Docker assets need to be automatically generated

[wmalpica]

We have talked with Nvidia and we should be able to leverage their system to generate our docker images. Meeting notes from 5/27/2020:

image types they generate:

• base image
• runtime image (conda packages) --> this is what we would want to build
• devel images (source build github repos. all rapids suite)

generation matrix:

• cuda: 10.0, 10.1, 10.2
• python: 3.6, 3.7
• gcc 5, gcc 7 (certain gcc are only for certain OS)
• OS: 16.04, 18.04, centos7

They build docker containers 1 a day (at night)

TODO:

• we would want to give them a dockerfile
  • we would need to do some work on our docker image to have the full matrix
  • image base should be gpuci/miniconda-cuda
    • extends from the cuda images but adds conda
    • links: DockerHub: https://hub.docker.com/r/gpuci/miniconda-cuda GitHub: https://github.com/rapidsai/gpuci-build-environment (edited)
  • AJ Schmidt has done a lot of dockerfile templating he can help us with that
    • they use jinja (.j2 files) https://github.com/rapidsai/build CONTRIBUTING.md has a good overview of the repo

[aucahuasi]

Review our cuda compilation flags and check how cuml is handling the compatibility support across gpus with a single build:

cuML example for GPU Archs: https://github.com/rapidsai/cuml/blob/branch-0.14/cpp/CMakeLists.txt#L225 And the conda build explicitly uses all architectures: https://github.com/rapidsai/cuml/blob/branch-0.14/conda/recipes/libcuml/build.sh#L13

[mario21ic]

In order to generate the docker image run this command:

./docker-build.sh 11.0 3.7 0.18 nightly