Re: Linux emulation, binary software, cluster/grid and SMP

["Bruce O'Neel" <edoneel%sdf.lonestar.org@localhost>]

On Wed, Jan 18, 2006 at 11:20:31AM -0500, George Georgalis wrote:
> Hi Masi,
> 
> On Wed, Jan 18, 2006 at 08:30:31AM +0100, Massimiliano Masi wrote:
> >Il giorno mar, 17/01/2006 alle 21.53 -0500, George Georgalis ha scritto:
> >
> >> Has anyone used linux emulation to run "The Intel Fortran Compiler
> >> 9.0" on "a system based on an AMD Athlon* or AMD Opteron*
> >> processor." Where there issues? How was performance?
> >> 
> >No, but if you know something that's using the intel's compiler, 
> >let us know! I'm very interested in this topic.
> 
> The intel site makes it look like you need Intel cpu, but thoes
> quotes are from Intel Fortran requirements, and I think it will
> work on AMD, plus Linux emul for NetBSD.
> 

We've run older versions on amd systems and they have worked fine.  The
resulting code (no processor specific switches set though) has also worked
fine.

You might want to think hard about g95.  The Intel compiler is likely
to produce faster code, of course, but g95 has gotten really good 
recently.  We also find that:

- it requres less patches than the intel compiler to deal with
odd corners of F90/F95.

- it's way less of a pain to install.

- you don't have a bizzaro website to navigate to get updates.

- the executables are much smaller.

oneel@isdclin6:~ > ls -hl 
/isdc/integration/osa_int/current_linux/bin/ii_skyimage
-rwxr-xr-x    1 oneel    obs_gen      7.9M Jan 11 16:02 
/isdc/integration/osa_int/current_linux/bin/ii_skyimage
oneel@isdclin6:~ > ls -hl $ISDC_ENV/bin/ii_skyimage
-rwxr-xr-x    1 isdc_cms obs_gen       23M Nov 29 14:38 
/isdc_sw/osa_sw/osa_sw-5.1/gcc/3.2.2/bin/ii_skyimage
oneel@isdclin6:~ > 


The bigger one is built with ifc.  Plus the ifc one requires fortran specific 
shared libraries.  Oh, and the ifc one is stripped and the g95 one isn't.

We've not seen huge differences between the generated code.  That said,
we don't set cpu specific optimization flags, nor has it been carefully 
measured.  Ie, we could have g95 be say 20% slower and we probably wouldn't
notice it.    We can't easily use CPU specific optimizations since our cluster
ends up not completely homogeneous.  Ie, we have some xenons, some amd systems,
and now we're getting some amd dual core systems.  

cheers

bruce


-- 
edoneel%sdf.lonestar.org@localhost
SDF Public Access UNIX System - http://sdf.lonestar.org