Re: MPICH-GM 1.2.5..9.

"Christopher D. Maestas" <[email protected]> Tue, 25 Feb 2003 13:20:23 -0700 (MST)
Newsgroups gmane.network.myrinet.general
Message-ID <[email protected]>
Chris,

Well we were interested in seeing how a compiler suite performs end to
end rather than in a hybrid approach.  I ran some initial hpl compiling with
a hyrbid approach, and there were some issues in the link stages already.  I can
see the headaches already from the user, to help desk, to maintainer ... ;-)

I still maintain that on an end to end approach, the intel compiler suite would
give better performance than the gcc suite, especially when trying to take
advantage of Itanium2 and XEON intructions sets.  So with that in mind, there is
still an advantage in compiling mpich-gm with the intel compilers over gcc.

Regards,

On Tue, 25 Feb 2003, Christopher Jordan wrote:

> 
> Yeah, you've got to test a hybrid, otherwise you're not really testing 
> mpich-gm performance with a given compiler, you're testing the compiler's 
> optimization of the code being compiled.  
> 
> I assume that mpich-gm supports the use of the $CC variable and the 
> $CC-conf facility of the newer mpiches? These allow you to set up 
> multiple compilers to link with one version of mpich (for any device, 
> AFAIK).
> 
> On Mon, 24 Feb 2003, Christopher D. Maestas wrote:
> 
> > Patrick,
> > 
> > In trying to keep things as simple as possible we used:
> > 	mpicc
> > which called icc, that linked to the mpich library compiled with the intel compilers.
> > 
> > And the same for gcc,
> > 	mpicc -> gcc -lmpich (compiled with gcc)
> > And the same for pgi,
> > 	mpicc -> pgcc -lmpich (compiled with pgi suite)
> > 
> > I haven't tried a hybrid or can't remember trying it :-) 
> > 	mpicc -> icc -lmpich (compiled with gcc)
> > 
> > I can try it ... or rather put it on the list to try ...
> > 
> > Regards,
> > 
> > On 24 Feb 2003, Patrick Geoffray wrote:
> > 
> > > Hi Christopher,
> > > 
> > > On Mon, 2003-02-24 at 13:45, Christopher D. Maestas wrote:
> > > > Yes there is a performance benefit in using mpich-gm with intel versus gcc.
> > > > mpich-gm compilations with the intel compiler are comparable to mpich-gm
> > > > compilations with the pgi compilers.  When testing hpl runs with the differently
> > > > compiled mpich-gm releases, the intel and pgi compiled versions gave better results
> > > > than the gcc compiled version.
> > > 
> > > Do you compile HPL with gcc or with the Intel compiler when linking with
> > > MPICH-GM compiled with gcc ?
> > > 
> > > Patrick
> > > 
> > 
> > --  Chris
> > 
> > 
> > 
> > _______________________________________________
> > Myrinet mailing list
> > [email protected]
> > http://email.osc.edu/mailman/listinfo/myrinet
> > 
> 
> 

-- Chris