On Tuesday 24 May 2005 06:09, Jason Bradley Nance wrote:
> Hello everyone,
> I was wondering what everyone's experience has been with mixing
> Intel's EM64T and AMD's Opteron in a clustered environment. Does
> commercial software like lsdyna have any problems? What about compilers
> such as PGI and Intel? I've had troubles getting the Intel compilers to
> even produce 64-bit binaries on Opterons.
>
> I know that the processors are supposed to be basically the same, but
> I'm sure that if there are any little inconsistencies we are sure to
> find them in the HPC world.
The problem with a homogeneous EM64T and Opteron environment is the
differences in compiler optimizations required to get best performance on
each. LS-DYNA for EM64T is compiled with an option which prevents it from
running on Opteron. It is not too difficult technically to produce a build
with ifort which should perform within 5% of optimum on both. That is
unlikely to happen, with the pressures to produce best performance with
separate releases. The Opteron release will run on EM64T with a 10 to 15%
performance penalty. The best argument you could give for a single build
supporting both, at a small loss in performance, would be the reduced burden
of QA.
I don't know what you mean about producing 64-bit binaries. Are you using the
l_fce_... ifort 8.1, which makes x86-64 code? If you use the pure 32-bit
compiler, you get 32-bit binaries. To run on both Opteron and EM64T, you
normally compile with the -xW option. If you want both the 32-bit and x86-64
ifort (along with ia64) in the same humongous tar file distribution, that is
coming with ifort 9.0 in a month.
--
Tim Prince
|
