ABINIT Discussion Forums

Hello abiniters of the world,

running certain response function phonon calculations in parallel and with mkmem 0 (or mk1mem 0) makes the program segfault (up to and including 6.11 development version). The input below will do it, at least on gfortran 4.5 on mac OSX and also on scientific linux with gfortran 4.3

with 8 kpt we should be able to use at least 2 or 4 processors. 2 goes fine, but using 4 with just mkmem 0 (or just mk1mem 0) gives the following on LOG_0002 (needless to say we are not spin polarizing):

>>>
-P-0002
-P-0002 kpgsph.F90:124:BUG
-P-0002 The variables ikg, mkmem, and mpw must satisfy ikg<=(mkmem-1)*mpw,
-P-0002 while the arguments of the routine are
-P-0002 ikg =182, mkmem =2, mpw =113
-P-0002 Probable cause: Known error in invars1 for parallel spin-polarized case.
-P-0002 Temporary solution: Change the number of parallel processes.
-P-0002
<<<

The effective number of k-points is 5, so this is probably once again a problem with abinit packing the processes too well and leaving a processor empty. However, it isn't the empty processor who is complaining, but the last full one (2) so there might be a distribution-of-work bug somewhere.

cheers

Matthieu







input file: to be de-commented. Add your favorite aluminium pseudopotential.
>>>
ndtset 2

#mkmem 0
#mk1mem 0
kptopt 3

getwfk2 1
rfphon2 1
rfatpol2 1 1
rfdir2 1 1 1
nqpt2 1
qpt2 0 0 0


#### common info
ngkpt 2 2 2
nshiftk 1
shiftk 0 0 0

nband 5
tsmear 0.001
occopt 7
tolvrs 1.e-8
nstep 100
ecut 2
ntypat 1
znucl 13
typat 1
xred 0 0 0
natom 1
acell 10 10 10
angdeg 60 60 60
<<<