Re: New prerequisites to support multi image COARRAY in gfortran

[Jerry DeLisle]

On 04/04/2017 10:44 AM, FX wrote:
>> We choose mpich as a default only because it is very stable.
> 
> Why are why tying ourselves to one MPI implementation?
> 
> FX
> 

Not tying ourselves at all. This just gives users who install gcc manually with
the ./configure process a default to use and only if they are using the download
prerequisites script which is optional. Distributions can still provide multiple
choices with package managers.

The final tie is at link time where a user may specify any compatible library
they choose. We do have a base of users who do not use package managers by the
way. So the goal is a minimal installation of a fully functional Fortran.

Jerry

Re: New prerequisites to support multi image COARRAY in gfortran

[Damian Rouson]

 



On April 4, 2017 at 9:48:23 AM, Jerry DeLisle 
(jvdeli...@charter.net(mailto:jvdeli...@charter.net)) wrote:

> Gerald, (or who does this)  
>  
> Since shared memory parallel programming with Fortran is now a Standard 
> feature  
> of the language, we would like to support full parallelism through the use of 
>  
> mpich and OpenCorrays.  


Clarification: Fortran’s parallel programming model is agnostic about the 
memory architecture. The 2008 standard made Fortran the only widely used 
language with a parallel programming model that works in shared- or 
distributed-memory systems.  I’ve seen papers in which people have run coarray 
Fortran (CAF) on over 100,000 cores with impressive results.

As a side note, the draft 2015 standard makes Fortran also the only widely used 
language with support for fault tolerance, which makes Alessandro’s and Andre’s 
work on failed-image support truly exciting and I’m amazed that there already 
exists an MPI limitation with some support for what is needed to support 
failed-image detection. 

Damian

Re: New prerequisites to support multi image COARRAY in gfortran

[Damian Rouson]

 
Resending as plain text (although now I realize my reply is at least partially 
redundant since reading Andre’s email):

On April 4, 2017 at 10:44:09 AM, FX 
(fxcoud...@gmail.com(mailto:fxcoud...@gmail.com)) wrote:

> > We choose mpich as a default only because it is very stable.
>  
> Why are why tying ourselves to one MPI implementation?
>  


Jerry’s request is just to put an MPICH tar ball on the GCC ftp server as a 
default.  When we modify “configure,” there will be a --with-mpi option that 
enables users to choose their own MPI implementation. 

Damian

Re: New prerequisites to support multi image COARRAY in gfortran

[Andre Vehreschild]

Hi FX,

we don't really do. You can use other MPI implementation as well, albeit not all
features are implemented in all the others. For failed images support we needed
the fault tolerance support for MPI, which I see only in mpich from 3.2 on. The
other MPI implementations publicy available do not support ft at the moment. So
as soon as they do in their stable versions, we might give the user more choice.

- Andre

On Tue, 4 Apr 2017 19:44:06 +0200
FX  wrote:

> > We choose mpich as a default only because it is very stable.  
> 
> Why are why tying ourselves to one MPI implementation?
> 
> FX


-- 
Andre Vehreschild * Email: vehre ad gmx dot de 

Re: New prerequisites to support multi image COARRAY in gfortran

[FX]

> We choose mpich as a default only because it is very stable.

Why are why tying ourselves to one MPI implementation?

FX

New prerequisites to support multi image COARRAY in gfortran

[Jerry DeLisle]

Gerald, (or who does this)

Since shared memory parallel programming with Fortran is now a Standard feature
of the language, we would like to support full parallelism through the use of
mpich and OpenCorrays.

We choose mpich as a default only because it is very stable. We choose
OpenCoarrays because it works well with gfortran and the developers have also
supported significant portions of the gfortran frontend (not to mention that it
is the only such library available). Choosing these as defaults for gfortran
does not exclude any user from using any other libraries in the future since the
final connection occurs at link time when a user compiles with -fcoarray=lib.

In order to support testing of installation of the new packages we are
requesting that the package tarballs be placed on the gcc.gnu.org along with the
other gcc prerequisite packages here:

ftp://gcc.gnu.org/pub/gcc/infrastructure/

There will be two stable released packages with appropriate shasum information
to be added. I can provide the links to those packages for download or send them
to you directly.

Requesting approval for this first step. (Later steps (more than one) will
provide patches for review of the contrib/download_prerequisites script. Another
step will be modifying OpenCoarrays build structure to align with the Gnu/gcc
build process. I believe mpich is already suitable, but will be testing.)

Is this OK?

Regards,

Jerry