Re: AMBER: AIX 5 test failures: -NaNQ

From: David A. Case <case.scripps.edu>
Date: Thu, 19 Apr 2007 18:13:46 -0700

On Thu, Apr 19, 2007, Patrick McCarren wrote:

> AMBER 9 (patched up to bug 34) compiled with no errors on an IBM P655+
> machine running AIX 5.2.0.0 with xlf enterprise edition 10.1 at ARSC,
> but some tests failed: crambin_divcon, 2pk4_stan, amoeba_jac, tp.mol2,
> ash.mol2, sustiva.mol2, fluorescein.mol2, frcmod, DGN.mol2 with NaNQ
> values. We didn't have this problem on our commodity intel and opteron
> machines. See http://www.chem.ucla.edu/~mccarren/TEST_FAILURES.dff.txt

Thanks for the report. The above URL doesn't work for me, although that is
not real important.

Scott had suggestions about how to track this down. You might also consider
using a lower optimization level, at least to see what changes.

Everything that you mention above (except amoeba_jac) relies on divcon, and
that is known to be the most non-portable part of Amber. This doesn't excuse
anything or anyone, but just says that I'm not surprised (too much) that
a AIX compiler might either give bad code, or expose bugs. Even Intel's
compilers (which are the ones most tested by the divcon folks) sometimes
have troubles (see, for example, bugfix.21). [Basically, divcon has lots
of constructs that are frustrate optimizing compilers, sometimes to the point
of having them emit bad code. But some things may also be code errors, where
other compilers are more forgiving than they need to be to satisfy the
language standard.]

You might not have plans to run divcon as QM/MM inside sander, as that is a
pretty specialized task. If you need antechamber on this machine, you might
consider (as a workaround) downloading the antechamber from
http://amber.scripps.edu/antechamber). This "stand-alone" version uses mopac6
rather than divcon for the QM calculations, and maybe the IBM compilers will
be happier with that.

The amoeba_jac failure is probably something different, and any details you
have there may help us (although a very different implementation is being
prepared for Amber 10). Clearly, you don't need to worry if you are not
planning to run amoeba calculations on this machine. But it probably is
something that we might want to try to track down.

Thanks again for the report....dac

-----------------------------------------------------------------------
The AMBER Mail Reflector
To post, send mail to amber.scripps.edu
To unsubscribe, send "unsubscribe amber" to majordomo.scripps.edu
Received on Sun Apr 22 2007 - 06:07:37 PDT
Custom Search