules in your YAML file but this shouldn’t be necessary.

For examples from various supported machines, compilers and MPI libraries, see the mache spack directory.

Building the Spack Environment

The next step is to try setting up polaris and asking it to build the Spack environment with a command something like:

  ./configure_polaris_envs.py --verbose --update_spack --conda <conda_path> -c gnu -i openmpi ...

The --update_spack flag tells polaris to create (or update) a Spack environment. You can specify a directory for testing Spack with the --spack flag. You can specify a temporary directory for building spack packages with --tmpdir (this directory must already exist). This is useful if your /tmp space is small (Spack will use several GB of temporary space).

Creating the Spack environment may take anywhere from minutes to hours, depending on your system.

If dependencies don’t build as expected, you may get an error message suggesting that your operating_system or target aren’t right. Here’s an example:

==> Error: concretization failed for the following reasons:
   1. zlib compiler '%gcc@9.4.0' incompatible with 'os=ubuntu20.04'
   2. readline compiler '%gcc@9.4.0' incompatible with 'os=ubuntu20.04'
   3. pkgconf compiler '%gcc@9.4.0' incompatible with 'os=ubuntu20.04'

In this example, I had specified operating_system: ubuntu22.04 in the YAML file but in fact my operating system is ubuntu20.04 as shown in the error message.

You can run:

source $SPACKDIR/share/spack/setup-env.sh
spack arch -o
spack arch -g

where $SPACKDIR is the directory where the Spack repository was cloned by polaris (you should see Cloning into <$SPACKDIR> in the terminal, which will hopefully help you find the right directory). This should hopefully give you something close to what Spack wants. If you get something like x86_64_v4 for the target, use x86_64 instead.

If you are getting other error messages, do your best to debug them but also feel free to get in touch with the polaris development team and we’ll help if we can.

If you get everything working well, please feel free to make a pull request into the polaris main repo to add your supported machine.