Closed SeanBryan51 closed 9 months ago
Building CABLE manually seems to work so it looks like this is a bug in benchcab. Could be to do with how we are loading modules?
@bschroeter @ccarouge
We haven't changed how we are loading modules, have we?
If I create my own conda environment with benchcab v3.0.0 in it, it works. So it looks to be a combination of benchcab loading modules and the way the hh5 conda environments are managed.
Interestingly, we get a different set of errors with a different compiler. If benchcab loads intel-compiler/2021.10.1
, we get the errors:
./serial_cable "ifort" "-O2 -fp-model precise" "-L/apps/netcdf/4.8.0/lib/Intel -O0" "-lnetcdf -lnetcdff" "/apps/netcdf/4.8.0/include/Intel"
/bin/ld: /g/data/hh5/public/apps/miniconda3/envs/analysis3-23.10/bin/../lib/gcc/x86_64-conda-linux-gnu/13.2.0//libgcc.a(_muldi3.o): unable to initialize decompress status for section .debug_info
/bin/ld: /g/data/hh5/public/apps/miniconda3/envs/analysis3-23.10/bin/../lib/gcc/x86_64-conda-linux-gnu/13.2.0//libgcc.a(_muldi3.o): unable to initialize decompress status for section .debug_info
/bin/ld: /g/data/hh5/public/apps/miniconda3/envs/analysis3-23.10/bin/../lib/gcc/x86_64-conda-linux-gnu/13.2.0//libgcc.a(_muldi3.o): unable to initialize decompress status for section .debug_info
/bin/ld: /g/data/hh5/public/apps/miniconda3/envs/analysis3-23.10/bin/../lib/gcc/x86_64-conda-linux-gnu/13.2.0//libgcc.a(_muldi3.o): unable to initialize decompress status for section .debug_info
/g/data/hh5/public/apps/miniconda3/envs/analysis3-23.10/bin/../x86_64-conda-linux-gnu/sysroot/usr/lib/../lib64/libm.so: undefined reference to `__get_cpu_features@GLIBC_PRIVATE'
//apps/hdf5/1.10.7/lib/libhdf5.so.103: undefined reference to `powf@GLIBC_2.27'
Hi all
@ccarouge put in a ticket on cws_help about this. The issue is that ifort
searches along $PATH
to find ld
, so it winds up using the ld
packaged with the analysis3 environments (or any conda environment that brings in binutils
in our case its a dependency of parcels
). There are a couple of fixes I can think of. The most expedient fix would be to just remove $CONDA_PREFIX/bin
from $PATH
immediately before invoking make
. Another option might be to create a config option for compiler location, if its set to system
, modify the path, otherwise if its set to conda
, leave $PATH
as is.
Thanks @dsroberts , we'll look at how we want to modify $PATH
.
The build step fails when running benchcab from the hh5 conda environment. See output below.
Steps to reproduce