[molpro-user] Can I compile molpro2009.1 parallel version without GA?
Manhui Wang
wangm9 at cardiff.ac.uk
Fri Oct 9 15:56:22 BST 2009
Dear He Ping,
GA may take full advantage of shared memory on an MPP node, but
MPI-2 doesn't. On the other hand, MPI-2 may take advantage of the
built-in MPI-2 library with fast connection. The performance depends on
lots of facts, including MPI library, machine, network etc. It is better
to build both versions of Molpro, and then choose the better one on that
machine.
It doesn't seem to be hard to make GA4-2 work on machine like
yours. Details were shown in my previous email.
Best wishes,
Manhui
He Ping wrote:
> Hi Manhui,
>
> Thanks. I will try these patches first.
> But would you like to tell me something about GA's effect on molpro? So
> far I can not get a GA version molpro, so I can not know the performance
> of GA version. If GA version is not much better than non-GA version, I
> will not take much time building it.
> Thanks a lot
>
> On Fri, Oct 9, 2009 at 7:25 PM, Manhui Wang <wangm9 at cardiff.ac.uk
> <mailto:wangm9 at cardiff.ac.uk>> wrote:
>
> Dear He Ping,
> Recent patches include some bugfixes for intel compiler 11,
> OopenMPI, and running molpro across nodes with InfiniBand. If you have
> not updated them, please do it now. It may resolve your existing
> problems.
>
> He Ping wrote:
> > Dear Manhui,
> >
> > Thanks a lot for your detailed reply, that's very helpful. Very
> sorry to
> > answer later, for I have to do a lot of tests. So far, one version of
> > molpro2009.1 is basically ok, but I still have some questions.
> >
> > 1. Compile Part.
> > Openmpi 1.3.3 can pass compile and link for both w GA and w/o GA
> > 4.2. I do not use my own blas, so I use default, this is my
> > configure step,
> >
> > ./configure -batch -ifort -icc -mppbase $MPI_HOME/include64 -var
> > LIBS="-L/usr/lib64 -libverbs -lm" -mpp (in your letter, I guess
> > you forget this necessary option.)
> >
> > But *intelmpi failed for both*, I can show the err message
> > seperately below.
> >
> > *Intelmpi w/o GA: *
> >
> > make[1]: Nothing to be done for `default'.
> > make[1]: Leaving directory
> >
> `/datastore/workspace/scheping/molpro2009.1_intelmpi/molpro2009.1/utilities'
> > make[1]: Entering directory
> >
> `/datastore/workspace/scheping/molpro2009.1_intelmpi/molpro2009.1/src'
> > Preprocessing include files
> > make[1]: *** [common.log] Error 1
> > make[1]: *** Deleting file `common.log'
> > make[1]: Leaving directory
> >
> `/datastore/workspace/scheping/molpro2009.1_intelmpi/molpro2009.1/src'
> > make: *** [src] Error 2
> >
> > *Intelmpi w GA:*
> >
> > compiling molpro_cvb.f
> > failed
> > molpro_cvb.f(1360): error #5102: Cannot open include file
> > 'common/ploc'
> > include "common/ploc"
> > --------------^
> > compilation aborted for molpro_cvb.f (code 1)
> > make[3]: *** [molpro_cvb.o] Error 1
> > preprocessing perfloc.f
> > compiling perfloc.f
> > failed
> > perfloc.f(14): error #5102: Cannot open include file
> 'common/ploc'
> > include "common/ploc"
> > --------------^
> > perfloc.f(42): error #6385: The highest data type rank permitted
> > is INTEGER(KIND=8). [VARIAT]
> > if(.not.variat)then
> > --------------^
> > perfloc.f(42): error #6385: The highest data type rank permitted
> > is INTEGER(KIND=8).
>
> Which version of intel compilers are you using? Has your GA worked fine?
> We have tested Molpro2009.1 with
> (1) intel/compilers/10.1.015 (11.0.074), GA 4-2 hosted by intel/mpi/3.1
> (3.2)
> (2) without GA, intel/compilers/10.1.015 (11.0.074), intel/mpi/3.1 (3.2)
>
> all work fine. CONFIG files will be helpful to see the problems.
> >
> > 2. No .out file when I use more than about 12 processes, but I can
> > get .xml file. It's very strange, everything is ok when process
> > number is less than 12, but once exceed this number, such as 16
> > cpus, molpro always gets this err message,
> >
> > orrtl: severe (174): SIGSEGV, segmentation fault occurred
> > Image PC Routine
> > Line Source
> > libopen-pal.so.0 00002AAAAB4805C6 Unknown
> > Unknown Unknown
> > libopen-pal.so.0 00002AAAAB482152 Unknown
> > Unknown Unknown
> > libc.so.6 000000310FC5F07A Unknown
> > Unknown Unknown
> > molprop_2009_1_Li 00000000005A4C36 Unknown
> > Unknown Unknown
> > molprop_2009_1_Li 00000000005A4B84 Unknown
> > Unknown Unknown
> > molprop_2009_1_Li 000000000053E57B Unknown
> > Unknown Unknown
> > molprop_2009_1_Li 0000000000540A8C Unknown
> > Unknown Unknown
> > molprop_2009_1_Li 000000000053C5E5 Unknown
> > Unknown Unknown
> > molprop_2009_1_Li 00000000004BCA5C Unknown
> > Unknown Unknown
> > libc.so.6 000000310FC1D8A4 Unknown
> > Unknown Unknown
> > molprop_2009_1_Li 00000000004BC969 Unknown
> > Unknown Unknown
> >
> > Can I ignore this message?
> Have you seen this on one or multiple nodes? If on multiple nodes, the
> problem has been fixed by recent patches. By default, both *.out and
> .xml can be obtained, but you can use option --no-xml-output to disable
> the *xml.
> In addition, OpenMPI seems to be unstable sometime. When lots of jobs
> are run with OpenMPI, some jobs hang up unexpectedly. This behavior is
> not seen for Intel MPI.
>
> >
> > 3. Script Err. For molpro openmpi version, the script
> > molpro_openmpi1.3.3/bin/molprop_2009_1_Linux_x86_64_i8 seems not
> > to work.
> > When I call this script, only one process is started, even if I
> > use -np 8. So I have to run it manually, such as
> > mpirun -np 8 -machinefile ./hosts
> > molprop_2009_1_Linux_x86_64_i8.exe test.com
> <http://test.com> <http://test.com>
> Have your ./bin/molpro worked?. For me, it works fine. In ./bin/molpro,
> some environmental settings are included. In the case that ./bin/molpro
> doesn't work properly, you might want to directly use
> molprop_2009_1_Linux_x86_64_i8.exe, then it is your responsibility to
> set up these environmental variables.
> > 4. Molpro w GA can not cross over nodes. One node is ok, but if
> cross
> > over nodes, I will get "molpro ARMCI DASSERT fail" err, and
> molpro
> > can not be terminated normally. Do you know the difference
> between
> > w GA and w/o GA? If GA is not better than w/o GA, I will
> pass this
> > GA version.
> I think this problem has been fixed by recent patches.
> As the difference between molpro w GA and w/o GA, it is hard to make a
> simple conclusion. For calculations with a small number of processes(
> eg. < 8), molpro w GA might be somewhat fast, but molpro without GA is
> quite competitive in performance when it is run with a large number of
> processes. Please refer to the benchmark
> results(http://www.molpro.net/info/bench.php).
> >
> >
> >
> > Sorry for packing up so many questions, answer is any one
> question
> > will help me a lot. And I think question 1 and 2 will be more
> > important to me. Thanks.
> >
> >
> >
>
> Best wishes,
> Manhui
>
>
>
> > On Thu, Sep 24, 2009 at 5:33 PM, Manhui Wang <wangm9 at cardiff.ac.uk
> <mailto:wangm9 at cardiff.ac.uk>
> > <mailto:wangm9 at cardiff.ac.uk <mailto:wangm9 at cardiff.ac.uk>>> wrote:
> >
> > Hi He Ping,
> > Yes, you can build parallel Molpro wthout GA for 2009.1.
> Please see
> > the manual A..3.3 Configuration
> >
> > For the case of using the MPI-2 library, one example can be
> >
> > ./configure -mpp -mppbase /usr/local/mpich2-install/include
> >
> > and the -mppbase directory should contain file mpi.h. Please
> ensure the
> > built-in or freshly built MPI-2 library fully supports MPI-2
> standard
> > and works properly.
> >
> >
> > Actually we have tested molpro2009.1 on almost the same system
> as what
> > you mentioned (EMT64, Red Hat Enterprise Linux Server release 5.3
> > (Tikanga), Intel MPI, ifort, icc, Infiniband). For both GA and
> MPI-2
> > buildings, all work fine. The configurations are shown as
> follows(beware
> > of lines wrapping):
> > (1) For Molpro2009.1 built with MPI-2
> > ./configure -batch -ifort -icc -blaspath
> > /software/intel/mkl/10.0.1.014/lib/em64t
> <http://10.0.1.014/lib/em64t>
> > <http://10.0.1.014/lib/em64t> -mppbase $MPI_HOME/include64
> > -var LIBS="-L/usr/lib64 -libverbs -lm"
> >
> > (2) For Molpro built with GA 4-2:
> > Build GA4-2:
> > make TARGET=LINUX64 USE_MPI=y CC=icc FC=ifort COPT='-O3'
> > FOPT='-O3' \
> > MPI_INCLUDE=$MPI_HOME/include64 MPI_LIB=$MPI_HOME/lib64 \
> > ARMCI_NETWORK=OPENIB MA_USE_ARMCI_MEM=y
> > IB_INCLUDE=/usr/include/infiniband IB_LIB=/usr/lib64
> >
> > mpirun ./global/testing/test.x
> > Build Molpro
> > ./configure -batch -ifort -icc -blaspath
> > /software/intel/mkl/10.0.1.014/lib/em64t
> <http://10.0.1.014/lib/em64t>
> > <http://10.0.1.014/lib/em64t> -mppbase /GA4-2path -var
> > LIBS="-L/usr/lib64 -libverbs -lm"
> >
> > (LIBS="-L/usr/lib64 -libverbs -lm" will make molpro link with
> Infiniband
> > library)
> >
> > (some note about MOLPRO built with MPI-2 library can also been
> in manual
> > 2.2.1 Specifying parallel execution)
> > Note: for MOLPRO built with MPI-2 library, when n processes are
> > specified, n-1 processes are used to compute and one process
> is used to
> > act as shared counter server (in the case of n=1, one process
> is used to
> > compute and no shared counter server is needed). Even so, it
> is quite
> > competitive in performance when it is run with a large number of
> > processes.
> > If you have built both versions, you can also compare the
> performance
> > yourself.
> >
> >
> > Best wishes,
> > Manhui
> >
> > He Ping wrote:
> > > Hello,
> > >
> > > I want to run molpro2009.1 parallel version on infiniband
> network.
> > I met
> > > some problems when using GA, from the manual, section 3.2, there
> > is one
> > > line to say,
> > >
> > > If the program is to be built for parallel execution then
> the Global
> > > Arrays toolkit *or* the
> > > MPI-2 library is needed.
> > >
> > > Does that mean I can build molpro parallel version without
> GA? If so,
> > > who can tell me some more about how to configure?
> > > My system is EM64T, Red Hat Enterprise Linux Server release 5.1
> > > (Tikanga), intel mpi, intel ifort and icc.
> > >
> > > Thanks a lot.
> > >
> > > --
> > >
> > > He Ping
> > >
> > >
> > >
> >
> ------------------------------------------------------------------------
> > >
> > > _______________________________________________
> > > Molpro-user mailing list
> > > Molpro-user at molpro.net <mailto:Molpro-user at molpro.net>
> <mailto:Molpro-user at molpro.net <mailto:Molpro-user at molpro.net>>
> > > http://www.molpro.net/mailman/listinfo/molpro-user
> >
> > --
> > -----------
> > Manhui Wang
> > School of Chemistry, Cardiff University,
> > Main Building, Park Place,
> > Cardiff CF10 3AT, UK
> > Telephone: +44 (0)29208 76637
> >
> >
> >
> >
> > --
> >
> > He Ping
> > [O] 010-58813311
>
> --
> -----------
> Manhui Wang
> School of Chemistry, Cardiff University,
> Main Building, Park Place,
> Cardiff CF10 3AT, UK
> Telephone: +44 (0)29208 76637
>
>
>
>
> --
>
> He Ping
> [O] 010-58813311
--
-----------
Manhui Wang
School of Chemistry, Cardiff University,
Main Building, Park Place,
Cardiff CF10 3AT, UK
Telephone: +44 (0)29208 76637
More information about the Molpro-user
mailing list