Re: [moose-generic] MOOSE_installation_problem

[Subhasis R. <ray...@gm...>]

Dear Vadim,
 Did you run "make clean" before rebuilding each time? If you specify
USE_HDF5=0, then the entire content of HDF5 related files get excluded from
the compilation process. But from the build log for the first case it seems
the object files were compiled with HDF5 support but the linking is being
done without.
 In the second case, is it possible that your version of HDF5 library is
linked to a different MPI library than what you set in MOOSE Makefile?
Otherwise, it is also possible that HDF5 lib is dynamically linked to MPI
and when you are building MOOSE without MPI support the MPI-lib location is
not passed to compiler. In that case you may want to use mpicxx as the
compiler at line 250:

# Here we automagically change compilers to deal with MPI.
ifdef USE_MPI
    CXX = mpicxx
#    CXX = /usr/local/mvapich2/bin/mpicxx
#    PARALLEL_DIR = parallel
#    PARALLEL_LIB = parallel/parallel.o
else
    CXX = g++ # <-- replace with mpicxx
#    CXX = CC    # Choose between Solaris CC and g++ on a Solaris machine
endif


 For the third case, we use the command:

python-config --includes

in the Makefile in order to get the location of header files for numpy and
python. Looks like that does not work on your system. So you need to
explicitly add
-I/usr/include/python2.7 -I/usr/include/python2.7

and

-I{location of numpy/arrayobject.h}

to CXXFLAGS

and add the following to LIBS (output of python-config --ldflags) updated
according to local system paths:

-L/usr/lib/python2.7/config -lpthread -ldl -lutil -lm -lpython2.7 -Xlinker
-export-dynamic -Wl,-O1 -Wl,-Bsymbolic-functions


hth,
  Subha


On Wed, Mar 6, 2013 at 1:43 AM, Vadim A <ast...@nc...> wrote:

> Dear Subha.
>
> I was trying three different ways to install MOOSE: without HDF5, with
> HDF5 and with MPI
>
> Unfortunatelly, I did fail in all. Here the list of problems which we
> have. We would very appreciate you comments or suggestions what we might
> missing here.
>
>
>
> (1) First I did
>
> make BUILD=release USE_HDF5=0
>
>
>
> g++ -O3 -pthread -fpermissive -fno-strict-aliasing -fPIC -Wall
> -Wno-long-long -pedantic -DNDEBUG -DUSE_GENESIS_PARSER -I
> /home/diag/opt/openmpi/1.4.3/gnu/include  -DSVN_REVISION=\"4401M\" -DLINUX
> -DUSE_GSL -I /home/diag/opt/gsl/1.15/gnu/include/    basecode/_basecode.o
> msg/_msg.o shell/_shell.o biophysics/_biophysics.o hsolve/_hsolve.o
> randnum/_randnum.o scheduling/_scheduling.o builtins/_builtins.o
> device/_device.o kinetics/_kinetics.o ksolve/_ksolve.o
> regressionTests/_rt.o utility/_utility.o geom/_geom.o mesh/_mesh.o
> manager/_manager.o signeur/_signeur.o     -L/usr/lib -L/usr/local/lib
> -lpthread -L/lib64 -L/usr/lib64 -L/usr/lib
> -L/home/diag/opt/gsl/1.15/gnu/lib/ -lgsl -lgslcblas -lm -o moose
>
>
>
> And getting a lot of error like this:
>
>
> builtins/_builtins.o: In function `HDF5WriterBase::setMode(unsigned int)':
> (.text+0x11f40): undefined reference to `H5check_version'
> ……………….
>
> builtins/_builtins.o:(.data.rel.ro._ZTVN3MPI8DatatypeE[vtable for
> MPI::Datatype]+0x78): undefined reference to `MPI::Datatype::Free()'
> collect2: ld returned 1 exit status
>
>
>
>
>
>
> (2) Then
>
> We have installed  HDF5 1.8.9 and I was trying to do build release:
>
>
> make BUILD=release
>
>
> It seems that all libs get compiled and it can find HDF5 1.8.9
>
>
> All Libs compiled
> g++ -O3 -pthread -fpermissive -fno-strict-aliasing -fPIC -Wall
> -Wno-long-long -pedantic -DNDEBUG -DUSE_GENESIS_PARSER -I
> /home/diag/opt/openmpi/1.4.3/gnu/include  -DSVN_REVISION=\"4401M\" -DLINUX
> -DUSE_GSL -I /home/diag/opt/gsl/1.15/gnu/include/    -DUSE_HDF5
>  -DH5_NO_DEPRECATED_SYMBOLS -*I
>  /home/diag/opt/hdf5/1.8.9/gnu/openmpi/include* basecode/_basecode.o
> msg/_msg.o shell/_shell.o biophysics/_biophysics.o hsolve/_hsolve.o
> randnum/_randnum.o scheduling/_scheduling.o builtins/_builtins.o
> device/_device.o kinetics/_kinetics.o ksolve/_ksolve.o
> regressionTests/_rt.o utility/_utility.o geom/_geom.o mesh/_mesh.o
> manager/_manager.o signeur/_signeur.o     -L/usr/lib -L/usr/local/lib
> -lpthread -L/lib64 -L/usr/lib64 -L/usr/lib
> -L/home/diag/opt/gsl/1.15/gnu/lib/ -lgsl -lgslcblas -lm -lhdf5
> -L/home/diag/opt/hdf5/1.8.9/gnu/openmpi/lib    -o moose
>
>
>
> but it complain about a lot of "undefined references like this:
>
>
> builtins/_builtins.o:(.data.rel.ro._ZTVN3MPI8DatatypeE[vtable for
> MPI::Datatype]+0x78): undefined reference to `MPI::Datatype::Free()'
> collect2: ld returned 1 exit status
> make: *** [moose] Error 1
>
> Looks like we still missing something. Should we have exactly same HDF5
> library as you are using -> *libhdf5-serial   ?*
>
>
>
>
> (3)
>
> Also I did try to compile with MPI:
>
>
> make BUILD=release USE_MPI=1
>
> All Libs compiled
> mpicxx -O3 -pthread -fpermissive -fno-strict-aliasing -fPIC -Wall
> -Wno-long-long -pedantic -DNDEBUG -DUSE_GENESIS_PARSER -I
> /home/diag/opt/openmpi/1.4.3/gnu/include  -DSVN_REVISION=\"4401M\" -DLINUX
> -DUSE_MPI -DMPICH_IGNORE_CXX_SEEK -DUSE_GSL -I
> /home/diag/opt/gsl/1.15/gnu/include/    -DUSE_HDF5
>  -DH5_NO_DEPRECATED_SYMBOLS -I
>  /home/diag/opt/hdf5/1.8.9/gnu/openmpi/include basecode/_basecode.o
> msg/_msg.o shell/_shell.o biophysics/_biophysics.o hsolve/_hsolve.o
> randnum/_randnum.o scheduling/_scheduling.o builtins/_builtins.o
> device/_device.o kinetics/_kinetics.o ksolve/_ksolve.o
> regressionTests/_rt.o utility/_utility.o geom/_geom.o mesh/_mesh.o
> manager/_manager.o signeur/_signeur.o     -L/usr/lib -L/usr/local/lib
> -lpthread -L/lib64 -L/usr/lib64 -L/usr/lib
> -L/home/diag/opt/gsl/1.15/gnu/lib/ -lgsl -lgslcblas -lm -lhdf5
> -L/home/diag/opt/hdf5/1.8.9/gnu/openmpi/lib    -o moose
>
>
>
> But getting error:
>
>
> moosemodule.cpp:80:31: error: numpy/arrayobject.h: No such file or
> directory
>
>
>
> But We have "numpy 1.6" loaded.  Should we set some other flag or should
> we have different version of "numpy" ?
>
>
>
> Thanks a lot in advance for any advice.
>
> Vadim.
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
> On Mar 2, 2013, at 10:58 PM, Subhasis Ray wrote:
>
> Hi,
> In the serial version, HDF5 1.8.9 should work fine. It is used in
> HDF5DataWriter class to incrementally dump data into an HDF5 file. It was
> implemented to reduce memory usage for long running simulations generating
> gigabytes of data.
> Best,
>   Subha
>
>
> On Sun, Mar 3, 2013 at 1:22 AM, Majumdar, Amitava <maj...@sd...>wrote:
>
>>  If HDF5 version is the only issue, we can make HDF51.8.0 or higher
>> available (actually we have HDF5 1.8.9 available
>> on another parallel computer at SDSC).
>>
>> -Amitvai
>>  ------------------------------
>> *From:* Subhasis Ray [ray...@gm...]
>> *Sent:* Saturday, March 02, 2013 1:35 AM
>> *To:* Vadim A
>>
>> *Cc:* Niraj Dudani; Yoshimoto, Kenneth; Majumdar, Amitava;
>> moo...@li...
>> *Subject:* Re: [moose-generic] MOOSE_installation_problem
>>
>>    Dear Vadim,
>>    I would recommend that you avoid the MPI build at this point. The
>> errors again are due to a minor issue with HDF5 library that can vary from
>> package to package. These functions were introduced in HDF5 1.8.0. Either
>> you need the same or a newer version installed. Otherwise, you can skip
>> HDF5 support (which is somewhat experimental anyways) altogether by passing
>> USE_HDF5=0 to "make" command.
>>    Hope this helps,
>>      Subha
>>
>>
>> On Sat, Mar 2, 2013 at 5:35 AM, Vadim A <ast...@nc...> wrote:
>>
>>> Dear Subha.
>>>
>>>  I set CXXFLAG to include mpi.h in "ifeq ($(BUILD),release)"
>>>
>>>
>>>  like you've suggested  and it seems that it can find all libraries now
>>> but unfortunately
>>>
>>>
>>>  make BUILD=release USE_MPI=1
>>>
>>>
>>>  produce another error
>>>
>>>
>>>  mpicxx -O3 -pthread -fpermissive -fno-strict-aliasing -fPIC -Wall
>>> -Wno-long-long -pedantic -DNDEBUG -DUSE_GENESIS_PARSER -I
>>> /home/diag/opt/openmpi/1.4.3/gnu/include  -DSVN_REVISION=\"4401M\" -DLINUX
>>> -DUSE_MPI -DMPICH_IGNORE_CXX_SEEK -DUSE_GSL -I
>>> /home/diag/opt/gsl/1.15/gnu/include/    -DUSE_HDF5
>>>  -DH5_NO_DEPRECATED_SYMBOLS -I  /home/diag/opt/hdf5/1.6/gnu/include  -I.
>>> -I../basecode -I../msg HDF5DataWriter.cpp -c
>>>
>>>
>>>
>>>
>>>  ../basecode/Conv.h:1009: warning: converting to 'unsigned int' from
>>> 'const double'
>>> HDF5DataWriter.cpp: In member function 'hid_t
>>> HDF5DataWriter::get_dataset(std::string)':
>>> HDF5DataWriter.cpp:222: error: 'H5Gcreate2' was not declared in this
>>> scope
>>> HDF5DataWriter.cpp:225: error: 'H5Gopen2' was not declared in this scope
>>> HDF5DataWriter.cpp:258: error: 'H5Dopen2' was not declared in this scope
>>>
>>>
>>>
>>>
>>>
>>>  HDF5  and openmpi  libraries included. Do I still missing something ?
>>>
>>>  Thanks a lot in advance for any comments.
>>>
>>>  Vadim.
>>>
>>>
>>>
>>>
>>>
>>>  On Feb 28, 2013, at 7:37 PM, Subhasis Ray wrote:
>>>
>>>
>>>
>>>
>>> Do not enable MPI unless you want to try the parallel version. To enable
>>> MPI, you have to pass the argument "USE_MPI=1" to make command.
>>>
>>>
>>>>
>>>>
>>>>  # Here we automagically change compilers to deal with MPI.
>>>> ifdef USE_MPI
>>>>         CXX = mpicxx
>>>> #       CXX = /usr/local/mvapich2/bin/mpicxx
>>>>         PARALLEL_DIR = /home/diag/opt/openmpi/1.4.3/gnu/  #parallel
>>>>          PARALLEL_LIB = /home/diag/opt/openmpi/1.4.3/gnu/lib/
>>>> #parallel/parallel.o
>>>> else
>>>>         CXX = g++
>>>> #       CXX = CC        # Choose between Solaris CC and g++ on a
>>>> Solaris machine
>>>> endif
>>>>
>>>>
>>>>
>>>  In the above, the PARALLEL_DIR and PARALLEL_LIB are MOOSE source
>>> directory and object file for the parallel version. They are not to be
>>> modified.
>>>
>>>>
>>>>
>>>>
>>>>  We are using "openmpi" which does not have "parallel.o"
>>>>
>>>
>>>  As mentioned above, parallel.o will be generated by the build process.
>>> It looks like your version of hdf5 includes mpi.h and the compiler cannot
>>> find it. So you have to add -I{location of mpi.h} to the CXXFLAGS variable
>>> somewhere.Depending on your build type, you can append this to the CXXFLAG
>>> assignment under :
>>>
>>> ifeq ($(BUILD),debug)
>>>
>>>  and/or under
>>>
>>> ifeq ($(BUILD),release)
>>>
>>>
>>>  Best,
>>>    Subha
>>>
>>>
>>>
>>
>
>