ratel-runner 0.2.2

Utilities for building and running the Ratel finite element library

Ratel Runner

CLI tools for building and running Ratel experiments.

Questions, comments, or concerns? Contact Zach Atkins or leave an issue.

Installation

UV (Recommended)

First, install uv, an open-source Python package manager written in Rust. Installing uv does not require root privileges and takes only a few seconds.

Then, you can either run ratel-runner without installing or install it using uv:

uvx ratel-runner
uv tool install 'ratel-runner@latest'

If you want to run iMPM experiments, specify the [mpm] optional dependency:

uvx ratel-runner
uv tool install 'ratel-runner[mpm]@latest'

Virtual Environment

If building on Tioga or Tuolumne, ensure you have a new enough Python version:

ml +cray-python

To install the python package, run:

pip install --user --upgrade ratel-runner

On Lassen, you should first make a virtual environment and ensure new enough compilers are set for building numpy:

ml +python/3.11.5
ml +base-gcc/11.2.1

# create virtual environment
python -m virtualenv .venv
# activate virtual environment
. .venv/bin/activate
# install ratel-runner
CC=gcc CXX=g++ pip install --upgrade git+https://github.com/zatkins-dev/Ratel-iMPM-Press.git

If you want to run implicit MPM experiments, you will need to use the --with gmsh flag to ensure the dependency is installed:

pip install --with gmsh --user --upgrade git+https://github.com/zatkins-dev/Ratel-iMPM-Press.git

Note, the gmsh package is unavailable on Lassen.

Building Ratel

Supported Machines

This package supports automatically building Ratel and its dependencies with optimal configurations on:

• Tioga
• Tuolumne
• Lassen

If building on Tioga or Tuolumne, ensure you have a new enough Python version:

ml +cray-python

For Lassen, instead use

ml +python/3.11.5

Tioga

If you are building on Tioga, add these commands to your ~/.bashrc or ~/.zshrc file and ensure they are run before building or acquiring a debug node:

if [[ "$(hostname)" == "tioga"* ]]; then
	module reset
	ml +rocmcc/6.4.0-cce-19.0.0d-magic
	ml +rocm/6.4.0
	ml +craype-accel-amd-gfx90a
	ml +cray-python
	ml +cray-libsci_acc
	ml +cray-hdf5-parallel/1.14.3.5
	ml +flux_wrappers
	ml +cray-mpich/8.1.32
	export HSA_XNACK=1
	export MPICH_GPU_SUPPORT_ENABLED=1
fi

ALWAYS build on a debug node. For Tioga, you can get such a node with the command:

flux alloc --queue=pdebug --setattr=thp=always -x -N1 -n1 -t 1h

For Tioga, the scratch directory defaults to the lustre2 parallel filesystem:

/p/lustre2/$USER/ratel-scratch

Tuolumne

If you are building on Tuolumne, add these commands to your ~/.bashrc or ~/.zshrc file and ensure they are run before building or acquiring a debug node:

if [[ "$(hostname)" == "tuolumne"* ]]; then
	module reset
	ml +rocmcc/6.4.0-cce-19.0.0d-magic
	ml +rocm/6.4.0
	ml +craype-accel-amd-gfx942
	ml +cray-python
	ml +cray-libsci_acc
	ml +cray-hdf5-parallel/1.14.3.5
	ml +flux_wrappers
	ml +cray-mpich/8.1.32
	export HSA_XNACK=1
	export MPICH_GPU_SUPPORT_ENABLED=1
	export MPICH_SMP_SINGLE_COPY_MODE=XPMEM
fi

ALWAYS build on a debug node. For Tuolumne, you can get such a node with the command:

flux alloc --queue=pdebug --setattr=thp=always --setattr=hugepages=512GB -x -N1 -n1 -t 1h

For Tuolumne, the scratch directory defaults to the lustre5 parallel filesystem:

/p/lustre5/$USER/ratel-scratch

Lassen

if [[ "$(hostname)" == "lassen"* ]]; then
	ml +clang/ibm-18.1.8-cuda-11.8.0-gcc-11.2.1
	ml +cuda/11.8.0
	ml +base-gcc/11.2.1
	ml +essl
	ml +lapack
	ml +python/3.11.5
fi

ALWAYS build on a debug node. For Lassen, you can get such a node with the command:

lalloc 1

For Lassen, the scratch directory defaults to the gpfs1 parallel filesystem:

/p/gpfs1/$USER/ratel-scratch

General Build instructions

Set an appropriate SCRATCH_DIR and OUTPUT_DIR, e.g.

# on supported machines, defaults to /parallel/filesystem/path/$USER/ratel-scratch
ratel-runner config set SCRATCH_DIR /p/lustre5/$USER/ratel-scratch
# typically defaults to the directory where commands are run
ratel-runner config set OUTPUT_DIR /usr/workspace/$USER/ratel-runner

If you are running on a supported machine, these configuration variables are optional. If you are building on an unsupported machine, you must also set PETSC_CONFIG to the path to a Python PETSc configuration script:

ratel-runner config set PETSC_CONFIG /path/to/reconfigure.py

Examples can be found in the PETSc repository.

If you are building on a machine with job scheduling, you should now acquire an interactive allocation, see Supported Machines for examples.

Then, Ratel and its dependencies can be built via:

ratel-runner build ratel

Configuration Variables

The following configuration variables are used to build Ratel and run experiments. The preferred way to set configuration variables is through the ratel-runner config command.

> ratel-runner config --help

 Usage: ratel-runner config [OPTIONS] COMMAND [ARGS]...

 Read/write values in the application configuration file.

╭─ Options ─────────────────────────────────────────────────────────────╮
│ --machine        [tuolumne|tioga|default]  [default: None]            │
│ --help                                     Show this message and      │
│                                            exit.                      │
╰───────────────────────────────────────────────────────────────────────╯
╭─ Commands ────────────────────────────────────────────────────────────╮
│ unset   Remove a key from the configuration file.                     │
│ set     Set a key-value pair in the configuration file.               │
│ get     Get the value of a key in the configuration file.             │
│ list    List all keys and values in the configuration file.           │
│ copy    Copy all configuration variables from one machine to another. │
╰───────────────────────────────────────────────────────────────────────╯

Alternatively, you can set the variables as environmental variables.

The list of relevant variables is given below.

Variable	Description	Default
SCRATCH_DIR	Location to clone and build repositories, store output files from experiments, etc. This should be on a parallel filesystem for most supercomputers.	See Supported Machines
OUTPUT_DIR	Location in which symbolic links to experiment result directories will be created.	Current runtime directory.
PETSC_DIR	Location of cloned PETSc repository. This can be an existing repository, or PETSc will be cloned to this directory if it does not exist.	$SCRATCH_DIR/build/petsc
PETSC_ARCH	PETSc arch/build to use.	Machine-dependent
PETSC_CONFIG	Python configuration file to use when building PETSc.	Machine-dependent
LIBCEED_DIR	Location of cloned libCEED repository. This can be an existing repository, or libCEED will be cloned to this directory if it does not exist.	$SCRATCH_DIR/build/libCEED
RATEL_DIR	Location of cloned Ratel repository. This can be an existing repository, or Ratel will be cloned to this directory if it does not exist.	$SCRATCH_DIR/build/ratel

Running Implicit Material Point Method (iMPM) Experiments

Experiments are run through the ratel-runner press command, use the help flag for a list of options.

ratel-runner mpm press --help

Press - Sticky Air

The "sticky-air" experiment models voids as a soft, perfectly compressible solid. The experiment requires the path to the voxel data file, the characteristic length (in mm), and the desired load fraction. See the help pages for the run and flux-run subcommands for other options:

ratel-runner mpm press sticky-air run --help
ratel-runner mpm press sticky-air flux-run --help

Note: The flux-run subcommand will only launch a batch job after the background mesh is generated. The background mesh generation may be quite expensive for characteristic lengths below 0.02, so you should first acquire an interactive allocation, generate the background mesh with a dry run, then finally submit the job from a login node. For example, to run an experiment with CL 0.02 and load fraction 0.4:

# Get allocation
flux alloc --queue=pdebug --setattr=thp=always --setattr=hugepages=512GB -x -N1 -n1 -t 1h
# Pre-generate mesh (only use 1 process, since we aren't launching the job)
ratel-runner mpm press sticky-air flux-run /path/to/voxel/data 0.02 0.4 -n 1 --dry-run
# Return allocation
exit

# Submit job to queue using generated mesh (note, use 16 processes)
ratel-runner mpm press sticky-air flux-run /path/to/voxel/data 0.02 0.4 -n 16

Alternate material properties can be provided as additional flags to the flux-run command. For example, to change the fracture toughness of the binder material, you could run

ratel-runner mpm press sticky-air flux-run /path/to/voxel/data 0.02 0.4 -n 1 --mpm_binder_fracture_toughness 1e2

Note: an extra - is required when compared to executing Ratel directly.