NEMO 3.7/4.0 + XIOS 2.0¶
Tested with
gcc4.9
,gcc5.4
on a linux systemgcc4.8
on a Mac (El Capitan OSX 10.11)
This is the version I first implemented GEOMETRIC in, which is a development version I guess (?) that eventually led to NEMO 4.0. The code structure largely follows NEMO 3.6 but the commands are slightly different.
If you get errors that are not documented here, see if the XIOS1.0 NEMO3.6 page contain the relevant errors.
The assumption here is that the compiler is fixed and the packages (e.g.,
NetCDF4 and a MPI bindings) are configured to be consistent with the compilers.
See here to check whether the binaries exist, where they
are, and how they might be installed separately if need be. All the #CHANGE
ME
highlighted below needs be modified to point to the appropriate paths or
binaries (soft links with ln -s
are ok).
The instructions below uses gcc4.9
for demonstration (modifications with
gcc5.4
as appropriate). I defined some extra variables on a Linux machine:
export $BD=/home/julian/testing/gcc4.9-builds # CHANGE ME
export C_INCLUDE_PATH=$BD/install/include:$C_INCLUDE_PATH
export CPLUS_INCLUDE_PATH=$BD/install/include:$CPLUS_INCLUDE_PATH
export LIBRARY_PATH=$BD/install/lib:$LIBRARY_PATH
export LD_LIBRARY_PATH=$BD/install/lib:$LD_LIBRARY_PATH
You shouldn’t need to do the above if the packages are forced to look at the
right place (e.g. via -L
and/or -I
flags with path to libraries and
include files respectively). Not all of these are necessary depending on whether
you choose to build/have static or dynamic libraries, and the
LD_LIBRARY_PATH
seems to sort out a lot of problems with linking libraries.
On a Mac done through anaconda the above was not necessary. My understanding is that setting these variables might not actually do anything unless an option is specifically enabled in Xcode.
XIOS 2.0 (svn v1322)¶
Do the following:
mkdir XIOS
cd XIOS
svn checkout -r 1322 http://forge.ipsl.jussieu.fr/ioserver/svn/XIOS/trunk xios-2.0
Note
svn checkout
http://forge.ipsl.jussieu.fr/ioserver/svn/XIOS/branchs/xios-2.0 xios-2.0
also works with instructions below.
To get XIOS to compile, the compilers and packages need to be pointed to first,
via modifying files in arch
. Since I am using gcc
, I did the following
just to make a fresh copy:
cd xios2.0/arch
cp arch-GCC_LINUX.env arch-GCC_local.env
cp arch-GCC_LINUX.fcm arch-GCC_local.fcm
cp arch-GCC_LINUX.path arch-GCC_local.path
The *.env
file specifies where HDF5 and NetCDF4 libraries live. The
*.fcm
file specifies which compilers and options to use. The *.path
file
specifies which paths and options to include. My files look like the following:
# arch-GCC_local.env
export HDF5_INC_DIR=/usr/local/include # CHANGE ME
export HDF5_LIB_DIR=/usr/local/lib # CHANGE ME
export NETCDF_INC_DIR=/usr/local/include # CHANGE ME
export NETCDF_LIB_DIR=/usr/local/lib # CHANGE ME
You could get an idea where the HDF5 and NetCDF4 directories are by doing
which h5copy
and which nc-config
(assuming these are on $PATH
),
which should give you a directory/bin
, and it is the directory
part you
want. If you did install the libraries somewhere else as in other packages, say, then make sure the which
commands are pointing to
the right place.
# arch-GCC_local.fcm
################################################################################
################### Projet XIOS ###################
################################################################################
%CCOMPILER /usr/local/bin/mpicc # CHANGE ME
%FCOMPILER /usr/local/bin/mpif90 # CHANGE ME
%LINKER /usr/local/bin/mpif90 # CHANGE ME
%BASE_CFLAGS -ansi -w
%PROD_CFLAGS -O3 -DBOOST_DISABLE_ASSERTS
%DEV_CFLAGS -g -O2
%DEBUG_CFLAGS -g
%BASE_FFLAGS -D__NONE__
%PROD_FFLAGS -O3
%DEV_FFLAGS -g -O2
%DEBUG_FFLAGS -g
%BASE_INC -D__NONE__
%BASE_LD -lstdc++
%CPP cpp-4.9 # CHANGE ME
%FPP cpp-4.9 -P # CHANGE ME
%MAKE make
Check the MPI locations and versions by doing which mpicc
and mpicc
--version
say. If they are the right ones you could just have mpicc
instead of the full path as given above. MPI bindings are used here to avoid a
possible error that may pop up in relation to the build trying to find
mpi.h
. The gmake
command was swapped out by the make
command (I
don’t have cmake
on the laptop).
Note
For gcc5.4
and maybe newer versions, doing just the above when compiling
leads to a whole load of errors about clashing in C++:
.../include/boost/functional/hash/extensions.hpp:69:33: error: ‘template<class T, class A> std::size_t boost::hash_value’ conflicts with a previous declaration
std::size_t hash_value(std::list<T, A> const& v)
^
Adding -D_GLIBCXX_USE_CXX11_ABI=0
to %BASE_CFLAGS
fixes these.
# arch-GCC_local.path
NETCDF_INCDIR="-I$NETCDF_INC_DIR"
NETCDF_LIBDIR="-Wl,'--allow-multiple-definition' -L$NETCDF_LIB_DIR"
NETCDF_LIB="-lnetcdff -lnetcdf"
MPI_INCDIR=""
MPI_LIBDIR=""
MPI_LIB=""
HDF5_INCDIR="-I$HDF5_INC_DIR"
HDF5_LIBDIR="-L$HDF5_LIB_DIR"
HDF5_LIB="-lhdf5_hl -lhdf5 -lhdf5 -lz"
The above has all the OASIS (the atmosphere / ocean coupler) keys removed. I
added the -Wl,'--allow-multiple-definition'
key for reasons I don’t remember
anymore…
I went into bld.cfg
, found the line
bld::tool::cflags %CFLAGS %CBASE_INC -I${PWD}/extern/src_netcdf -I${PWD}/extern/boost/include -I${PWD}/extern/rapidxml/include -I${PWD}/extern/blitz/include
and changed src_netcdf
to src_netcdf4
(see XIOS1.0 stuff for the reason).
Now it should be ready to compile. Assuming the current directory is
xios2.0/arch
:
cd ../
./make_xios --full --prod --arch GCC_local -j2 |& tee compile_log.txt
The -j2
option uses two processors to build. The tee
command is to keep
logs of potential errors (the |&
is short for 2>&1 |
) for debugging the
compiler issues that may arise. It should work and takes around 5 mins to
compile for me. The main end result is are binaries in xios2.0/bin/
which
NEMO will call.
NEMO 3.7/4.0 (svn v8666)¶
Check out a version of NEMO. I have another folder separate to the XIOS folders to contain the NEMO codes and binaries:
mkdir NEMO
cd NEMO
svn checkout -r 8666 http://forge.ipsl.jussieu.fr/nemo/svn/NEMO/trunk nemo3.7-8666
This checks out version 8666 (NEMO 3.7/4.0) and dumps it into a folder called
nemo3.7-8666
(change the target path to whatever you like). A similar
procedure to specify compilers and where XIOS lives needs to be done for NEMO.
Again, because of the compilers I am using:
cd nemo3.7-8666/NEMOGCM/ARCH
cp OLD/arch-gfortran_linux.fcm ./arch-gfortran_local.fcm
None of the fcm files associated with gfortran actually worked for me out of the box so here is my build of it (click HERE for a detailed log of how I got to the following):
# gfortran_local.fcm
# generic gfortran compiler options for linux
# NCDF_INC netcdf include file
# NCDF_LIB netcdf library
# FC Fortran compiler command
# FCFLAGS Fortran compiler flags
# FFLAGS Fortran 77 compiler flags
# LD linker
# LDFLAGS linker flags, e.g. -L<lib dir> if you have libraries in a
# FPPFLAGS pre-processing flags
# AR assembler
# ARFLAGS assembler flags
# MK make
# USER_INC additional include files for the compiler, e.g. -I<include dir>
# USER_LIB additional libraries to pass to the linker, e.g. -l<library>
%NCDF_HOME /usr/local # CHANGE ME
%XIOS_HOME /home/julian/testing/gcc4.9-builds/XIOS/xios-2.0 # CHANGE ME
%CPP cpp-4.9 # CHANGE ME
%CPPFLAGS -P -traditional
%XIOS_INC -I%XIOS_HOME/inc
%XIOS_LIB -L%XIOS_HOME/lib -lxios
%NCDF_INC -I%NCDF_HOME/include
%NCDF_LIB -L%NCDF_HOME/lib -lnetcdf -lnetcdff -lstdc++
%FC mpif90 # CHANGE ME
%FCFLAGS -fdefault-real-8 -O3 -funroll-all-loops -fcray-pointer -cpp -ffree-line-length-none
%FFLAGS %FCFLAGS
%LD %FC
%LDFLAGS
%FPPFLAGS -P -C -traditional
%AR ar
%ARFLAGS -rs
%MK make
%USER_INC %XIOS_INC %NCDF_INC
%USER_LIB %XIOS_LIB %NCDF_LIB
The main changes are (see here for an attempt at the reasoning and a log of errors that motivates the changes):
added
%NCDF_HOME
to point to where NetCDF livesadded
%XIOS_*
keys to point to where XIOS livesadded
%CPP
and flags, consistent with usinggcc4.9
added the
-lnetcdff
and-lstdc++
flags to NetCDF flagsusing
mpif90
which is a MPI binding ofgfortran-4.9
added
-cpp
and-ffree-line-length-none
to Fortran flagsswapped out
gmake
withmake
Then, I did (see NEMO 3.6 for the reason):
cd ../CONFIG/
./makenemo -j0 -r GYRE_PISCES -n GYRE_testing -m gfortran_local
Edit /GYRE_testing/cpp_GYRE_testing.fcm
and replaced key_top
with
key_nosignedzero
(does not compile TOP for speed reasons, and make sure
zeros are not signed). Then
./makenemo -j2 -n GYRE_testing -m gfortran_local |& tee compile_log.txt
This uses two processors, with GYRE
as a reference, builds a new folder
called GYRE_testing
, with the specified architecture file, and outputs a
log.
Note
The -r GYRE
flag here only needs to be done once to create an extra folder
and add GYRE_testing
to cfg.txt
. The subsequent compilations should
then read, e.g., ./makenemo -n GYRE_testing -m gfortran_local
.
Check that it does run with the following:
cd GYRE_testing/EXP00
mpiexec -n 1 ./opa
This may be mpirun
instead of mpiexec
, and -n 1
just runs it as a
single core process. Change nn_itend = 4320
in nn_itend = 120
to only
run it for 10 days (rdt = 7200
which is 2 hours). With all the defaults as
is, there should be some GYRE_5d_*.nc
data in the folder. You can read this
with ncview
(see the ncview page or, if you have
sudo
access, you can install it through sudo apt-get install ncview
),
bearing in mind that this is actually a rotated gyre configuration (see the
following NEMO forge page or search
for gyre
in the NEMO book).
Note
If your installation compiles but does not run with the following error
dyld: Library not loaded: @rpath/libnetcdff.6.dylib
Referenced from: /paths/./nemo
Reason: no suitable image found. Did find:
/usr/local/lib/libnetcdff.6.dylib: stat() failed with errno=13
then it is not finding the right libraries. These could be fixed by adding the
-Wl,-rpath,/fill me in/lib
flag to the relevant flags bit in the *.fcm
files (or possibly in XIOS the path
and/or env
) to specify exactly
where the libraries live. This can happen for example on a Mac or if the
libraries are installed not at the usual place.
Note
One infuriating problem I had specifically with a Mac (though it might be a
gcc4.8
issue) is that the run does not get beyond the initialisation
stage. Going into ocean.output
and searching for E R R O R
shows that
it complained about a misspelled namelist item (in my case it was in the
namberg
namelist). If you go into output.namelist.dyn
and look for the
offending namelist is that it might be reading in nonsense. This may happen if
the comment character !
is right next to a variable, e.g.
ln_icebergs = .true.!this is a comment
Fix this by adding a white space, i.e.
ln_icebergs = .true. !this is a comment