Installer for Aarch64 (and ppc64le)

[hmaarrfk]

Now that the aarch migration is well underway, it seems like it would be to create a Miniconda installer for aarch64

Not too sure why this one is not working
Archiconda/build-tools#23

[jjhelmus]

I'd be in favor of this. The question is where to host the installer? Perhaps a GitHub release would be reasonable with a link to the installer from the conda-forge.org.

[mariusvniekerk]

Github release seems prudent.

[isuruf]

@hmaarrfk, let us know what we can do to help like creating a github repo, activating a CI, etc.

[jakirkham]

Thoughts @msarahan?

[hmaarrfk]

Jewlerconda3/construct.yaml

constructor --platform=linux-aarch64  Jewlerconda3/  

# what to call this???
name: Jewlerconda3
version: 0.1.0

channels:
  - https://conda.anaconda.org/conda-forge
  - https://conda.anaconda.org/anaconda
  # There still seems to be a dependency needed from c4aarch64
  - https://conda.anaconda.org/c4aarch64

write_condarc: True
#license_file: installer_license.txt

specs:
  - python 3.7.3
  - conda
  - pip

[mariusvniekerk]

forge-conda?

[hadim]

An aarch64 flavor of conda provided by Anaconda or conda-forge would indeed be super nice. Any news or ETA on this? What does remain to be done? Anything I can do to help.

In the meantime, I use Archiconda which does the job well.

[hmaarrfk]

somebody needs to test it on real metal hardware.

I got RPi4, but still haven't had a chance to connect it to a monitor.

[hadim]

I just tested the last Archiconda3 installer on an Rpi4 with Ubuntu 19.10 arm64 and it works perfectly.

I couldn't installed either update conda because of an issue with the university network here (DNS won't resolve) but I'll report once it works.

[hadim]

Waiting for the DNS issue to be solved I tried to build an installer with the packages I need:

name: Archiconda
version: 0.3.0

channels:
  - https://conda.anaconda.org/conda-forge
  - https://conda.anaconda.org/anaconda
  # There still seems to be a dependency needed from c4aarch64
  - https://conda.anaconda.org/c4aarch64

write_condarc: True

specs:
  - python
  - conda
  - pip
  - numpy
  - scipy
  - pandas
  - sqlalchemy
  - pyserial
  - slacker

constructor --platform=linux-aarch64 .

I built it on my host machine (Ubuntu x86_64) but when trying the installation on RPi 4 I have the following error:

$ bash Archiconda-0.3.0-Linux-aarch64.sh 

Welcome to Archiconda 0.3.0

Archiconda will now be installed into this location:
/home/ubuntu/archiconda

  - Press ENTER to confirm the location
  - Press CTRL-C to abort the installation
  - Or specify a different location below

[/home/ubuntu/archiconda] >>> 
PREFIX=/home/ubuntu/archiconda
Unpacking payload ...
Archiconda-0.3.0-Linux-aarch64.sh: line 259: /home/ubuntu/archiconda/conda.exe: cannot execute binary file: Exec format error
ERROR: could not extract tar starting at offset 000000000000011467+9231072+2

I guess it's because the installer is x86_64 only compatible. Any idea of how I could build an installer that works on arm64? Docker (but then how do you initially install constructor without conda)?

Also, I noticed a pretty important package is not available yet for arm64: matplotlib. I have open an issue about it: conda-forge/matplotlib-feedstock#221

[isuruf]

Docker (but then how do you initially install constructor without conda)?

Use condaforge/linux-anvil-aarch64 docker image which can be run on x86_64 after running https://github.com/conda-forge/conda-smithy/blob/84652a0239b1ffa0660bed8374f809b2dca390f1/conda_smithy/templates/azure-pipelines-linux.yml.tmpl#L25

[isuruf]

Btw, we can get rid of c4aarch64 if we built https://github.com/conda-forge/ctng-compilers-feedstock

[hmaarrfk]

I just setup a repo (for my own company) that would use Travis to upload binaries to GitHub on tags.

If we have a name setup for the installer, I'm glad to setup a CI to get everything going at this stage.

We can eventually build up automated tests to test it invarious docker instances for different flavors of Linux.

[hadim]

So I have been able to build my own installer and everything works smoothly (except for matplotlib that needs to installed manually).

An official (or at least semi-official) installer would be nice indeed. Happy to help/test/debug if needed.

[hmaarrfk]

The Archiconda one comes pretty close. I would only rename it and move it locations. Is anything (other than matplotlib) missing?