[CP2K:2670] compile the latest version of cp2k with g95

[zh]

Dear all,

I tried to compile the latest version (e.g. 2.1.223) of cp2k with g95,
scalapck (0.96), openmpi-1.4.1. The g95 installed on my computer
cluster is the 0.92 version built on Linux x86_64/EMT64 (32 bit
D.I.). The fftw3, openmpi, and scalapack are compiled with this
version of g95 and installed on my machines.The following arch file
was used for compiling cp2k:
CC = cc
CPP = /usr/bin/cpp
FC = /work/z-hou/soft/openmpi-g95/bin/mpif90
LD = /work/z-hou/soft/openmpi-g95/bin/mpif90
AR = ar -r
DFLAGS = -D__G95 -D__FFTSG -D__parallel -D__BLACS -D__SCALAPACK -
D__FFTW3
G95_INC = /work/z-hou/soft/fftw3-g95/include
CPPFLAGS = -C -traditional $(DFLAGS) -I$(G95_INC)
FCFLAGS = $(DFLAGS) -I$(G95_INC) -ffree-form -fno-underscoring -O1 -
ffast-math -cpp -g -fbounds-check -ftrace=full -funroll-loops
LDFLAGS = $(FCFLAGS)
G95_LIB = /work/z-hou/scalapack_installer_0.96/lib
LIBS = $(G95_LIB)/libscalapack.a \
$(G95_LIB)/blacsC.a \
$(G95_LIB)/blacsF77.a \
$(G95_LIB)/blacs.a \
$(G95_LIB)/libreflapack.a \
$(G95_LIB)/librefblas.a \
/work/z-hou/soft/fftw3-g95/lib/libfftw3.a

OBJECTS_ARCHITECTURE = machine_g95.o

The binary module of cp2k (i.e., cp2k.popt) is successfully obtained
with the above arch file. But when I run it in parallel manner, the
following errors were met:

[pcc-hs36:16109] *** Process received signal ***
[pcc-hs36:16109] Signal: Segmentation fault (11)
[pcc-hs36:16109] Signal code: (128)
[pcc-hs36:16109] Failing at address: (nil)
[pcc-hs36:16109] [ 0] /lib64/libpthread.so.0 [0x382700e7c0]
[pcc-hs36:16109] [ 1] /work/z-hou/soft/openmpi-g95/lib/libmpi.so.0
[0x2b66ad6d1c9b]
[pcc-hs36:16109] [ 2] /work/z-hou/soft/openmpi-g95/lib/libmpi.so.
0(PMPI_Group_incl+0x172) [0x2b66ad6f87d2]
[pcc-hs36:16109] [ 3] /work/z-hou/bin/cp2k.sopt(BI_TransUserComm+0x52)
[0x34957e2]
[pcc-hs36:16109] [ 4] /work/z-hou/bin/cp2k.sopt(blacs_gridmap+0x10d)
[0x348dc0d]
[pcc-hs36:16109] [ 5] /work/z-hou/bin/cp2k.sopt(blacs_gridinit+0xbb)
[0x348dadb]
[pcc-hs36:16109] [ 6] /work/z-hou/bin/
cp2k.sopt(cp_blacs_calls_MP_cp_blacs_gridinit+0x4c) [0x2cc830c]
[pcc-hs36:16109] [ 7] /work/z-hou/bin/
cp2k.sopt(cp_blacs_env_MP_cp_blacs_env_create+0x542) [0x1b1984a]
[pcc-hs36:16109] [ 8] /work/z-hou/bin/
cp2k.sopt(qs_environment_types_MP_init_qs_env+0x1098) [0xb1ef24]
[pcc-hs36:16109] [ 9] /work/z-hou/bin/
cp2k.sopt(qs_environment_types_MP_qs_env_create+0x122) [0xb21b4d]
[pcc-hs36:16109] [10] /work/z-hou/bin/
cp2k.sopt(qs_main_MP_quickstep_create_force_env+0x12b) [0xcff143]
[pcc-hs36:16109] [11] /work/z-hou/bin/
cp2k.sopt(f77_interface_MP_create_force_env+0x1c15) [0x4f515b]
[pcc-hs36:16109] [12] /work/z-hou/bin/cp2k.sopt(cp2k_runs_MP_cp2k_run
+0x1b83) [0x4afc53]
[pcc-hs36:16109] [13] /work/z-hou/bin/cp2k.sopt(cp2k_runs_MP_run_input
+0x24d) [0x4bab8b]
[pcc-hs36:16109] [14] /work/z-hou/bin/cp2k.sopt(MAIN_+0x154b)
[0x4ad1cb]
[pcc-hs36:16109] [15] /work/z-hou/bin/cp2k.sopt(main+0xe) [0x366f97e]
[pcc-hs36:16109] [16] /lib64/libc.so.6(__libc_start_main+0xf4)
[0x382641d994]
[pcc-hs36:16109] [17] /work/z-hou/bin/cp2k.sopt [0x4abbe9]
[pcc-hs36:16109] *** End of error message ***
--------------------------------------------------------------------------
mpirun noticed that process rank 0 with PID 16109 on node pcc-hs36
exited on signal 11 (Segmentation fault).
--------------------------------------------------------------------------

From searching the solution of the similar problem on internet, it
seems to be related to how fortran compiler handles the 32 bit and 64
bit in mpi. But I did not find out the exact solution.

I also tried the 0.92 version of g95 built on Linux x86_64/EMT64 (64
bit D.I.). The same errors were met. I also suffered this terrible
problem when I used the intel fortran compiler on a same machine and
the pgi fortran compiler on a CARY-XT5 machine to compile the latest
version of cp2k. If anyone can give me some helps or hints to fix this
problem, I will greatly appreciate him.

--
You received this message because you are subscribed to the Google Groups "cp2k" group.
To post to this group, send email to cp...@googlegroups.com.
To unsubscribe from this group, send email to cp2k+uns...@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/cp2k?hl=en.

[Axel]

On Apr 21, 8:43 am, zh <valen...@gmail.com> wrote:
> Dear all,
>
> I tried to compile the latest version (e.g. 2.1.223) of cp2k with g95,
> scalapck (0.96), openmpi-1.4.1. The g95 installed on my computer
> cluster is the 0.92 version built on Linux x86_64/EMT64 (32 bit
> D.I.).  The fftw3, openmpi, and scalapack are compiled with this
> version of g95 and installed on my machines.The following arch file
> was used for compiling cp2k:

as you can see from the stack trace, the error
is coming from openmpi and thus unlikely to be
compiler dependent.

what you see on the XT5 is very likely a different issue.

axel.

[Hou Zhufeng]

Dear Alex,

Thank you very much for your reply.

In the source codes of cp2k "message_r_passing.F", there is a comment as follows:
"
    ! requires that your MPI installation matches your Fortran compiler... be aware
    ! e.g. on a 64 bit machine, NAG will require MPI_ADDRESS_KIND=4, and gfortran MPI_ADDRESS_KIND=8
"
Are the error messages related to the "MPI_ADDRESS_KIND"?  How can we make MPI_ADDRESS_KIND=8 when installing openmpi with g95?
Actually, when the "message_passing.f90" and "dbcsr_message_passing.f90" were compiled with openmpi and g95, some warning messages appeared:
for example:
    Included at dbcsr_message_passing.f90:1822

    CALL MPI_WIN_CREATE(range, winsize, data_type_size,&
                        2
Warning (155): Inconsistent types (COMPLEX(8)/INTEGER(4)) in actual argument lists at (1) and (2)


Zh

[Hou Zhufeng]

Dear axel,

Sorry for spelling your name wrong in my last reply.

After trying many times for the compilation and installment of scalpack, openmpi, and cp2k with g95, finally I obtained an useable cp2k.popt. The problem comes from the compilation of BLACS with g95.  After removing "TRANSCOMM = -DCSameF77" in the Bmake.inc and recompiling BLACS, the problem is solved.  If "TRANSCOMM = -DCSameF77" is added into "Bmake.inc, it will lead to a warning message when BI_MPI_F77_to_c_trans_comm.c in BLACS is compiled.

Cheers

Zh