[gentoo-soc] Weekly Report: MPI Overlay, Week 6

[gentoo-soc] Weekly Report: MPI Overlay, Week 6

[Michael Gilroy]

[-- Attachment #1: Type: text/plain, Size: 4328 bytes --]

Hi all,

*Summary:*

Last week's plans:
1) Successfully export src_* and test with mpich.
2) Continue testing against mpi-providers and adding more versions of
openmpi and mpich to the overlay.
3) Continue adding to wiki/README as functionality is added. Will talk to
mentors about other means of documentation as well.

There were some change in plans, as I got some clarification with regards
to design. Essentially, mpi-select will have little focus on the mpi
implementations themselves, as the implementations will have no need to
inherit mpi-select, but rather the software the depends on various
mpi-select.

Next week's plan:
1) Successfully test mpi-select against hpl, disregarding having only
mpi-dependent libs installed.
2) Implement proper DEPEND= function based on users' make.conf settings.
3) Clarify design goals for mpi-dependent installations and start work on
executing this.

*Day-by-day breakdown:*

*Monday's Goals:*

Research and emulate PYTHON_TARGETS make.conf execution.
Work in progress. From Justin's email, I gathered that I have some more
research to do with Python's implementation in make.conf. Today I looked at
python code and tested different make.conf setups to see how they were
handled. I'm hoping this will give me a better framework to work with
similar design when I go to code some of this tomorrow. I will likely have
some questions about this when it comes to combining this with multilib's
handling of different ABIs.

Add userspace toggles (hopefully using eselect). I may have questions about
how to do this exactly.
Change of plans. This will be covered later with userspace MPI selection,
which I will tackle later on after the 2nd eclass is more developed.
However, I will have to keep this userspace implementation in mind when
writing the 2nd eclass.


*Tuesday's Goals:*

Implement a mixture of Python's make.conf setup while being conscious of
different multilib ABIs.

I believe there exists a much better way to do this as opposed to what I
have currently implemented. More specifically: is there a variable that
tracks the location of make.conf on any given system? Right now I am
searching for make.conf in a static grep search, which I assume is a bit
gross. I plan on continuing the email discussion early tomorrow.

- NOTE: as tomorrow is a national holiday in the US, while I will still be
working, my hours will be shorter. I am planning on 4-6 hours of work
tomorrow, and will continue regular shifts for the rest of the week to hit
35+ hours.



*Wednesday's Goals:*

Properly test and debug everything in mpi-select. There's been quite a bit
of code added that has not yet been tested, it is about time I got
mpi-select back to a stable build.
I am still struggling with gcc/fortran compilation issues. I have tried
installing and reinstalling gcc (versions 4.8.5 & 5.4.0) with
USE="+fortran" with no avail when it comes to compilation errors. Any help
with this would be much appreciated, as this has been a bit of a barrier
that I have held onto for too long thinking I could resolve it myself.

Continue design conversations. While things are starting to come together
more, some concepts are still foggy as there are many things to consider,
making tasks more complex.
Incomplete. I hope to set up a proper weekly meeting in the near future to
settle this.


*Thursday's Plans:*

Make mpi_wrapper in particular more useful in terms of environment
variables. Start deveopment on implementation-specific case statement in
src_* steps by referencing recent ebuilds.

Work in progress. I have started conditional src_* for primarily mpich.
Will look for openmpi ebuild trends tomorrow.

Resolve testing/compiler issues.
Work in progress. I anticipate on using a different bare metal Gentoo
environment for testing more.



*Friday's Plans:*

Add openmpi conditional src_* functionality to mpi-select.
Change in plans, as per our weekly meeting. I was a bit confused as to how
to execute this, but I will be working on the correct version this weekend.
I want to work on it while our conversation is fresh in memory ;)

Iron out bugs via testing, properly comment/document functions. Clean up &
refactor code that could look prettier.
Work in progress. Again, I will be using some of this weekend to work on
this while it is fresh in my mind.


Thanks,
Michael

[-- Attachment #2: Type: text/html, Size: 7397 bytes --]