Minified, multi-architectural container images

[jennydaman]

Is your feature request related to a problem?

We are interested in using N4BiasFieldCorrection as part of our pipelines and run it in containers on PowerPC.

Proposed new features

A) In Dockerfile, by installing cmake using apt instead of directly from Kitware's website, we can build multi-architectural images using docker buildx. See here:

https://github.com/FNNDSC/ANTs/blob/04c2938830f6c1233bde9c4784ad2f58d6a8f058/Dockerfile

I intend to make a PR after figuring out a good solution to #1069

B) Small (<60MB) images can be produced by creating individual images for each ANTs program. See

https://github.com/FNNDSC/microANTs

Using this script, I created a multi-architectural image for N4BiasFieldCorrection on DockerHub: https://hub.docker.com/layers/fnndsc/n4biasfieldcorrection/2.5.0/images/sha256-7e4da5a9f606aa170ae8d1187a6de915f5a1d03ca2dd7e773fc24c411c2227ab?context=explore (the ppc64le one doesn't work yet, as aforementioned)

If there is more interest in the community, I can use a GitHub Actions matrix to automatically build these minified images for all ANTs programs. I would also like to know whether the ANTsX team is interested in hosting these images and CI infrastructure in their own official namespace.

Alternatives you've considered

docker buildx is a very inefficient solution because it uses QEMU emulation. Cross-compilation would be a much more efficient solution, however it would require in-depth knowledge of the Cmake configurations of ANTs which I lack.

Alternatively, it would also be even better if the ants package in conda-forge was cross-compiled for linux-aarch64 and linux-ppc64le.

Additional context

[cookpa]

Hi,

A few thoughts:

ANTsPy has moved to using a conda cmake install for cross-platform compatibility, and it probably would be good to have ANTs do the same.

I'm not sure about the use case for the mini ANTs images. What does having a separate image for each program do? Would it make sense to just have one static build per platform, so people could copy the program(s) they want in their Dockerfile?

Cross-platform builds would be awesome. QEMU takes too long for Github Actions, so this would be a good alternative. But I don't have an arm64 machine to work on.

Regarding conda forge, you'd have to reach out to the developers of that package, we don't maintain that directly - though we've collaborated on updating ANTs to help facilitate conda builds and are happy to discuss further here.

[jennydaman]

Yeah, I was the one who sent the ANTsPy PR those years ago.

We're interested in using N4BiasFieldCorrection in "pipelines" which run on Kubernetes, i.e. run a sequence of containers to process a dataset on a cluster of computers. Small image sizes can greatly decrease cold boot time. Since N4BiasFieldCorrection might run on any number of the nodes in our cluster, the image needs to be pulled quite often.

[cookpa]

Right, the apt conda install was better than the hard-coded kitware download, but it eventually ran into problems when we had to update the CMake version. Another user suggested using conda to install it, which has worked well since then. So I was thinking to copy the same procedure (#1646)

[jennydaman]

@cookpa sure! I actually have a branch for that already: https://github.com/FNNDSC/ANTs/blob/multiarch-dockerfile/Dockerfile

Do you want a PR?

[jennydaman]

Comparing the Dockerfile I wrote to #1646, I am using micromamba which is a C++ reimplementation of conda. Also, I am using cache mounts to improve speed and minimize the image size. You can try

docker pull docker.io/fnndsc/ants:2.5.0-micromamba

[cookpa]

This looks interesting, but I'm inclined to keep things simple for ANTs, I like your Dockerfile here

https://github.com/FNNDSC/ANTs/blob/04c2938830f6c1233bde9c4784ad2f58d6a8f058/Dockerfile

Using Debian and shorter than the existing one. Could you make a PR for that?

I was also thinking of making the shared libs option a build arg - then one could build statically, and then copy a program like N4BiasFieldCorrection into a micro distribution. The executable would be larger but I think on balance it would be smaller than copying the shared-lib executable and all of its shared libraries. Unless this wouldn't work for some reason I'm missing

[jennydaman]

I did a force-push to amend the commit message to follow the convention.

[jennydaman]

Also, static builds would be very helpful! (I lack the cmake expertise to do it myself)

[cookpa]

Hi @jennydaman , I made a few changes in #1665, among them adding a build arg for BUILD_SHARED_LIBS, set this to OFF and it will build static libraries and executables.

After building a static image with all the tools, I was then able to copy individual executables (no need to bring the libraries) into a debian:bookworm-slim base image, and then run them directly.

Thanks again for your help with the Dockerfile

[jennydaman]

I tried out BUILD_SHARED_LIBS=OFF and it works, many thanks!