Skip to content

Latest commit

 

History

History
277 lines (201 loc) · 9.27 KB

README.md

File metadata and controls

277 lines (201 loc) · 9.27 KB
Raw

Notice

This project has been superseded by miniforge, a project by conda-forge.

See:

Archiconda3

This repository holds many configuration scripts for the Archiconda3 distribution. Focused on porting conda-forge's work to 64 bit Arm processors.

The goal of this repository is develop the groundwork needed to compile conda-forge on aarch64. Once that is layed out, we will be working toward backporting much of this work to conda-forge

Tasks to do:

  1. Create an initial Archiconda installer.
  2. Upload it to github
  3. Create the shippable template that installs the Archiconda installer.
  4. Cereate an anaconda channel.
  5. Create the pinnings for Archiconda kinda done
  6. Create the templates for the different conda-smithy compatible
  7. Figure out how to generate the depency tree for all packages.
  8. Start rebuilding a few packages
  9. Create barebones Archiconda3 (3.7.1) and Archiconda2 (2.7.15) installers
  10. Build remaining packages necessary to install conda-build and anaconda-client using archiconda + conda-forge.
  11. Create a meta-channel for conda-forge for things to resolve faster
  12. Create a new docker image for archiarm
    • Must have conda-build + anaconda-client installed
    • Prefetch the compilers
    • Prefetch numpy + dependencies???? (we never compiled this yet)
    • Prefetch pytest???

How far along the stack do we need to go:

For users to install Archiconda3, they need to be able to install the conda package. conda-build, and anaconda-client do not need to be in that package, they can be obtained from conda-forge directly.

What this means, is that every package that conda depends on, to run, needs to be an arch specific package. The packages used to build those package do not need to be arch specific packages.

For now, these are the necessary packages:

  • python-3.7.0-hab5db58_3
  • ca-certificates-2018.11.29-ha4d7672_0
  • conda-env-2.6.0-1
  • libgcc-ng-7.3.0-h5c90dd9_0
  • libstdcxx-ng-7.3.0-h5c90dd9_0
  • libffi-3.2.1-h38784ca_1005
  • ncurses-6.1-hf484d3e_1002
  • openssl-1.0.2p-h14c3975_1002
  • xz-5.2.4-h14c3975_1001
  • yaml-0.1.7-h14c3975_3
  • zlib-1.2.11-h14c3975_1003
  • libedit-3.1.20170329-hf8c457e_1001
  • readline-7.0-h7ce4240_5 no, nobody wants readline.
  • tk-8.6.8-h14c3975_0
  • sqlite-3.26.0-hf8c457e_1000
  • asn1crypto-0.24.0-py37_1003
  • certifi-2018.11.29-py37_1000
  • chardet-3.0.4-py37_1
  • idna-2.7-py37_0
  • pycosat-0.6.3-py37h14c3975_0
  • pycparser-2.19-py37_0
  • pysocks-1.6.8-py37_0
  • ruamel_yaml-0.15.64-py37h14c3975_0
  • six-1.12.0-py37_1000
  • cffi-1.11.5-py37hb7f436b_1
  • setuptools-40.6.3-py37_0
  • cryptography-2.3.1-py37hb7f436b_1
  • wheel-0.32.3-py37_0
  • pip-18.1-py37_1000
  • pyopenssl-18.0.0-py37_0
  • urllib3-1.23-py37_0
  • requests-2.19.1-py37_0
  • conda-4.5.12-py37_1000

Limitations of the approach

We are basically compute bound at this point. shippable gives us 1 CI to use at a time (for one organization).

1 build, even for a nearly empty pure python package, takes about 3 minutes.

See for example the package setuptools: https://app.shippable.com/github/Archiconda/setuptools-feedstock/dashboard

The 3 minutes (180 seconds) are broken up as follows:

  • 15s Shippable things we have no control over
  • 5 seconds updating ubuntu's cache of apt
  • 4 seconds installing bzip2 and curl
  • 2 seconds downloading Archiconda3
  • 40 seconds installing Archiconda3 (including conda-build and anaconda-client)
  • 100 seconds building the package
  • 5 seconds getting the package name
  • 5 seconds to upload the package
  • 1 second leaning up.

We can maybe cut 50 seconds of this by uploading our own container.

But ultimately, we only have 1 CI, with 1 parallel job at a time, so we cannot run too many feedstocks at once.

It does have 96 cores, so maybe we can find a different way to parallelize things? I really feel like that might be over complicating things.

How to speed things up

I haven't had many problems compiling standard software.

For example, python compiled on the first shot, and the issues were primarely due to the fact that some software hardcodes binutil dependencies.

Therefore, we can potentially not have shippable automatically get triggered.

Users would have to wait until a regular linux 64 bit build passes, before triggering the shippable build.

How to start:

system requirements

python
conda
conda constructor
selenium
chromedriver-binary
python-chromedriver-binary

Be friends with jjhelmus

He will build the following critial packages:

The compilers and libstdc

These include the following:

binutils_impl_linux-aarch64-2.29.1-hc862510_0.tar.bz2
binutils_linux-aarch64-2.29.1-h1dbaa89_0.tar.bz2
crosstool-ng-1.23.0.451_g5888cf1-5.tar.bz2
gcc_impl_linux-aarch64-7.3.0-h68995b2_0.tar.bz2
gcc_linux-aarch64-7.3.0-h98564e2_0.tar.bz2
gdb_linux-aarch64-7.12.1-h6bc79d0_0.tar.bz2
gfortran_impl_linux-aarch64-7.3.0-h5c90dd9_0.tar.bz2
gfortran_linux-aarch64-7.3.0-h98564e2_0.tar.bz2
gxx_impl_linux-aarch64-7.3.0-h5c90dd9_0.tar.bz2
gxx_linux-aarch64-7.3.0-h98564e2_0.tar.bz2
libgcc-ng-7.3.0-h5c90dd9_0.tar.bz2
libgfortran-ng-7.3.0-h6bc79d0_0.tar.bz2
libstdcxx-ng-7.3.0-h5c90dd9_0.tar.bz2

Technically you don't need make (listed below), but i've had so much trouble building it, I'm just going to take jjhelmus'

make-4.2.1-h7b6447c_1.tar.bz2

These packages are available from his c4aarch64 anaconda channel.

He also created a tag on anaconda.org, I assume through cross compilation, where you can bootstrap a conda for aarch64.

https://anaconda.org/jjhelmus/repo?label=aarch64_bootstrap

Now the fun stuff

  1. Create a docker image that uses Centos7, and has Archiconda installed.

  2. There is a bug in the version of conda in the bootstrap, that doesn't allow you to have multiple channels. The maximum number of channels is 2. This is why this next step is important.

  3. Use that anaconda channel to create an anaconda installer. See the installer directory in this repositiory.

  4. Start building pacakges.

  5. You can try and build make, but after you do that, try and build m4 with your newly cut version of make. I couldn't get it to work. I think jjhelmus did something special. Or I did something wrong :/

  6. The build order is documented #4

Building recipes

Most of the heavy lifting is done by fork_conda_forge.py, a python script that does:

  1. Uses the github API to check if Archiconda already has the desired feedstock.
  2. If not, it will fork it from conda-forge
  3. Sets up an aarch64 branch.
  4. Enables building the repository on shippable.
    • To enable the repository on shippable, you must first request access to aarch64 machines.
    • Once you get access, you need to pay for an API key trust that I'm not stealing your creditials, inspect the code in fork_conda_forge.py to believe that I'm not stealing your credentials.
    • You will see a chrome window popup, and some explination on the command line as to what is happening.
    • Basically, I screen scrape and click buttons for you to enable the repository.
    • You need to log in manually the first time so that the script can save your cookies.
  5. Rerenders the recipe with conda smithy rerender
    • You should install conda-smithy from Archiconda to rerender things correctly.
  6. Pushes the changes to the aarch64 branch.

Because the repository has been enabled, pushing to that branch will trigger a build on shippable.

Docker image

The Docker image should:

  1. Set the locale
# Set the locale
ENV LANG en_US.UTF-8
ENV LANGUAGE en_US:en
ENV LC_ALL en_US.UTF-8

  1. Install a few basic pacakges. bzip2 and curl or wget are necessary to help bootstrap the process.

  2. Install Archiconda3.

  3. Add the archiconda3 to the path.

aarch64 docker on x86 using qemu

If you are creating the docker image while running on your personal laptop, it is useful to use qemu-static to run an aarch64 inside x86.

Put the line

# needed to build this on x86
COPY qemu-aarch64-static /usr/bin/

and you should be gravy.

Bulding using qemu and docker

I've found that it can be helpful to build and test using qemu and docker. It might be as simple as iterating quickly to find what dependencies were missing from the original packages.

Often, binutils are hardcoded, so you have to pass compilation flags to tell the build system what ar command to use.

It might be useful to mount a local directory to use with the docker build system.

For example, the following command mounts the registration directory which contains a bunch of feedstocks and runs the archiconda/centos7 image. The first command it runs is bash allows you to interact with the system.

docker run -v /home/mark2/git/aarch64/build-tools/registration:/feedstocks -i -t archiconda/centos7 bash

Be warned, CPU emulation is SLOW and will make your computer crawl.

I started compiling cmake at the same time it did on shippable, and didn't even get through the bootstrap when it has finished compiling.