The ABINIT Users Mailing List ( CLOSED )

["Allan, Douglas C Dr" <AllanDC@corning.com>]

dekoning:
I am afraid I do not know what the problem is, but the symptom is
reporting something that ordinarily should not happen, and that is why
the code reports a warning.  If the line minimization is proceeding
correctly, the new trial energies naturally should be lower than the
previous energies.  The line minimizations are nearly an innermost loop
of the conjugate gradient optimization.  I think they can occasionally
go uphill but should not repeatedly go uphill.  Your final result is a
clear indication of a failure of the minimization upon comparing it with
the correct answer.
Someone closer to recent versions of the code will have to comment on
why this might happen.  Most likely there is a bug somehow in the
calculation of <G|H|C>, i.e. the gradient, or in its renormalization
into the projected gradient (conjugate gradient) that is used to form
the line minimization step.  Or maybe a more recent programmer has a
better guess.
If the gradients were routinely in error then the whole code would not
work, so whatever is wrong must be pretty isolated to your version and
build.  Can anyone comment who has experience on a similar machine?

Good luck,
Doug Allan

-----Original Message-----
From: dekoning@ifi.unicamp.br [mailto:dekoning@ifi.unicamp.br] 
Sent: Thursday, March 29, 2007 4:32 PM
To: forum@abinit.org
Subject: [abinit-forum] Problem running ABINIT on Altix

I just compiled ABINIT version 5.3.3 on an Altix system, based on the
Intel Itanium 2 processor using the Intel Fortran Compiler, version 9.0.
I did not use any special flags, just running "./configure" and "make".
It produced the executables (serial version to begin with) and then I
ran the test run #1 from the Tutorial
(http://www.abinit.org/Infos_v5.2/tutorial/lesson_1.html). It runs OK
and completes the job within a second or so. Then, as suggested in the
tutorial, I compared the output file t11.out to that in the Refs
directory. There should not be any large differences, but to my surprise
there are in fact very large differences between the two results. For
instance, the converged total energy of my run (9th iteration) is
-0.32750890496212 Hartree, while the reference result (6 iterations) is
-1.1037224213136 Hartree. One thing that I noticed is that there are 22
warnings in my log file. They all read: 


cgwf : WARNING: New trial energy at line is higher than former.


This seems to be something related to the conjugate-gradient
minimization.



To check I also built the same code on my own P4 Linux box using the
same compiler, also running only "./configure" and "make". For this
machine there is no problem, and  test run #1 gives the same results as
the reference file in the directory /Refs. 

Any idea what the problem on the Altix system might be?