AMBER Archive (2005)

Subject: Re: AMBER: compiler for amber8 on opteron

From: Robert Duke (rduke_at_email.unc.edu)
Date: Fri Mar 18 2005 - 07:11:22 CST


Marc -
I had the same concerns about the intel compiler working well with opterons,
but with the introduction of their 64 bit chips (EM64 stuff, not itanium),
they have ended up producing a compiler that is roughly equivalent to the
pgi compiler in performance, based on some benchmarking I did at Utah on
pmemd. Now, there is another compiler, rather new, that is being very well
received for use on opterons; of the two generic opteron cluster facilities
I have access to, they are both using it as their primary compiler. In my
hands, it was faster and easier to work with. This is the pathscale
compiler, pathf90. I posted a couple of long emails to the reflector, first
about benchmarking pmemd on the utah clusters with pathscale, and then I
released some updated configuration info for a pmemd (new_configure)
utility (the actual config source is available as a link from
amber.scripps.edu) that included the opteron and also included some
comparison between ifort, pgi, and pathscale. You might want to look on the
reflector and web site for this info. I need to get Dave to post the
benchmarks on the website so everyone can see what the opterons can do. I
did not originally respond to your mail because I have not evolved
configuration support for the entire amber package for the opteron, and I
know the original release there included pgi on the opteron but not ifort
(where you should get the EM64 version) or pathscale. I would suggest you
can adapt from the pmemd config.h to set up the overall config.h, however.
Be aware that there is a sander optimization problem with ifort, whereby the
extra points code does not work properly if you optimize sander to the same
level that pmemd is optimized (one of the reasons I retained separate
configuration for pmemd).
Regards - Bob Duke

----- Original Message -----
From: "Marc Petitmermet" <petitmermet_at_mat.ethz.ch>
To: <amber_at_scripps.edu>
Sent: Friday, March 18, 2005 3:22 AM
Subject: Re: AMBER: compiler for amber8 on opteron

>> Performance on Opteron
>> ifort >> pgi >> gcc
>
> I never thought of the Intel compilers because, well, doesn't Intel
> compilers optimize for Intel CPUs only? You seem to have experience with
> these different compilers, therefore, I will proceed as you suggested.
> Thanks for the advice.
>
> Regards,
> Marc
>
> -----------------------------------------------------------------------
> The AMBER Mail Reflector
> To post, send mail to amber_at_scripps.edu
> To unsubscribe, send "unsubscribe amber" to majordomo_at_scripps.edu
>

-----------------------------------------------------------------------
The AMBER Mail Reflector
To post, send mail to amber_at_scripps.edu
To unsubscribe, send "unsubscribe amber" to majordomo_at_scripps.edu