[CPMD-list] trying to confirm correct compilation of cpmd binary

[Axel Kohlmeyer]

On Fri, 27 May 2005, Isaac Tamblyn wrote:

IS> Hello,

IS> 
IS> I'm having trouble determining whether or not the binary I've compiled is
IS> working properly. 

isaac,

it is very commendable, that you check your binaries first.
as simple as it is, we've seen a lot of cases, where people
just start with their most complicated system right away...

IS> 3 questions: 
IS> 
IS> 1.	The program finishes by saying "FORTRAN STOP", is this normal?

yes.

IS> 2.	I used the files in CPMD-test/linear-scaling/ to try and test my
IS> binary. The readme (and output from the test file) both claim that the run
IS> should take approximately 5 min. Mine took ~ 70. I used the file
IS> libatlas_x86_64.a from Axel's website.

please have a closer look at the reference output. the reference run
was done on a parallel machine with 32 processors (search for the 
PARAPARA... block). if you account for the fact that a parallel binary
has roughly 10% overhead and assume ideal scaling, your cpu is about
2.5 times as fast than those of the reference machine...

IS> 3.	Also, when I compare the output I get for several of the test cases,
IS> I find that there can be significant differences. I include as an example
IS> both my output and the reference output for the test linear scaling inp-1.
IS> Is this something that I should be worried about, or is this just a
IS> difference due to the fact that I'm using 3.9.2 and the test cases were for
IS> an earlier version?

well, there are several possible sources of differences in the outputs.
- a bug was found and fixed in the newer cpmd release (this will
   for example apply to open shell calculations of systems with the PBE 
   functional and a significant 'vacuum region'). quite a few bugs have
   been fixed since cpmd-v3.7.1.
- you have to be careful comparing certain values. the accuracy of 
   some properties, e.g. forces, energies, depend on the convergence, 
   but also on the fact of whether you are using identical grids or
   you are running a parallel or serial job.
- also there have been corrections to the calculation of the 
   electrostatic energy (with no difference in the forces, IIRC).
- finally, also aggressive optimization and different numerical
   libraries can result in small differences. with heavy numerical
   codes you always have to find a compromise between speed and
   accuracy. there are however some compilers, that miscompile
   some parts of cpmd (resulting in large differences). there was
   a report about this in combination with the PGI compiler on x86_64. 
   please make sure, that you have at least version 6.0, which is 
   supposed to have this bug fixed.

IS> I used the PGI compiler, as I had trouble similar to someone else on this
IS> list when I tried using the Intel fortran compiler (the program would crash
IS> at subroutine fileopen). The system I'm using this on is an AMD Opteron
IS> platform. (I've also tried it on a P4, and run into similar problems.I used
IS> the library libatlas_p4 in that case).

please try the workaround i just sent to cpmd-list.

regards,
	axel.

IS>  
IS> 
IS> Thanks in advance,
IS> 
IS>  
IS> 
IS> Isaac
IS> 
IS>  
IS> 
IS>  
IS> 
IS> 

-- 

=======================================================================
Dr. Axel Kohlmeyer   e-mail: axel.kohlmeyer at theochem.ruhr-uni-bochum.de
Lehrstuhl fuer Theoretische Chemie          Phone: ++49 (0)234/32-26673
Ruhr-Universitaet Bochum - NC 03/53         Fax:   ++49 (0)234/32-14045
D-44780 Bochum  http://www.theochem.ruhr-uni-bochum.de/~axel.kohlmeyer/
=======================================================================
If you make something idiot-proof, the universe creates a better idiot.