cannot build NVHPC on ubuntu (and workaround)

[pescobar]

Hi,

I am using easybuild 4.9.1 on ubuntu jammy (22.04) but I couldn't build NVHPC until I patched the easyblock.

With the default easyblock I get this error in the sanity_check_step

== 2024-04-16 16:56:44,636 build_log.py:171 ERROR EasyBuild crashed with an error (at easybuild/apps/EasyBuild/4.9.1/lib/python3.10/site-packages/easybuild/base/exceptions.py:126 in __init__): Sanity check failed: sanity check com
mand cd /tmp/eb-bk_6m8dc/tmp7fpol1sc && nvc++ -std=c++20 minimal.cpp -o minimal exited with code 2 (output: /scicore/soft/easybuild/apps/binutils/2.40/bin/ld: cannot find -ldl: No such file or directory
/scicore/soft/easybuild/apps/binutils/2.40/bin/ld: cannot find -lpthread: No such file or directory
/scicore/soft/easybuild/apps/binutils/2.40/bin/ld: cannot find -lc: No such file or directory
) (at easybuild/apps/EasyBuild/4.9.1/lib/python3.10/site-packages/easybuild/framework/easyblock.py:3669 in _sanity_check_step)

After some debugging I managed to build it patching the easyblock like this (had to add -L/lib/x86_64-linux-gnu)

diff -ru $EBROOTEASYBUILD/lib/python3.10/site-packages/easybuild/easyblocks/n/nvhpc.py /tmp/nvhpc.py
--- /scicore/soft/easybuild/apps/EasyBuild/4.9.1/lib/python3.10/site-packages/easybuild/easyblocks/n/nvhpc.py   2024-04-05 09:30:35.000000000 +0200
+++ /tmp/nvhpc.py       2024-04-16 16:47:59.268066132 +0200
@@ -221,7 +221,7 @@
             # see: https://github.com/easybuilders/easybuild-easyblocks/pull/3240
             tmpdir = tempfile.mkdtemp()
             write_file(os.path.join(tmpdir, 'minimal.cpp'), NVHPC_MINIMAL_EXAMPLE)
-            minimal_compiler_cmd = "cd %s && nvc++ -std=c++20 minimal.cpp -o minimal" % tmpdir
+            minimal_compiler_cmd = "cd %s && nvc++ -L/lib/x86_64-linux-gnu -std=c++20 minimal.cpp -o minimal" % tmpdir
             custom_commands.append(minimal_compiler_cmd)

         super(EB_NVHPC, self).sanity_check_step(custom_paths=custom_paths, custom_commands=custom_commands)

I am not sure how to fix it upstream. If you give me some advice I can send a proper PR

[pescobar]

CCing those who contributed to this easyblock in case they can give any feedback @AndiH @jfgrimm @appolloford @boegel

[jfgrimm]

hmm, don't think I've seen that issue before. I'll have a go at building on ubuntu 22.04

[jfgrimm]

I can't reproduce this on any of my Ubuntu 22.04 systems

[cgross95]

We see this exact same issue on Ubuntu 22.04 building NVHPC-23.7-CUDA-12.1.1.eb. We haven't tried the workaround yet. We can also provide any other information from our configuration if it would be helpful for reproducing.

[cgross95]

Here is a test report.

[AndiH]

I'm really no EB expert, just one of the authors of the EasyBlock.

To me, it sounds like the compiler doesn't look in the right directories for the library objects; i.e. the LD_LIBRARY_PATH may be incomplete. Not sure, though, if this falls into EasyBuild's responsibility…

[pescobar]

@jfgrimm can you try this to check what libraries you link?

$> module purge

$> module load NVHPC/23.7-CUDA-12.3.0

$> cat <<EOF > minimal.cpp
#include <ranges>
int main(){ return 0; }
EOF

$>  nvc++ minimal.cpp -o minimal

$> ldd minimal

This is what I get in my system. The only difference is that I had to compile the binary using nvc++ -L/lib/x86_64-linux-gnu minimal.cpp -o minimal

$> ldd minimal
        linux-vdso.so.1 (0x00007ffdf57f9000)
        libatomic.so.1 => /scicore/soft/easybuild/apps/GCCcore/12.3.0/lib64/libatomic.so.1 (0x00007f1ff6e24000)
        libnvhpcatm.so => /scicore/soft/easybuild/apps/NVHPC/23.7-CUDA-12.3.0/Linux_x86_64/23.7/compilers/lib/libnvhpcatm.so (0x00007f1ff6c00000)
        libstdc++.so.6 => /scicore/soft/easybuild/apps/GCCcore/12.3.0/lib64/libstdc++.so.6 (0x00007f1ff69d8000)
        libnvomp.so => /scicore/soft/easybuild/apps/NVHPC/23.7-CUDA-12.3.0/Linux_x86_64/23.7/compilers/lib/libnvomp.so (0x00007f1ff5800000)
        libnvcpumath.so => /scicore/soft/easybuild/apps/NVHPC/23.7-CUDA-12.3.0/Linux_x86_64/23.7/compilers/lib/libnvcpumath.so (0x00007f1ff5200000)
        libnvc.so => /scicore/soft/easybuild/apps/NVHPC/23.7-CUDA-12.3.0/Linux_x86_64/23.7/compilers/lib/libnvc.so (0x00007f1ff4e00000)
        libc.so.6 => /lib/x86_64-linux-gnu/libc.so.6 (0x00007f1ff4bd8000)
        libgcc_s.so.1 => /scicore/soft/easybuild/apps/GCCcore/12.3.0/lib64/libgcc_s.so.1 (0x00007f1ff69b7000)
        libm.so.6 => /lib/x86_64-linux-gnu/libm.so.6 (0x00007f1ff68d0000)
        /lib64/ld-linux-x86-64.so.2 (0x00007f1ff6e30000)
        libdl.so.2 => /lib/x86_64-linux-gnu/libdl.so.2 (0x00007f1ff68cb000)
        libpthread.so.0 => /lib/x86_64-linux-gnu/libpthread.so.0 (0x00007f1ff68c6000)

[jfgrimm]

oh, I just realised that although I installed EB 4.9.1, I was still building with 4.9.0...

I now get the same error 🤦

[jfgrimm]

I'm guessing that the issue is with the generated localrc. I don't think this is actually on easybuild doing something weird, I get the same issue if I manually run makelocalrc -x -gcc $(which gcc) -gpp $(which g++) -g77 $(which gfortran)

Adding the following line to the generated localrc fixes the issue for me:

set DEFLIBDIR=/lib/x86_64-linux-gnu;

[pescobar]

I noticed this line which is specific to debian systems

but it's not applied for versions < 21 . See here

patching the easyblock like this also workarounds the issue but to be honest I am not sure what would be the right fix

 diff -ru /scicore/soft/easybuild/apps/EasyBuild/4.9.1/lib/python3.10/site-packages/easybuild/easyblocks/n/nvhpc.py ~/tmp/nvhpc.py
--- /scicore/soft/easybuild/apps/EasyBuild/4.9.1/lib/python3.10/site-packages/easybuild/easyblocks/n/nvhpc.py   2024-04-17 11:23:00.892718404 +0200
+++ /scicore/home/scicore/easybuild/tmp/nvhpc.py        2024-04-17 12:51:54.115386938 +0200
@@ -54,9 +54,6 @@
 # contents for siterc file to make PGI/NVHPC pick up $LIBRARY_PATH
 # cfr. https://www.pgroup.com/support/link.htm#lib_path_ldflags
 SITERC_LIBRARY_PATH = """
-# get the value of the environment variable LIBRARY_PATH
-variable LIBRARY_PATH is environment(LIBRARY_PATH);
-
 # split this value at colons, separate by -L, prepend 1st one by -L
 variable library_path is
 default($if($LIBRARY_PATH,-L$replace($LIBRARY_PATH,":", -L)));
@@ -188,12 +185,11 @@
                 if os.path.islink(path):
                     os.remove(path)

-        if LooseVersion(self.version) < LooseVersion('21.3'):
-            # install (or update) siterc file to make NVHPC consider $LIBRARY_PATH
-            siterc_path = os.path.join(compilers_subdir, 'bin', 'siterc')
-            write_file(siterc_path, SITERC_LIBRARY_PATH, append=True)
-            self.log.info("Appended instructions to pick up $LIBRARY_PATH to siterc file at %s: %s",
-                          siterc_path, SITERC_LIBRARY_PATH)
+        # install (or update) siterc file to make NVHPC consider $LIBRARY_PATH
+        siterc_path = os.path.join(compilers_subdir, 'bin', 'siterc')
+        write_file(siterc_path, SITERC_LIBRARY_PATH, append=True)
+        self.log.info("Appended instructions to pick up $LIBRARY_PATH to siterc file at %s: %s",
+                      siterc_path, SITERC_LIBRARY_PATH)

         # The cuda nvvp tar file has broken permissions
         adjust_permissions(self.installdir, stat.S_IWUSR, add=True, onlydirs=True)

[pescobar]

To summarize, the problem is that debian systems need to have folder /usr/lib/x86_64-linux-gnu/ in env var LIBRARY_PATH. This is needed to compile the example in the sanity_check step and also to build any other easyconfig using the nvhpc toolchain

possible solutions are:

• update the easyblock so that the generated rc file fixes it

• update the easyblock to use the right LIBRARY_PATH for sanity_check step and define the required env vars in the nvhpc easyconfig .e.g add this to the easyconfigs

allow_append_abs_path = True
modextrapaths = {
    'LD_LIBRARY_PATH':  'Linux_%(arch)s/%(version)s/compilers/extras/qd/lib/',
    'LIBRARY_PATH':  'Linux_%(arch)s/%(version)s/compilers/extras/qd/lib/'
    }

modextrapaths_append = {
    'LD_LIBRARY_PATH':  '/usr/lib/x86_64-linux-gnu',
    'LIBRARY_PATH':  '/usr/lib/x86_64-linux-gnu'
    }