JUSTUS2/Software/Singularity: Difference between revisions

From bwHPC Wiki
Jump to navigation Jump to search
Line 94: Line 94:
</pre>
</pre>


Note: Keep in mind that other modules you may have loaded will not be available inside the container.
<font color=red>Keep in mind that other modules you may have loaded will not be available inside the container.</font>


=== Using GPUs ===
=== Using GPUs ===

Revision as of 18:06, 22 March 2021

Description Content
module load system/singularity
Availability BwForCluster_JUSTUS_2
License Open-source software, distributed under the 3-clause BSD License. More...
Citing ---
Links Homepage | Documentation
Graphical Interface No

Description

Singularity is a container platform.

License

Singularity is free, open-source software released under the 3-clause BSD license. Please read the license for additional information about Singularity.

Usage

Loading the module

You can load the default version of Singularity with the following command:

$ module load system/singularity

If you wish to load another (older) version of Singularity, you can do so using

$ module load system/singularity/<version>

with <version> specifying the desired version.

Important: On JUSTUS 2, Singularity is available on all compute nodes. You do not have to load a module.

Program Binaries

The binary singularity is main program of the container platform.

To get help using Singularity execute the following command:

$ singularity --help

Furthermore, a man page is available and can be accessed by typing:

$ man singularity

For additional information about how to use Singularity, please consult the documentation.

Important: On JUSTUS 2, Singularity is only available on compute nodes. You must switch to such a node to invoke any singularity commands.

Batch jobs with containers

Batch jobs utilizing Singularity containers are generally built the same way as all other batch jobs, where the job script contains singularity commands. For example:

#!/bin/bash
# Allocate one node
#SBATCH --nodes=1
# Number of program instances to be executed
#SBATCH --tasks-per-node=4
# 8 GB memory required per node
#SBATCH --mem=16G
# Maximum run time of job
#SBATCH --time=1:00:00
# Give job a reasonable name
#SBATCH --job-name=Singularity
# File name for standard output (%j will be replaced by job id)
#SBATCH --output=singularity_job-%j.out
# File name for error output
#SBATCH --error=singularity_job-%j.err

module load your/singularity/version #(not needed on JUSTUS 2, but could be necessary on other system)

cd your/workspace

# Run container (two options to start a container)
singularity run [options] <container>
singularity exec [options] <container> <command>

Keep in mind that other modules you may have loaded will not be available inside the container.

Using GPUs

#!/bin/bash
[…]
# Allocate one GPU per node
#SBATCH --partition=gpu
#SBATCH --gres=gpu:1
[…]

module load your/singularity/version #(not needed on JUSTUS 2, but could be necessary on other system)

cd your/workspace

# Run container (two options to start a container)
singularity run --nv [options] <container>
singularity exec --nv [options] <container> <command>

Using the flag is advisable, but may be omitted if the correct GPU- and driver-APIs are available on the container.

Examples

Run your first container on JUSTUS 2

Build a TensorFlow container with Singularity and execute a Python command:

# request interactive node with GPUs
$ srun --nodes=1 --exclusive --gres=gpu:2 --pty bash

# create workspace and navigate into it
$ WORKSPACE=`ws_allocate tensorflow 3`
$ cd $WORKSPACE

# build container
$ singularity build tensorflow-20.11-tf2-py3.sif docker://nvcr.io/nvidia/tensorflow:20.11-tf2-py3

# execute Python command
$ singularity exec --nv tensorflow-20.11-tf2-py3.sif python -c 'import tensorflow as tf; \
  print("Num GPUs Available: ",len(tf.config.experimental.list_physical_devices("GPU")))'

Note: Ready-to-use containers can be pulled from the NVIDIA GPU CLOUD (NGC) catalog.

NGC environment modules on JUSTUS 2

1) Prepare a workspace to store the container

$ srun --nodes=1 --exclusive --gres=gpu:2 --pty bash
$ WORKSPACE=`ws_allocate npc 3`
$ cd $WORKSPACE
$ export NGC_IMAGE_DIR=$(pwd)

Important: Containers can only run in workspaces.

2) PyTorch container

$ module load ngc/.numlib
$ module load 20.12-py3
$ python3
>>> import torch
>>> x = torch.randn(2,3)
>>> print(x)
>>> quit()
$ module unload 20.12-py3
$ module unload ngc/.numlib

Note: Use the container in the same manner as an interactive shell.

3) LAMMPS container

$ module load ngc/.chem
$ module load 29Oct2020
$ wget https://lammps.sandia.gov/inputs/in.lj.txt
$ export SINGULARITY_BINDPATH=$(pwd)
$ mpirun -n 2 lmp -in in.lj.txt -var x 8 -var y 8 -var z 8 -k on g 2 -sf kk -pk kokkos cuda/aware on neigh full \
  comm device binsize 2.8
$ module unload 29Oct2020
$ module unload ngc/.chem

Note: Use SINGULARITY_BINDPATH=<PATH> to mount the directory with the input file.


Currently, module load ngc/.numlib for numeric libraries and module load ngc/.chem for chemistry programs can be selected.

Batch jobs with containers on JUSTUS 2

Run a GROMACS container with Singularity as a batch job:

$ WORKSPACE=`ws_allocate gromacs 3`                                         # allocate workspace
$ cd $WORKSPACE                                                             # change to workspace
$ singularity pull gromacs-2020_2.sif docker://nvcr.io/hpc/gromacs:2020.2   # pull container from NGC
$ cp -r /opt/bwhpc/common/chem/ngc/gromacs/ ./bwhpc-examples/               # copy example to workspace
$ cd ./bwhpc-examples                                                       # change to example directory
$ sbatch gromacs-2020.2_gpu.slurm                                           # submit job
$ squeue                                                                    # obtain JOBID
$ scontrol show job <JOBID>                                                 # check state of job

More batch job examples are located at /opt/bwhpc/common/chem/ngc.

Useful links