[Wien] problem in l2main.F (parallel version) for large systems

[Valerio Bellini]

Dear all,

I am reporting here something which I call a 'problem' in the l2main 
subroutine,
since I am not sure whether it happens systematically to be named 'bug'.
I am running a very large system (>12000 APW's), and use the mpi
parallelized version by allocating many (32) processors on a
IBM SP RS/6000Power 4 multiprocessor/multinode machine compiling
with the (mp)xlf90 standard IBM Fortran compiler.
When the option -O3 is used to compile the lapw2 program, it happens
somehow random that in the blocked loop in l2main.F routine, some or all
the ALM and BLM coefficients attain NaNq values.
This happens for random value of the inequivalent atoms.
Writing NaNq values both in the charge and potentials files, i.e. 
.clmval(up/dn)
and .vrespval(up/dn) files, the code substitute them with zero's when it
tries to read them later, and loose for some atoms the charge, spoiling
obviously all the self-consistent procedure.
This problem is avoided by imposing the compiler to keep the semantic
of the program, i.e. by using the option -qstrict in addition to the -O3 
option
in the compilation step. Probably something goes fishy in the blocked
loop cited above, when the compiler tries to optimize it by inverting 
the order
of the loops, or maybe there is some variable that should be initialized
(for instance the 'alm' and 'blm' matrices themselves?)
The use of -qstrict does not deteriorate the speed-up given
by the -O3 option to the lapw2 program, so that it can be used without any
loss of performance.
The use of -qstrict might slow down the the lapw1 program (I didn't try),
but since everything goes well for lapw1, I suggest to  recompile only 
the lapw2
program with the -qstrict option.
I am wondering whether some of you has encountered such a problem already.
I have to mention here that in a similar calculation, on a different 
system and
allocating a lower number of CPU's, this problem did not show up!
Best regards,
Valerio

-- 

*******************************************************************************
  Valerio Bellini
  INFM-S3 National Research Center on nanoStructures and bioSystems at Surfaces
  and Department of Physics, University of Modena and Reggio Emilia
  Via Campi 213/A, 41100 Modena, Italy.
  Phone:   ++39 059 2055301
  Fax:     ++39 059 374794
  E-mail:  bellini.valerio at unimore.it
  WWW:     http://www.s3.infm.it
*******************************************************************************

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://zeus.theochem.tuwien.ac.at/pipermail/wien/attachments/20041122/98e03db0/attachment.html