Temporary addition of GNU 7.3 to GHA about atlas HOT 13 CLOSED

@wdeconinck since this involves eckit and odc too I would prefer this is done in the new CI/CD framework.
Please liaise with @dtip or @figi44 to inject GNU 7.3 in our build matrix.

from atlas.

Thanks for this info. @wdeconinck I think we must then also add this to the rest of the stack.
Can you please liaise with @dtip ?

from atlas.

We're now testing the ECMWF stack with gnu 7.3.1 on CentOS 7.9

PR: ecmwf-actions/downstream-ci#14

Shout if you run into any issues.

from atlas.

@tlmquintino if you think that is necessary. Otherwise I can add just a few lines in the older atlas GitHub actions that I till have that uses Github-owned runners, for the remaining 6 months or so. Depends how high in the stack the UKMO wants this requirement to be.

from atlas.

I'm just tagging @yaswant here as he handles the software stack on our side.

from atlas.

Looking at the release notes for GCC 7.3 to 7.5 it looks like they were only bug fixes.
See https://gcc.gnu.org/gcc-7/changes.html

So unless you want to reproduce buggy behaviour, I would argue that using 7.5 is a very good proxy to using 7.3 for the purposes of ensuring the builds dont break and unit tests are passing.

from atlas.

As some fluke experience in recent pull-request showed (see #145), the gnu 7.5 had a c++ 14 compiler bug fixed compared to 7.3 (https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64095); thanks @odlomax !
That shows that we cannot rely on gnu-7.5 for gnu-7.3 compatibility.

from atlas.

@tlmquintino if you think that is necessary. Otherwise I can add just a few lines in the older atlas GitHub actions that I till have that uses Github-owned runners, for the remaining 6 months or so. Depends how high in the stack the UKMO wants this requirement to be.

@tlmquintino @wdeconinck

Thanks for getting back. UKMO is basically restricted to GNU 7.3 until we're fully up and running with with our new HPC -- that should happen by the end of this year.

We're currently using the following ECMWF packages in our JEDI work with JCSDA:

• eckit
• fckit
• atlas
• fiat
• ectrans
• odc

Our main concern is that some exotic parts of C++17 might slip into Atlas/eckit/fckit and we lose our capability to run it on the HPC. Once we're on the new HPC, this problem goes away. (until we start having fun with C++20...)

from atlas.

In the mean time of having it added to the new CI/CD framework, I have added it to the older ci with Github-hosted runners (see above referenced commit).
That is currently testing eckit/fckit/atlas with gnu-7 installed via apt-get install gcc-7 g++-7 gfortran-7.
Note that this is gcc 7.5 instead of 7.3 I don't know if that version difference is important though.

from atlas.

Thank you @wdeconinck - is there any way we can set minor version to 3? Its not required from CXX17 perspective and not relevant in the scope of the current discussion. Sharing this for information only.

Fortran-mpi related complication: F2018 features needed to support the build of the mpi_f08 module - we cant build this module with mpich or openmpi in gcc-7.3 environment. And this is the gnu compiler version we are stuck with current Cray.

from atlas.

@yaswant It seems not trivial to install gcc-7.3 when the default 'gcc-7' is 7.5 (see https://askubuntu.com/questions/1225026/how-can-i-install-a-specific-version-of-gcc-on-ubuntu-18-04)
The route is probably compiling a specific gcc version manually.
I'll leave that for the newer custom CI/CD self-hosted runners that are being developed.

Which code is using the mpi_f08 module?

from atlas.

@wdeconinck no problem.

My colleague was testing mpi_f08 module in ops based JEDI interface for the UM (ops-um-jedi). If I understand correctly fckit does not support shared memory so he used raw mpi there¹. It is unlikely that anyone using fckit will be doing that.

from atlas.

We're now testing the ECMWF stack with gnu 7.3.1 on CentOS 7.9

PR: ecmwf-actions/downstream-ci#14

Shout if you run into any issues.

Thanks @dtip, @wdeconinck, @tlmquintino!

from atlas.