[CP2K:3299] regression?

[Urban Borštnik]

Dear Alin,

I ran your input with a gfortran-compiled version of the latest cvs (the
2.2 development branch) with 24 and 48 processes and could not reproduce
a crash (with bounds checking).

The first thing I would try is to either reduce the optimization level
for all files or selectively for "distribution_1d_types" and
"distribution_methods.F" (in the same way as for "graphcon.F"--take a
look at your arch file).

The other thing is to get a stack trace to see where the segmentation
fault occurs...

Cheers,
Urban

On Fri, 2011-06-17 at 20:32 +0100, Alin Marin Elena wrote:
> Hi All,
> 
> I have just compiled cp2k, the last cvs and I discovered a very strange 
> behaviour...
> I have a small system, 23 atoms and I want to perform some simple DFT 
> calculation. I try to run it in parallel.
> If I use 12 cores I get my numbers... (HF12good)
> If I use 24 cores I get a segmentation fault... (HF24broken)
> If I use 23 cores I get my numbers... (HF23good)
> 
> When I use an older version of cp2k few months...
> If I use 24 cores I get my numbers... (HF24good)
> 
> for all the successfull runs I got the same energy
> alin at stokes2:~/playground/alin> grep "ENERGY| " HF*good/HF.out 
> HF12good/HF.out: ENERGY| Total FORCE_EVAL ( QS ) energy (a.u.):             
> -146.062369186077234
> HF23good/HF.out: ENERGY| Total FORCE_EVAL ( QS ) energy (a.u.):             
> -146.062369186076523
> HF24good/HF.out: ENERGY| Total FORCE_EVAL ( QS ) energy (a.u.):             
> -146.062369186077234
> 
> Inputs and outputs for all tests can be found in the attachement (or link,if 
> the list rejects it)... to make the attach smaller I removed the .cube files.
> https://rapidshare.com/files/1317606759/hftests.tar.bz2
> 
> All the binaries are compiled with the same settings compilers and libs
> alin at stokes2:~/playground/alin/HF24broken> module list
> Currently Loaded Modulefiles:
>   1) intel-mkl/10.2.6.038   2) intel-fc/2011.1.107    3) intel-cc/2011.1.107    
> 4) mpt/2.01 (sgi mpi implementation)
> 
> alin at stokes2:~/playground/cp2k> cat arch/stokes-intel.popt 
> # this works on my setup on abbaton and baphomet
> # intel toolchain/mkl/openmpi... parallel threaded and optimised
> #
> #
> CC       = icc
> CPP      =
> FC       = ifort 
> LD       = ifort 
> AR       = ar -r
> DFLAGS   = -D__INTEL -D__FFTSG -D__parallel -D__BLACS -D__SCALAPACK -D__FFTMKL 
> -I$(MPI_ROOT)/include
> CPPFLAGS =
> FCFLAGS  = $(DFLAGS)  -O3 -xSSE4.2 -heap-arrays  -fpp -free
> FCFLAGS2 = $(DFLAGS)  -O1  -heap-arrays  -fpp -free
> LDFLAGS  = $(FCFLAGS) -L{$MKLROOT}/lib/em64t
> LIBS     = -lmkl_scalapack_lp64 -lmkl_intel_lp64 -lmkl_sequential -lmkl_core -
> lmkl_blacs_sgimpt_lp64 -lmpi
> 
> OBJECTS_ARCHITECTURE = machine_intel.o
> 
> 
> graphcon.o: graphcon.F
>         $(FC) -c $(FCFLAGS2) $<
> 
> 
> 
> Any suggestion much appreciated,
> 
> regards,
> 
> Alin
> 
> 
> 
> 
> 
>