version `GFORTRAN_10' not found

[chenyubi14]

Hi,

I compiled the packages successfully. However, when I tried to do the test example by running the binary elphbolt.x, I ran into this issue

elphbolt.x: /lib64/libgfortran.so.5: version `GFORTRAN_10' not found (required by elphbolt.x)

I searched this issue online, found two methods in this website , but both of them don't work.

Here is the setup of my system. I am working on a cluster (pod on cnsi.ucsb.edu) where I don't have the privilege to edit /lib64/. The installed libgfortran.so in the cluster has versions lower than GFORTRAN_10. I found that the package OpenCoarrays/ has libgfortran.so that supports GFORTRAN_10 in two paths: OpenCoarrays/prerequisites/installations/gnu/12.1.0/lib64/libgfortran.so and OpenCoarrays/prerequisites/builds/gcc-12.1.0/x86_64-pc-linux-gnu/libgfortran/.libs/libgfortran.so.

1. The first method is to link libgfortran.so to the libgfortran.so in OpenCoarrays/ . It failed because I cannot edit /lib64/
2. The second method is to use LD_PRELOAD. I put LD_PRELOAD=${HOME}/OpenCoarrays/prerequisites/builds/gcc-12.1.0/x86_64-pc-linux-gnu/libgfortran/.libs/libgfortran.so in my ~/.bashrc file. I faced this error

ERROR: ld.so: object '~/OpenCoarrays/prerequisites/builds/gcc-12.1.0/x86_64-pc-linux-gnu/libgfortran/.libs/libgfortran.so' from LD_PRELOAD cannot be preloaded: ignored.

Both methods failed, but I think there might be a way to change the path to libgfortran.so.5 in elphbolt.x. I tried to search where this path is defined in the source files, but I failed. I wonder whether you could point out a way to solve this issue.

THANKS!

[nakib]

Hi chenyubi14,

Thanks for using elphbolt and opening this issue.

I think the following combination worked for me for building OpenCoarrays 2.9.2: gnu_gcc/9.2.0, openmpi/4.0.5gcc9.2.0, and cmake/3.18.4. You can specify the versions of these dependencies when you build OpenCoarrays. The documentation of elphbolt gives a couple of examples. Let me know if this fixes your issue.

Have you managed to successfully build elphbolt on a desktop or a laptop?

Best,
Nakib

[chenyubi14]

Hi Nakib,

Thanks a lot for your response!

I don't really think the suggestions could resolve my issue. In any case, I have successfully compiled elphbolt on my ubuntu system, and it works smoothly. Thus, it is not urgent for me to use elphbolt in the cluster.

If you want to hear my long-winded problem, here it is. My problem cannot be solved by building OpenCoarrays as your suggested, because I have already successfully compiled OpenCoarrays. My cluster does not have the new versions of gcc as required by OpenCoarrays, so I can not load modules like what you did. As a result, OpenCoarrays installed gcc, mpich, etc, in its subfolder with the new versions. Therefore, I have GFORTRAN_10 supported in a subfolder of OpenCoarrays/, but elphbot.x cannot use a libgfortran.so.5 that supports GFORTRAN_10.

I think there is a way to link elphbot.x to the fortran in the subfolders of OpenCoarrays/, so this is why I am posting this issue. If my descriptions are not clear enough, you may also ignore it, and I can continue using elphbolt with my ubuntu.

THANKS!

[nakib]

Good to know that you were able to build the program on your Ubuntu machine. For production level calculations, however, you will probably have to move on to a cluster at some point. To that end, I can suggest that you ask your cluster maintainer if they can create a singularity container from a dockerfile (https://docs.sylabs.io/guides/2.6/user-guide/singularity_and_docker.html). Here's a dockerfile which I have tested on my laptop: https://github.com/nakib/elphbolt/blob/develop/Dockerfile. This should be a completely portable solution. Let me know if this is a possibility for your cluster.

Another option is a spack (https://spack.readthedocs.io/en/latest/) package. Let me know if you are interested and I will share with you an example, but I am not an expert on it by any means. :)

Best,
Nakib

[chenyubi14]

Hi Nakib,

I have tried building docker image by the dockerfile, and install it by singularity. However, I seem to have the preload issues, so we gave up on this method.

With the help of the cluster managers, we finally have elphbolt.x working properly. The final working method is by using "module load intel/20;module load openmpi/2.1.0" and "include msi_intel.make" on Pod.

Thanks,
Yubi

[nakib]

Dear Yubi

Good to know that you managed to get the code running on your cluster. Have fun!

Best
Nakib

[chenyubi14]

I need to point out one more issue, and my final solution.

If I use "msi_intel.make", there are other problems. One major problem is that no matter how many cores I assigned, like cafrun -n 2 elphbolt.x, it will always use 40 cores. In addition, the jobs may not properly finish.

The reason is that "msi_intel.make" uses FC=ifort, but FC=caf is needed to resolve the problems.

However, FC=caf like in "lp_gfortran.make" and "thickpad_gcc.make" will have the very first issue I mentioned.
elphbolt.x: /lib64/libgfortran.so.5: version GFORTRAN_10 not found (required by elphbolt.x)

My final solution is add export LD_LIBRARY_PATH=/<path-to-OpenCoarrays>/prerequisites/installations/gcc/<release-tag>/lib64:$LD_LIBRARY_PATH to ~/.bashrc. (I got the hint from this website https://github.com/sourceryinstitute/gcc/ )
It will have FC=caf working perfectly, and the number of cores exactly follows my command cafrun -n 2 elphbolt.x
Note OpenCoarrays should be compiled by ./install.sh

[nakib]

Hi Yubi,

For an intel build, there is no need to use cafrun. The caf and cafrun commands are for use with gcc+opencoarrays. For an intel build, you can try something like this in your run script:

export FOR_COARRAY_NUM_IMAGES=2
<path to your elphbolt executable>/elphbolt.x

Best,
Nakib