AMBER Archive (2007)

Subject: Re: AMBER: AMBER8 compiler with intel 9.1 compiler

From: Scott Brozell (sbrozell_at_scripps.edu)
Date: Wed May 16 2007 - 13:42:06 CDT


Hi,

On Tue, 15 May 2007, Peter Z. Qin wrote:

> I am running AMBER8 on our school mainframe (Linux cluster). Recent changes
> at this facility forces me to re-compile AMBER8 with an Intel 9.1 compiler.
> While my previous compilation of AMBER8 using an Intel 8.1 compiler passed
> all tests, the re-compilation (both the serial and the parallel version)
> fails testing on several cases. A typical example, generated from
> test/cytosine/Run.cytosine, is:
>
> 87,88c87,88
> < NSTEP = 1 TIME(PS) = 5420.002 TEMP(K) = 301.21 PRESS -5.8
> < Etot = -608.3990 EKtot = 1063.0604 EPtot = -1671.459
> ---
> > NSTEP = 1 TIME(PS) = 5420.002 TEMP(K) = 301.68 PRESS -651.0
> > Etot = -606.7460 EKtot = 1064.7135 EPtot = -1671.459
> 92c92
> < EKCMT = 176.9122 VIRIAL = 180.9720 VOLUME = 32637.922
> ---
> > EKCMT = 176.9122 VIRIAL = 635.6568 VOLUME = 32637.922
> 94c94
> < Ewald error estimate: 0.1512E-1
> ---
> > Ewald error estimate: 0.6367E+1
>
> My question is whether these differences are significant? and iff so, whar
> could be the fix?

These are sig. diffs.
What is the outpout of ifort -V ?
What config file was used ?
What are the machine details ?

The simplest fix is to use an older intel compiler.
I stay away from the latest from intel because I have found their
quality assurance to be lacking.
On the machines I use the latest versions are
ia32 intel-compilers-9.0-20050916
ia64 intel-compilers-9.1-20070111
but I have not built amber with these versions (using 8.1 instead).
I could not quickly find details about which intel versions
have been used to build Amber at the web site.
Other posters may be helpful here.

The likely cause is changes in code meaning due to compiler
optimizations. If you cannot use an older ifort then one
could reduce the optimization level. There are clues in the diffs
above, and more info about the other failing tests might enable
someone knowledgeable about Amber's software architecture
to identify a small number of files that have been affected.

Scott

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