The ABINIT Users Mailing List ( CLOSED )

["D. R. Hamann" <drhamann@mat-simresearch.com>]

Dear all,

I encountered two bugs in 5.2.4 while attempting to run the first GW 
tutorial on a parallel system. The first is in the routine 
08seqpar/outkss.F90 and is serious but subtle. The second is in 
11drive/screening.F90 and is only a nuisance, printing arrays of NaN's 
to the log file. The attached tarball includes patches to fix both, and 
a simple modification of tests/v4/t84.in which will make the serious bug 
happen when executed in parallel (if you want to convince yourself). A 
more detailed explanation of the bugs (for developers) follows. This 
text included in the tarball as a README.

The outkss bug looks like a typo in line 1186. It is subtle because 
depending on the value of the nbandkss input variable, parallel 
execution may give correct results, or may not. (I'll leave the reasons 
for this as an exercise for the interested reader.) Sequential execution 
or execution with abinip specifying one process will give correct 
results in any case.

The screening bug results from the fact that a block of code is executed 
in all processes using data in an array that is only initialized in 
process 0, the “master” process. While most printing by other processes 
had been expunged from this section of code, there is hidden printing in 
the routine matcginv (through the call to printcm, which probably should 
have been inside debug flags). My patch keeps the whole block of code 
from executing in all but process 0, although I could have simply 
enclosed the matcginv calls in appropriate if's.

Don Hamann

--
D. R. Hamann
Mat-Sim Research LLC    | Deptartment of Physics
P.O. Box 742            |  and Astronomy
Murray Hill, NJ 07974   | Rutgers University
phone: 908-370-8079     | 732-445-4381

email: drhamann@mat-simresearch.com

Attachment: drh_gw.tgz
Description: application/compressed