Re: [Myricom help #17130] Problems with Intel compilers for Linux and MPICH-GM

On Tue, Feb 11, 2003 at 04:45:47PM -0500, Patrick Geoffray wrote:
> Hi Craig,
> 
> On Sun, 2003-02-09 at 11:54, Craig Tierney wrote:
> > On Mon, Feb 03, 2003 at 07:53:12AM -0500, Susan Blackford wrote:
> 
> > > > You cannot use Intel 7.0 compilers with MPICH-GM 1.2.4..8a or earlier.
> > 
> > Why not?  I have tried using Intel 7.0 with mpich-gm 1.2.1..7b.  It seem
> > ok.  I do have problems but I cannot say for sure it is a problem with
> > mpich.  What types of problems could I be seeing with older mpich-gm versions?
> 
> There were 2 problems related to Intel 7.0 compilers in older mpich-gm
> release. The first one is the F90 deallocate not using free(), the
> second one is that the malloc lib (Doug Lea's malloc) was compiled with
> -O2, and it was too much for Intel compilers (the consistency of the
> malloc metadata would be compromised in rare cases).
> 
> So, if your code never calls F90 deallocate or malloc/free with a
> complex pattern, you should not have a problem using older mpich-gm
> versions with Intel 7.0. This incompatibility statement was broader on
> purpose, we didn't have all of the informations at that time.
> 
> What type of problems do you currently see ? numerical errors ?

I am having 2 outstanding problems with codes on the Intel system.
The first one is for WRF, which is the next MM5.  I have most issues
resolved but one user is having address errors when running the model.
I haven't been around for 2 weeks so I don't know the status.  

The second problem is numerical errors with MM5 when optimized.  The 
problem appears to be when amax and amin are optimized with SSE2.

The first problem might be related to a dealloc problem (WRF is f90)
but the MM5 is not.  

I am testing out mpich-1.2.5..9 on some Itanium2 right now.  When I 
get a chance I will grab a few nodes and try it out on the Xeons.

Craig

-- 
Craig Tierney ([email protected])