Re: [OMPI users] difference between OpenMPI - intel MPI -- how to understand where\why

2016-02-17 Thread Diego Avesani 
Dear all,

 I am trying to solve a huge linear system, which converges with open MPI
and not with MPI. The problem is that I do not know what is the exact
result.

Thanks

Diego


On 16 February 2016 at 16:13, Eugene Loh  wrote:

> Which one is producing correct (or at least reasonable) results? Are both
> results correct?  Do you have ways of assessing correctness of your results?
>
> On February 16, 2016 at 5:19:16 AM, Diego Avesani (diego.aves...@gmail.com)
>> wrote:
>>
>>> Dear all,
>>>   I have written an fortran-MPI code.
>>> Usually, I compile it in MPI or in openMPI according to the cluster where
>>> it runs.
>>> Unfortunately, I get complitly a different result and I do not know why.
>>>
>> ___
> users mailing list
> us...@open-mpi.org
> Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/users
> Link to this post:
> http://www.open-mpi.org/community/lists/users/2016/02/28541.php
>

Re: [OMPI users] difference between OpenMPI - intel MPI -- how to understand where\why

2016-02-17 Thread Jeff Squyres (jsquyres) 
-Original Message-
From: Diego Avesani 
Reply: Open MPI Users 
Date: February 17, 2016 at 8:13:35 AM
To: Open MPI Users 
Subject:  Re: [OMPI users] difference between OpenMPI - intel MPI -- how to 
understand where\why

> I am trying to solve a huge linear system, which converges with open MPI
> and not with MPI. 

I think you're trying to say "...which converges with Open MPI and not with 
*Intel* MPI", right?

> The problem is that I do not know what is the exact
> result.

What are the differences between your runs?

- you already identified Open MPI vs. Intel MPI
- are you running on different hardware?  (e.g., different clusters with 
different processors / accelerators, different network types, ...etc.)
- are you compiling with different compilers?
- are you using the same support libraries (e.g., numerical support libraries), 
and if so, were they compiled differently?

Usually, the best idea in these kinds of cases is to reduce the issue to a 
small program where you can reproduce the problem (i.e., reduce the unknowns to 
as small a number as possible), and then see where that leads to.

--  
Jeff Squyres
jsquy...@cisco.com  
For corporate legal information go to: 
http://www.cisco.com/web/about/doing_business/legal/cri/