hermit.hww.de (Cray XE6)

Here you find information about the the system http://www.hlrs.de/systems/platforms/cray-xe6-hermit/.

Note

These instructions are up-to-date as of November 18th 2013.

Scalable Python

As the Hermit system is intedend for simulations with thousands of CPU cores, a special Python interpreter is used here. The scalable Python reduces the import time by performing all import related I/O operations with a single CPU core and uses MPI for broadcasting the data. As a limitation, all the MPI tasks have to perform the same import statements.

As HLRS does not allow general internet access on compute system, e.g. version control repositories cannot be accessed directly (it is possible to setup ssh tunnel for some services). Here, we download the scalable Python first to to a local machine and use then scp for copying it to Hermit:

git clone git@gitorious.org:scalable-python/scalable-python.git scalable-python-src
scp -r scalable-python-src username@hermit1.hww.de:

We will build scalable Python with GNU compilers (other compilers can be used for actual GPAW build), so start by changing the default programming environment on Hermit:

module swap PrgEnv-cray PrgEnv-gnu

Due to cross-compile environment in Cray XE6, a normal Python interpreter is build for the front-end nodes and the MPI-enabled one for the compute nodes. The build can be accomplished by the following build_gcc script

#!/bin/bash -x

export CC=cc
export CXX=g++
export MPICC=cc
export LINKFORSHARED='-Wl,-export-dynamic -dynamic'
export MPI_LINKFORSHARED='-Wl,-export-dynamic -dynamic'

install_dir='/some_path/scalable-python-gcc'
./configure --prefix=$install_dir --enable-mpi --disable-ipv6 2>&1 | tee loki-conf

module swap craype-interlagos craype-istanbul
module list
make 2>&1 | tee log-make
make install 2>&1 | tee log-inst

make clean

module swap craype-istanbul craype-interlagos
make mpi 2>&1 | tee log-make-mpi
cp ${install_dir}/bin/python .
make install-mpi 2>&1 | tee log-inst-mpi

Python packages can now be built on the front-end node with /some_path/scalable-python-gcc/bin/python.

NumPy

As the performance of the HOME filesystem is not very good, we install all the other components than the pure Python to a disk within the workspace mechanism of HLRS (with disadvantage that the workspaces expire and have to be manually reallocated). Otherwise, no special tricks are needed for installing NumPy:

/some_path/scalable-python-gcc/bin/python setup.py install --home=/path_in_workspace

GPAW

On Hermit, Intel compiler together with ACML library seemed to give best performance for GPAW, in addition HDF5 will be used for parallel I/O. Thus, load the followgin modules:

module swap PrgEnv-gnu PrgEnv-intel
module load acml
module load hdf5-parallel

The compilation is relatively straightforward, however, as we build NumPy for compute nodes it does not work in front-end, and one has to specify NumPy include dirs in customize.py and provide --ignore-numpy flag when building. The system NumPy headers seem to work fine, but safer option is to use headers of own NumPy installation

extra_compile_args = ['-std=c99', '-O3'] #, '-O0']
compiler = 'cc'
mpicompiler = 'cc'
mpilinker= 'cc'
libraries = ['acml']
extra_link_args += ['-dynamic']
include_dirs += ['/usr/lib64/python2.6/site-packages/numpy/core/include']
# include_dirs += ['/path_in_workspace/lib/python/numpy/core/include/']

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)]
define_macros += [("GPAW_PERFORMANCE_REPORT",1)]

Buid with:

/some_path/scalable-python-gcc/bin/python setup.py install --home=/path_in_workspace --ignore-numpy

Creating a module

Users can define their own modules for making it easier to setup environment variables. First, create (or edit) a file .modulerc in the home directory:

#%Module1.0
##

append-path MODULEPATH $HOME/modules

Now, custom modules can be put to the modules directory, e.g. GPAW module in file modules/gpaw:

#%Module1.0

set     prog_root       "/root_folder_of_gpaw_installation"

# Change PrgEnv to intel
set confl_prgenvs [list "PrgEnv-cray" "PrgEnv-gnu"]
foreach prgenv $confl_prgenvs {
  if { [ is-loaded $prgenv] } {
    module swap $prgenv PrgEnv-intel
  }
}

setenv GPAW_SETUP_PATH                  "$prog_root/gpaw-setups-0.8.7929"
prepend-path    PYTHONPATH              "$prog_root/lib/python"
prepend-path    PATH                    "$prog_root/bin/"

Now, GPAW paths can be set as:

module load gpaw