louhi.csc.fi (Cray XT4/XT5)

Here you find information about the the system http://www.csc.fi/english/research/Computing_services/computing/servers/louhi.

Note

These instructions are up-to-date as of August 28th 2012.

GPAW

The recent operating system releases for Cray XT4/5 (CLE 2.2 UP01 and later) supports dynamic libraries which simplifies GPAW installation significantly.

These instructions for GPAW installation use Python 2.6.5 compiled with GNU compiler suite, see the end of this page for instructions for compiling Python.

First, load the Python module and set XTPE_LINK_TYPE environment variable for dynamic linking:

module load python
module load hdf5-parallel
setenv XTPE_LINK_TYPE dynamic

GPAW can now be build with a minimal customize.py

#User provided customizations for the gpaw setup

compiler = 'cc'
mpicompiler = 'cc'
mpilinker= 'cc'

extra_compile_args = ['-std=c99']
libraries = []

scalapack = True
hdf5 = True

define_macros += [('GPAW_NO_UNDERSCORE_CBLACS', '1')]
define_macros += [('GPAW_NO_UNDERSCORE_CSCALAPACK', '1')]
define_macros += [("GPAW_ASYNC",1)]
define_macros += [("GPAW_MPI2",1)]

Currently, there is a small bug in the Cray libraries which results in failure when trying to build the _gpaw.so library. As the library is not really needed in parallel calculations, the problem can be circumvented by creating the file after running the setup.py script:

python setup.py build_ext
...
touch build/lib.linux-x86_64-2.6/_gpaw.so
python setup.py install --home=...

Python and Numpy

Python can be compiled with PGI compiler as follows:

setenv XTPE_LINK_TYPE dynamic
./configure --prefix=path_to_install CC=cc CXX=cc OPT=-fastsse LINKFORSHARED=-Wl,--export-dynamic
make install

In order to use optimized BLAS with Numpy one has to first build a CBLAS which is linked with Cray’s optimized BLAS routines. First, download the CBLAS source from netlib:

wget http://www.netlib.org/blas/blast-forum/cblas.tgz
tar -xzf cblas.tgz

Change to the CBLAS directory and copy Makefile.LINUX to Makefile.in. Add correct compiler commands and paths to Makefile.in:

...
PLAT = louhi

#-----------------------------------------------------------------------------
# Libraries and includs
#-----------------------------------------------------------------------------

BLLIB =
CBDIR = $(HOME)/CBLAS
CBLIBDIR = $(CBDIR)/lib/$(PLAT)
CBLIB = $(CBLIBDIR)/libcblas.a

#-----------------------------------------------------------------------------
# Compilers
#-----------------------------------------------------------------------------

CC = cc
FC = ftn

LOADER = $(FC)

#-----------------------------------------------------------------------------
# Flags for Compilers
#-----------------------------------------------------------------------------

CFLAGS = -O3 -DADD_ -fPIC
FFLAGS = -O3 -fPIC

...

Finally, build CBLAS:

make alllib

You are now ready to build Numpy with the newly created CBLAS library. The standard Numpy tries to use only the ATLAS BLAS, and in order to use different BLAS one has to manually edit the file numpy/core/setup.py. Comment out an if statement as follows:

def get_dotblas_sources(ext, build_dir):
  if blas_info:
      # if ('NO_ATLAS_INFO',1) in blas_info.get('define_macros',[]):
      #     return None # dotblas needs ATLAS, Fortran compiled blas will not be sufficient.
      return ext.depends[:1]

Then, add the correct libraries and paths to the file site.cfg:

[blas]
blas_libs = cblas
library_dirs = /home/csc/jenkovaa/CBLAS/lib/louhi

[lapack]
lapack_libs = sci
library_dirs = /opt/xt-libsci/10.3.8/pgi/lib

Now, one should be able to build Numpy as usual:

python setup.py install