AMBER Archive (2008)

Subject: Re: AMBER: Amber 9: Question regarding Test Failures and Compilers

From: David A. Case (case_at_scripps.edu)
Date: Wed Jan 23 2008 - 10:22:04 CST

On Wed, Jan 23, 2008, Jan Schulze wrote:
>
> I'm trying to build the serial version of Amber 9 on Linux. I tried on
> x86_64 and x86_32 architectures with gfortran and g95 compilers and I
> applied the newest bug fixes before compilation. Compilation works on
> both architectures without errors but with quite a lot of warnings.
>
> Running the tests gives various test failures. On x86_64 the tests
> produce big deviations from the expected results, like for example:
>
> possible FAILURE: check cytosine.out.dif
> /nfs/wsi/bs/share/opt/x86_64_sl4/amber9/test/cytosine
> 99,100c99,100
> < NSTEP = 2 TIME(PS) = 5420.004 TEMP(K) = 300.46 PRESS
> = -32.0
> < Etot = -608.0823 EKtot = 1060.4055 EPtot =
> -1668.4878
> ---
> > NSTEP = 2 TIME(PS) = 5420.004 TEMP(K) = 300.49 PRESS
> = 33.2
> > Etot = -664.1291 EKtot = 1060.4894 EPtot =
> -1724.6185
> 103,104c103,104
> < EELEC = 12.8250 EHBOND = 0. RESTRAINT = 0.
> < EKCMT = 176.8830 VIRIAL = 199.4654 VOLUME =
> 32637.8239
> ---
> > EELEC = -43.3058 EHBOND = 0. RESTRAINT = 0.
> > EKCMT = 176.8830 VIRIAL = 153.4667 VOLUME =
> 32637.8239

This is indeed worrisome, and not at all what I get for the g95 x86_64
combination.

>
>
> On x86_32, I also get various test failures, however deviations are
> smaller. Failures are found in many of the tests, but most frequent in
> the pbsa_radi tests:

The pbsa_radii tests you show are all diffs between a true zero and something
like 10**-34 Ang. This must have to do with moving underflows to zero, and
are not really a concern.

>
> I used the "Goto" optimized blas library and tried with and without
> platform option compiler flags. Are there any other things, I could try?

Don't use Goto? It is certainly possible that they could cause a problem if
they were somehow not built correctly.

You might report the results of "g95 --version", and give the exact flags you
gave to configure.

> Maybe using a portland compiler would be more promising than gfortan/g95?

We've had more problems with portland and g95, not fewer.

....dac

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