The ABINIT Users Mailing List ( CLOSED )

[mjv500@york.ac.uk]

Hello,

I am having some trouble with (at least) 2 parts of the parallelism in 5.3.2

1) On IBM pserver AIX/xlf90 09.01.0000.0007, when my system becomes larger 
(8x8x8 kpt instead of 4x4x4) it looks like abinit chooses band parallelism 
instead of kpoint. I always run on 32 nodes (standard queue). Then I get

ERROR: 0032-160 Too many communicators  (2046) in MPI_Comm_create, task 0

from each node. The last lines of the log are

getdim_nloc : deduce lmnmax  =   9, lnmax  =   3,
                     lmnmaxso=   9, lnmaxso=   3.
 distrb2: enter 
 mpi_enreg%parareel= 0
 mpi_enreg%paralbd= 1
 mpi_enreg%paral_compil_respfn= 0
 distrb2: exit 

I have tried to force the parallelism by kpt in the input with

npfft 1
npband 32

but this had no effect. Am I doing something wrong? Is the band parallelism 
well tested (eventually on this type of machine)? And is it really possible 
to disable it as I tried?

2) On the same platform as above, but also on max (MacOSX /xlf90 v9/ g5 
processors) I sometimes get a crash at the end of the scf for RF runs. In 
my latest case, three ddk calculations worked fine, but then, at the end of 
the first phonon perturbation:

At SCF step   13       vres2   =  1.20E-10 < tolvrs=  1.00E-09 =>converged.
-open ddk wf file :GaAsout_DS2_1WF7
-open ddk wf file :GaAsout_DS2_1WF8
-open ddk wf file :GaAsout_DS2_1WF9
-P-0000  leave_test : synchronization done...
-P-0000  leave_test : exiting...

and the code stops without comment or error. Similar happened on max with 
an initial successful GS run, then a converging RF which crashes without 
comment.

Have things changed in 5.3.3? Other known problems?

Any insight will be most welcome. I can debug too, but this kind of thing 
is probably beyond my skills.

Matthieu