[Q-e-developers] [Pw_forum] parallel installation comments

[Scott Brozell]

Hi,

Great, thanks.

Here is a correction for item 3 on building with -mkl (i forgot to
include LAPACK_LIBS):

./configure BLAS_LIBS='-mkl=cluster' LAPACK_LIBS='-mkl=cluster' SCALAPACK_LIBS='-mkl=cluster'

scott

On Thu, May 04, 2017 at 09:53:47AM +0200, Paolo Giannozzi wrote:
> Thank you. Your proposed changes are now in the development version.
> 
> Paolo
> 
> On Wed, May 3, 2017 at 8:07 PM, Scott Brozell <srb at osc.edu> wrote:
> > Hi,
> >
> > Installation of quantum espresso on large commodity clusters, e.g,
> > https://www.osc.edu/resources/technical_support/supercomputers/oakley/technical_specifications
> > using intel compilers and mvapich mpi has been straightforward in
> > the past; for example, configure; make all; worked in 5.2.1.
> > However, in 5.4.0, 6.0, and 6.1 the simple procedure above produced
> > serial executables.
> >
> > Attached is a patch file for x_ac_qe_mpi.m4 in 6.1 which corrects
> > two problems.  In the four points below, these changes are described,
> > a comment on MKL is made, and a typo is reported.
> >
> > 1.  x_ac_qe_mpi.m4 parallel is used before initialization.
> > In the released install/configure, which is based on x_ac_qe_mpi.m4,
> > an error occurs because parallel is used before it is initialized:
> >
> > ./install/configure: line 11602: test: : integer expression expected
> >
> > This change to x_ac_qe_mpi.m4 fixes that:
> > #parallel=0
> > ->
> > parallel=0
> >
> > 2.  x_ac_qe_mpi.m4 mpi_libs can be non empty.
> > On our computer systems when an MPI is loaded, mvapich or otherwise,
> > the environment variable MPI_LIBS is defined.  This causes mpi_libs
> > in install/configure to be defined to a non empty string.  But
> > other espresso requirements for building parallel executables
> > may not be met; in particular, -D__MPI may not be.
> >
> > This addition at line 28 to x_ac_qe_mpi.m4 fixes that:
> > else
> >         if test "$use_parallel" -ne 0
> >         then
> >                 have_mpi=1
> >                 parallel=1
> >                 try_dflags="$try_dflags -D__MPI"
> >         fi
> >
> > 3.  -mkl for intel compilers is a better way.
> > configure's automatic detection of MKL failed on our systems maybe
> > because we have available multiple compilers, MPIs, and MKLs.
> > Intel compiler versions 11 or later provide the -mkl option.
> > This is a much better way since it avoids decrypting MKLs large
> > number of files to produce the appropriate -L and -l options.
> > Here is how we invoked configure:
> >
> > ./configure BLAS_LIBS='-mkl=cluster' SCALAPACK_LIBS='-mkl=cluster'
> >
> > 4.  User Guide typo on p5; user_guide.tex patch file is attached.
> >
> > Be short: no need to send 128 copies of the same error message just because you this is
> > ->
> > Be short: no need to send 128 copies of the same error message just because this is
> >
> > thanks,
> > scott
> >
> > Scott Brozell, Ph.D.
> > Scientific Applications Group
> > Ohio Supercomputer Center
> > Columbus, OH 43212