

9. Linux Containers

Singularity

Warning

We are going to stop offering Singularity 2.6 at the end of March.

Singularity is available on the ABCI System. Available versions are Singularity version 2.6 and SingularityPRO 3.5. To use Singularity, set up user environment by the module command.

Singularity 2.6

[username@g0001~]$ module load singularity/2.6.1

SingularityPRO 3.5

[username@g0001~]$ module load singularitypro/3.5

More comprehensive user guide for Singularity will be found:

• Singularity 2.6 User Guide
• SingularityPRO 3.5 User Guide

To run NGC-provided Docker images on ABCI by using Singularity: NVIDIA NGC

Running a container with Singularity

When you use Singularity, you need to start Singularity container using singularity run command in job script. The container image is downloaded at first startup and cached in home area. The second and subsequent times startup is faster by using cached data.

Example) Execution of Singularity

The following sample is execution of Singularity using caffe2 container image published in Docker Hub. python sample.py is executed on Singularity container started by singularity run command.

Singularity 2.6

[username@es1 ~]$ qrsh -l rt_F=1 -l h_rt=1:00:00
[username@g0001~]$ module load singularity/2.6.1
[username@g0001~]$ singularity run --nv docker://caffe2ai/caffe2:latest
Docker image path: index.docker.io/caffe2ai/caffe2:latest
Cache folder set to /fs3/home/username/.singularity/docker
Creating container runtime...
...
[username@g0001~]$ python sample.py
True

SingularityPRO 3.5

[username@es1 ~]$ qrsh -l rt_F=1 -l h_rt=1:00:00
[username@g0001~]$ module load singularitypro/3.5
[username@g0001~]$ singularity run --nv docker://caffe2ai/caffe2:latest
...
Singularity> python sample.py
True

Create a Singularity image (pull)

Singularity container image can be stored as a file. This procedure shows how to create a Singularity image file using pull.

Example) Create a Singularity image file using pull

Singularity 2.6

[username@es1 ~]$ module load singularity/2.6.1
[username@es1 ~]$ singularity pull --name caffe2.img docker://caffe2ai/caffe2:latest
Docker image path: index.docker.io/caffe2ai/caffe2:latest
Cache folder set to /fs3/home/username/.singularity/docker
...
[username@es1 ~]$ ls caffe2.img
caffe2.img

SingularityPRO 3.5

[username@es1 ~]$ module load singularitypro/3.5
[username@es1 ~]$ singularity pull caffe2.img docker://caffe2ai/caffe2:latest
INFO:    Converting OCI blobs to SIF format
INFO:    Starting build...
...
[username@es1 ~]$ ls caffe2.img
caffe2.img

Example) Start a container using Singularity image file

Singularity 2.6

[username@es1 ~]$ module load singularity/2.6.1
[username@es1 ~]$ singularity run ./caffe2.img

SingularityPRO 3.5

[username@es1 ~]$ module load singularitypro/3.5
[username@es1 ~]$ singularity run ./caffe2.img

Example) Using a Singularity image file in a job script

Singularity 2.6

[username@es1 ~]$ cat job.sh
(snip)
source /etc/profile.d/modules.sh
module load singularity/2.6.1 openmpi/3.1.6

mpiexec -n 4 singularity exec --nv ./caffe2.img \
    python sample.py

SingularityPRO 3.5

[username@es1 ~]$ cat job.sh
(snip)
source /etc/profile.d/modules.sh
module load singularitypro/3.5 openmpi/3.1.6

mpiexec -n 4 singularity exec --nv ./caffe2.img \
    python sample.py

Create a Singularity image (build)

In the SingularityPRO 3.5 environment of the ABCI system, You can build container image files using fakeroot option.

Warning

You cannot build a container image from the recipe file in Singularity 2.6 environment. To use your custom container image, you adapt your own server environment to the ABCI environment (the version of singularity, framework, and mpi), build a container image on it, and then move the container image to ABCI system.

Example) Create a Singularity image file using build

SingularityPRO 3.5

[username@es1 ~]$ module load singularitypro/3.5
[username@es1 ~]$ singularity build --fakeroot ubuntu.sif ubuntu.def
INFO:    Starting build...
(snip)
INFO:    Creating SIF file...
INFO:    Build complete: ubuntu.sif
[username@es1 singularity]$

If the above command is executed under the group area (/ groups1,/groups2), an error occurs. This can be avoided by executing the newgrp command after checking the group to which the id -a command belongs, as shown below, before executing the singularity command. In the example below, gaa00000 is the group to which it belongs.

[username@es1 groupname]$ id -a
uid=0000(aaa00000aa) gid=0000(aaa00000aa) groups=0000(aaa00000aa),00000(gaa00000)
[username@es1 groupname]$ newgrp gaa00000

Build Singularity image from Dockerfile

On ABCI, you cannot build a Singularity image directly from Dockerfile. If you have only Dockerfile, you have two ways to build a Singularity image on ABCI.

Via Docker Hub

Build a Docker container image from Dockerfile on a system having Docker execution environment, and upload the image to Docker Hub. You can use the Docker container image on ABCI.

Following example shows how to build SSD300 v1.1 image developed by NVIDIA from Dockerfile and upload it to Docker Hub.

[user@pc ~]$ git clone https://github.com/NVIDIA/DeepLearningExamples
[user@pc ~]$ cd DeepLearningExamples/PyTorch/Detection/SSD
[user@pc SSD]$ cat Dockerfile
ARG FROM_IMAGE_NAME=nvcr.io/nvidia/pytorch:20.06-py3
FROM ${FROM_IMAGE_NAME}

# Set working directory
WORKDIR /workspace

ENV PYTHONPATH "${PYTHONPATH}:/workspace"

COPY requirements.txt .
RUN pip install --no-cache-dir git+https://github.com/NVIDIA/dllogger.git#egg=dllogger
RUN pip install -r requirements.txt
RUN python3 -m pip install pycocotools==2.0.0

# Copy SSD code
COPY ./setup.py .
COPY ./csrc ./csrc
RUN pip install .

COPY . .
[user@pc SSD]$ docker build -t user/docker_name .
[user@pc SSD]$ docker login && docker push user/docker_name

To run the built image on ABCI, please refer to Running a container with Singularity

Convert Dockerfile to Singularity recipe

By converting Dockerfile to Singularity recipe, you can build a Singularity container image which provides the same functionality defined in the Dockerfile on ABCI. You can manually convert Dockerfile, but using Singularity Python helps the conversion.

Warning

The conversion of Singularity Python is not perfect. If singularity build fails when the generated Singularity recipe file is used, modify the recipe file manually.

Example procedure for installing Singularity Python)

[username@es1 ~]$ module load python/3.6/3.6.5
[username@es1 ~]$ python3 -m venv work
[username@es1 ~]$ source work/bin/activate
(work) [username@es1 ~]$ pip3 install spython

Following example shows how to convert Dockerfile of SSD300 v1.1 image developed by NVIDIA using Singularity Python and modify the generated Singularity recipe (ssd.def) so that it can correctly generate a Singularity image.

Just converting Dockerfile results in a built time error. To avoid the problem, this example modifies the Singularity recipe as described below.

• Files in WORKDIR will not be copied => Set the copy destination to the absolute path of WORKDIR
• No path to pip => Add a setting to take over environment variables available in Docker image

[username@es1 ~]$ module load python/3.6/3.6.5
[username@es1 ~]$ source work/bin/activate
(work) [username@es1 ~]$ git clone https://github.com/NVIDIA/DeepLearningExamples
(work) [username@es1 ~]$ cd DeepLearningExamples/PyTorch/Detection/SSD
(work) [username@es1 SSD]$ spython recipe Dockerfile ssd.def
(work) [username@es1 SSD]$ cp -p ssd.def ssd_org.def
(work) [username@es1 SSD]$ vi ssd.def
Bootstrap: docker
From: nvcr.io/nvidia/pytorch:20.06-py3
Stage: spython-base

%files
requirements.txt /workspace                     <- Change path
./setup.py /workspace                           <- Change path
./csrc /workspace/csrc                          <- Change path
. /workspace                                    <- Change path
%post
FROM_IMAGE_NAME=nvcr.io/nvidia/pytorch:20.06-py3
. /.singularity.d/env/10-docker2singularity.sh  <- Add

# Set working directory
cd /workspace

PYTHONPATH="${PYTHONPATH}:/workspace"

pip install --no-cache-dir git+https://github.com/NVIDIA/dllogger.git#egg=dllogger
pip install -r requirements.txt
python3 -m pip install pycocotools==2.0.0

# Copy SSD code
pip install .

%environment
export PYTHONPATH="${PYTHONPATH}:/workspace"
%runscript
cd /workspace
exec /bin/bash "$@"
%startscript
cd /workspace
exec /bin/bash "$@"

To create a Singularity image from the generated recipe file on ABCI, please refer to Create a Singularity image (build).

Docker

In the ABCI System, job can be executed on Docker container. When you use Docker, you need to set up user environment by the module command and specify -l docker option and -l docker_image option at job submission.

Warning

Docker container can not be used on memory-intensive node in the ABCI system.

option	description
-l docker	job is executed on Docker container
-l docker_images	specify using Docker image

The available Docker image can be referred by show_docker_images command.

[username@es1 ~]$ show_docker_images
REPOSITORY                TAG             IMAGE ID     CREATED       SIZE
jcm:5000/dhub/ubuntu      latest          113a43faa138 3 weeks ago   81.2MB

Warning

In the ABCI System, Users can use only Docker images provided in the system.

Example) job script using Docker

The following job script executes python3 ./test.py on Docker container.

[username@es1 ~]$ cat run.sh
#!/bin/sh
#$-cwd
#$-j y
#$-l rt_F=1
#$-l docker=1
#$-l docker_images="*jcm:5000/dhub/ubuntu*"

python3 ./sample.py

Example) Submission of job script using Docker

[username@es1 ~]$ qsub run.sh
Your job 12345 ("run.sh") has been submitted

Warning

Docker container is only available on a node-exclusive job.