Re: [mpi-21] ABI

["Yann K." <yann.kalemkarian@xxxxxxxx>]

Greg,

I've just read your talk, and was wondering when you gave it ? When did
you started thinking of this issue ?

I don't know if I understood all of the issue, but I was just imagining
something I feel is not real scifi, and might happen. Just questions to
be discussed if relevant :
+ If we chose to do an MPI ABI, will the ABI change accordingly to
processors evolution ? Will we have to publish a new ABI when we will
have 128bit cpus available in order to avoid performance issues ? Can an
ABI be really a standard thus, or only a kind of point-in-time-agreement ?
+ I read somewhere in this forum that MPI would be one day integrated in
OSes like sockets are today. If this becomes true, therefore, with time,
people from other industries (not HPC) might start using MPI to avoid
reprogramming communication libraries, or using a heavy CORBA-like
framework, because it's true it's easy.  If we have an ABI, will those
people be excluded if they use embedded boards using 32 bits
technologies, like the 32 bits chips Freescale does (we still find in a
lot of industries), because of the 64bit oriented ABI ?

My 2c.

Yann

Greg Lindahl wrote:
> On Tue, Dec 04, 2007 at 02:45:44PM -0500, Jeff Squyres wrote:
>
>   
>> Regardless of the fact that you correctly originally identified  
>> testing as still-an-open-issue, many others have somehow latched on to  
>> the claim that an ABI "reduces the ISV testing matrix"
>>     
>
> Well, fortunately, we're going to sit down in the future and have a
> good discussion about all of this. Meanwhile, you might do better
> to point people at one of my talks:
>
> http://www.pbm.com/~lindahl/Case_for_an_MPI_ABI.pdf
>
> This issue is addressed on slide 15.
>
> -- greg
>
>
>
>