Build examples

The SingularityCE Definition Files provided at rat-containers/singularityCE use chaining in order to build several containers for different purposes.

        ---
align: center
alt: SingularityCE container chaining
name: singularityCE_container_chaining
config:
    theme: 'neutral'
---

    flowchart TB
        CUDA(nvidia/cuda_11.5.2-devel-rockylinux8)
        ROCKYCUDA(rocky-cuda)
        ROCKYCUDAPR(rocky-cuda-project-rat)
        ROCKYCUDAPY(rocky-cuda-pyrat)

        CUDA -.->|pulls from base| ROCKYCUDA
        ROCKYCUDA ==> ROCKYCUDAPR
        ROCKYCUDA ==> ROCKYCUDAPY

Overview

  • nvidia/cuda_11.5.2-devel-rockylinux8: base image from Docker library to start bootstrap from, contains Rocky Linux 8 OS as well as packages required to compile CUDA related projects. It is however incomplete to compile Rat C++ libraries and more dependencies need to be installed on this system.

  • rocky-cuda: local image bootstrapped from previous image with more development dependencies required to compile Rat C++ libraries (e.g; Git, CMake, Ninja, LAPACK, OpenBLAS, Boost, etc…). This images installs all dependencies available in the package manager repositories of Rocky Linux using dnf.

  • rocky-cuda-project-rat: local image bootstrapped from rocky-cuda. This one installs all other libraries that could not be installed beforehand to compile Rat C++ libraries. It compiles from source TCLAP, Armadillo, SuiteSparse, Sundials, NLopt, and of course all Rat libraries.

  • rocky-cuda-pyrat: local image bootstrapped from rocky-cuda. Very similar to rocky-cuda-project-rat with the addition of Python related packages to simplify building and packaging of pyRat Python wheels.

The definition files contain all the necessary commands to build Rat C++ compliant containers with the adequate development environements, as well as a definition for pyRat.

Build examples

This subsection describes how to build the above images using the definition files available at rat-containers/singularityCE by chaining them like in the above diagram.

Hint

Feel free to take these as bases and implement images as see fit.

One could for instance:

  • “Flatten” the NVIDIA, rocky-cuda and the rocky-cuda-project-rat definitions into a single definition file to be more direct.

  • Update the libraries versions to newer ones (e.g. Rat v2.022.2) and patch some errors (if there are any).

  • Change the base OS from Rocky Linux 8 to another base image.

  • Install more libaries in the environment to compose Rat with other programs.

The following images are built within the rat-containers/singularityCE directory.

cd singularityCE

rocky-cuda image

This image serves as the base for rocky-cuda-project-rat and rocky-cuda-pyrat and should be built first if chaining.

  • Pull the NVIDIA CUDA Rocky Linux 8 image from Docker:

singularity pull docker://nvidia/cuda:11.5.2-devel-rockylinux8

This image contains the CUDA Toolkit (nvcc, libcublas, libcublasLt, …) needed to enable GPU acceleration for Rat.

It is also possible to install the CUDA Toolkit inside the container from conventional methods as described in CUDA Toolkit installation guide, however pulling it from an existing image seems quite faster.

Unfortunately this image does not contain basic C/C++ development tools and dependencies needed to compile Rat (yet).

  • Bootstrap the rocky-cuda image from the previous one:

sudo singularity build --force rocky-cuda.sif rocky-cuda.def

This will build the rocky-cuda.sif container from the rocky-cuda.def definition file.

It bootstraps from the previous Docker image and adds tools and libraries such as CMake, Boost, VTK and many others.

Dependency libraries are installed with -devel to include development files such as headers for compilation and linking.

Some other dependencies not being available in the dnf package manager repositories will need to be compiled locally for Rat. This will be done in the next images.

rocky-cuda-project-rat image

This image utilizes the previous rocky-cuda.sif image to compile and install the rest of the dependencies as well as the Rat C++ libraries.

sudo singularity build --force rocky-cuda-project-rat.sif rocky-cuda-project-rat.def

This definition file:

  • downloads needed tarball sources,

  • compiles and install last dependencies such as NLopt or SuiteSparse,

  • compiles and install Rat libraries.

The result is the rocky-cuda-project.sif container which includes everything needed to compile a Rat project with CUDA.

rocky-cuda-pyrat image

This image the very similar to the previous one. It simply adds a few utilities to easily compile pyRat and build wheels from within, such as UV and PatchELF.

sudo singularity build --force rocky-cuda-pyrat.sif rocky-cuda-pyrat.def

The result is the rocky-cuda-project.sif container which includes everything needed to compile a Rat project with CUDA, as well as tools to easily build pyRat and Python wheels.