Sticky thread for binary prerequisite uploads

[zbeekman]

Current status

Please note, this is not a bug at all. This issue is being kept open to make it easier to find, until a better/cleaner solution is implemented. This thread is a place to upload binary assets needed by our CI testing infrastructure. In this way, github issues are being used as a bootleg CDN, of sorts. At some point I think that I could create a project, using a combo of https://nightli.es, https://travis-ci.org and https://bintray.com to provide a free service for building and uploading MPICH and OpenMPI Homebrew bottles using the latest GCC. This has the potential to speed up or enable a lot of MPI projects that want to use GCC over clang on OS X using CI testing with Travis-CI.

Old Text

The MPI implementation for the Travis-CI builds on OS X cannot be built from source, and the only "bottled" installations use clang as the mpicc command. (Without verbose output the built times out after 10 minutes without anything being logged---i.e., quietly building MPI---and with verbose output the log length exceeds 50mb which also causes a failure.) This doesn't appear to impact the functionality of OCA, however, generating coverage data (#51) becomes nearly impossible due to some files requiring one version of gcov and other files requiring a different version. The codecov.io script is very good at "just do what I mean/want" and replacing it with manual coverage generation, collection, transmission etc. is just too daunting a task. (It requires a uniform version for gcov.)

Unless @rouson has a better idea of a server to host binary files (inappropriate for inclusion in a git repository) I'm going to try using the Github issues file uploader mechanism as a content delivery network (CDN) to provide pre compiled/bottled homebrew MPICH and OpenMPI implementations to the Travis-CI OS X workers that are built against gcc-5. My only other thought is to host it at my UMIACS.UMD.EDU public FTP page. This should greatly speed the OS X builds as apposed to building from source, and allow testing of the GCC-5 gcc compiler, along with generating coverage info.

Here is a test bottle upload: umobj-1.3.0.yosemite.bottle.tar.gz

This bottle is completely unrelated to OCA, but is a homebrew bottle.

[zbeekman]

Closing issue to see if file is still accessible.

[zbeekman]

Success! I'm going to work on building MPICH and OMPI with gcc-5.3, bottling them, then uploading them to this comment thread. I would guess that this is going to take at least 2 hours to finish, assuming all the Homebrew problems impacting MPI have finally been resolved...

[zbeekman]

Here is the mpich bottle:
mpich-3.2.yosemite.bottle.1.tar.gz

I think I should be able to tell Travis-CI to use an environment emulating yosemite.

[zbeekman]

Here is the OpenMPI bottle:

open-mpi-1.10.1_1.yosemite.bottle.1.tar.gz

[rouson]

This all sounds good and congrats on your successes so far! It sounds like you have things squared away and are well on your way to a solution. @zbeekman, if part of your question to me is where to host binaries online, I'll just mention that I find Dropbox convenient. One nice thing about is that I can put something in a Dropbox even if I'm offline and then it will automatically be uploaded once I'm online. And if the upload doesn't complete before I go offline again, it will seamlessly continue from where it left off the next time I'm online. When I need a URL to access the file via some remote mechanism, I command-click on the file and select "Copy Public Link" in the dropdown menu to put the URL on my clipboard.

Presumably your approach is better because it makes sense to have the files in archived on github.com somewhere and doing it through the Issues uploader enables you to add descriptive text and track related issues. I only mentioned Dropbox as a convenient alternative if the issues uploader approach proves to be cumbersome or is undesirable for any other reason.

[zbeekman]

mpich-3.2.x86_64_linux.bottle.1.tar.gz

This is a linux brew bottle. One problem is that the formula's bottle do block will need to be patched with
sha256 "6b2ef4f73d43bf40891aac2d1aabb78cf9dfaa692bafd07ee8c0d90eb3ccfdcb" => :x86_64_linux

[zbeekman]

OpenMPI on OS X needs updating:
open-mpi-1.10.2.yosemite.bottle.1.tar.gz

Edit: This seems not to work... but I'm investigating

[zbeekman]

open-mpi-1.10.2.yosemite.bottle.tar.gz

This bottle was rebuilt on a vagrant yosemite install, with a proper cellar location. It should work now.

[zbeekman]

I'm going to attempt, for now, to just use Homebrew bottled MPI implementations... these will use clang for the C compiler... hopefully that doesn't cause any problems.