Singularity
Background
Singularity is a container solution created by necessity for scientific and application driven workloads. The Singularity containers can be used to package entire scientific workflows, software and libraries, and even data. This means that you don’t have to ask your cluster admin to install anything for you - you can put it in a Singularity container and run.
(see webpage https://www.sylabs.io and/or http://singularity.lbl.gov/ for details)
Build a Singularity Container
Build a container from a Singularity recipe (ubuntu1604-cuda92-ompi400.def)
wget https://wiki.ncsa.illinois.edu/download/attachments/82518873/ubuntu1604-cuda92-ompi400.def singularity build ubuntu1604-cuda92-ompi400.simg ubuntu1604-cuda92-ompi400.def
Pull a container from Singularity Hub or Docker Hub
singularity pull docker://nvidia/cuda-ppc64le:9.2-cudnn7-devel-ubuntu16.04
Run a Singularity Container
Run a container with "exec" command
mpirun -n 4 singularity exec --nv /opt/apps/samples-image/ubuntu1604-cuda92-ompi400.simg hostname
Run a container with shell
singularity shell --nv /opt/apps/samples-image/ubuntu1604-cuda92-ompi400.simg
Build a writable Singularity container
By default, the Singularity container filesystem is read-only. There are 2 options to build a writable Singularity container.
The "--sandbox
" option allows users to create a container within a writable directory (called a sandbox). It’s possible to create a sandbox without root privileges, but to ensure proper file permissions it is recommended to do so as root.
singularity build --sandbox cuda-ppc64le/ docker://nvidia/cuda-ppc64le:9.2-cudnn7-devel-ubuntu16.0 singularity shell --writable cuda-ppc64le/
The "--writable"
option allows users to create a writable ext3 image. You must create writable containers as root.
singularity build --writable cuda-ppc64le.simg docker://nvidia/cuda-ppc64le:9.2-cudnn7-devel-ubuntu16.04 singularity shell --writable cuda-ppc64le.simg
Build a writable Singularity container based on a existing read-only image
Convert it to a writable directory (a sandbox):
singularity build --sandbox cuda-ppc64le-mod/ cuda-ppc64le.simg singularity shell --writable cuda-ppc64le-mod/
Use an existing container as a target to build a new container
singularity build --writable cuda-ppc64le-mod.simg cuda-ppc64le.simg singularity shell --writable cuda-ppc64le-mod.simg
Frequently Asked Questions
- Can I run a Singularity container with x86 architecture image on an IBM ppc64le system?
A: Singularity can not directly run x86 binaries on a IBM Power system because the bytes of machine code are different. Users need to make sure their own images are built for "ppc64le" rather than "x86". - What is the option "--nv" means?
A: The "--nv" allows the Singularity container to leverage the NVIDIA GPU on the host system.