[Wien] Support of 64-bit MPI addresses in LAPW0
Oleg Rubel
oleg.rubel at physik.uni-marburg.de
Tue Mar 3 19:57:52 CET 2009
Dear Wien2k Users and Developers,
I would like to share my experience in resolving a problem with MPI_ADDRESS in LAPW0 on 64-bit Intel Xeon. I use WIEN2k_08.3 (Release 18/9/2008) compiled with Intel Fortran (version 11.0) and MVAPICH2 1.2.0 2008-11-06. The calculation aborts in LAPW0 saying that "MPI_ADDRESS does not fit in Fortran integer". It is known that MPI_ADDRESS supports 32-bit addresses and the error occurs on 64-bit machines once the address does not fit 32-bit (http://publib.boulder.ibm.com/infocenter/clresctr/vxrx/index.jsp?topic=/com.ibm.cluster.pe432.mpisub.doc/am107_igadd.html). It is proposed to use MPI_GET_ADDRESS instead, which is compatible with 64-bit.
In my case, I get the error not every time, usually for relatively large (> 40 processes) MPI calculation. The problem have gone completely after I changed all instances of MPI_ADDRESS in lapw0.F to MPI_GET_ADDRESS and recompiled the parallel version of LAPW0.
Why not to change to MPI_GET_ADDRESS in the next release?
Sincerely,
Oleg
--
Oleg Rubel, PhD
Thunder Bay Regional Research Institute
290 Munro St
Thunder Bay, ON
P7A 7T1, Canada
Phone: +1-807-7663350
E-mail: rubelo at tbh.net
Homepage: http://www.staff.uni-marburg.de/~rubel/
More information about the Wien
mailing list