Incompatible with older OSes (RHEL6) #32

pelson · 2016-06-23T13:28:21Z

Looks like we are building a matplotlib which doesn't work with older libstdc++:

>>> import matplotlib.contour
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "matplotlib/contour.py", line 16, in <module>
    import matplotlib._contour as _contour
ImportError: /usr/lib64/libstdc++.so.6: version `GLIBCXX_3.4.15' not found (required by matplotlib/_contour.so)

ldd lib/python2.7/site-packages/matplotlib/_contour.so                                      
    linux-vdso.so.1 =>  (0x00007ffc59738000)
    libpython2.7.so.1.0 => /.../lib/python2.7/site-packages/matplotlib/../../../libpython2.7.so.1.0 (0x00007ff8d0eea000)
    libstdc++.so.6 => /usr/lib64/libstdc++.so.6 (0x00007ff8d0bbb000)
    libm.so.6 => /lib64/libm.so.6 (0x00007ff8d0937000)
    libgcc_s.so.1 => /lib64/libgcc_s.so.1 (0x00007ff8d0721000)
    libpthread.so.0 => /lib64/libpthread.so.0 (0x00007ff8d0503000)
    libc.so.6 => /lib64/libc.so.6 (0x00007ff8d016f000)
    libdl.so.2 => /lib64/libdl.so.2 (0x00007ff8cff6b000)
    libutil.so.1 => /lib64/libutil.so.1 (0x00007ff8cfd67000)
    /lib64/ld-linux-x86-64.so.2 (0x0000003998e00000)

ls -ltr /usr/lib64/libstdc++.s*                                                                                                      
-rwxr-xr-x. 1 root root 825480 Jun 28  2010 /usr/lib64/libstdc++.so.5.0.7
lrwxrwxrwx. 1 root root     18 Oct 31  2011 /usr/lib64/libstdc++.so.5 -> libstdc++.so.5.0.7
-rwxr-xr-x  1 root root 989840 Jun  1  2015 /usr/lib64/libstdc++.so.6.0.13
lrwxrwxrwx  1 root root     19 Aug 28  2015 /usr/lib64/libstdc++.so.6 -> libstdc++.so.6.0.13

This is for a RHEL6 machine.

Ping @jakirkham as I have been out of the loop a little on our compiler choices.

The text was updated successfully, but these errors were encountered:

ocefpaf · 2016-06-23T13:32:24Z

I am pretty sure that is b/c conda-forge matplotlib is picking up numpy with openblas and in the env libgcc is not being installed. I am getting a lot of similar reports everywhere. We have to either expand our build matrix to bake np_blas or remove conda-forge's numpy.

PS: can you try to install libgcc to confirm that is the same issue? And check which numpy was installed in the env mkl, vanilla, or conda-forge with openblas (which should have installed libgcc with it).

jakirkham · 2016-06-23T14:14:36Z

I am pretty sure that is b/c conda-forge matplotlib is picking up numpy with openblas and in the env libgcc is not being installed.

Doubtful. The libraries are pointing to the system directory.

ls -ltr /usr/lib64/libstdc++.s*                                                                                                      
-rwxr-xr-x. 1 root root 825480 Jun 28  2010 /usr/lib64/libstdc++.so.5.0.7
lrwxrwxrwx. 1 root root     18 Oct 31  2011 /usr/lib64/libstdc++.so.5 -> libstdc++.so.5.0.7
-rwxr-xr-x  1 root root 989840 Jun  1  2015 /usr/lib64/libstdc++.so.6.0.13
lrwxrwxrwx  1 root root     19 Aug 28  2015 /usr/lib64/libstdc++.so.6 -> libstdc++.so.6.0.13

This is interesting. So, why are there 2 libstdc++, @pelson?

ocefpaf · 2016-06-23T14:16:01Z

Doubtful. The libraries are pointing to the system directory.

That is exactly the issue! The conda versions, which numpy+openblas were compiled with, are not there!!

Using our docker image:

conda install -c conda-forge install matplotlib
python -m "import matplotlib.contour"
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/opt/conda/lib/python3.5/site-packages/matplotlib/contour.py", line 16, in <module>
    import matplotlib._contour as _contour
ImportError: /usr/lib64/libstdc++.so.6: version `GLIBCXX_3.4.15' not found (required by /opt/conda/lib/python3.5/site-packages/matplotlib/_contour.cpython-35m-x86_64-linux-gnu.so)

conda install libgcc
python -c "import matplotlib.contour"

Success. Or force install numpy with openblas.

ocefpaf · 2016-06-23T14:34:53Z

FYI here is the source of the problem (search for numpy in the CircleCI logs of the latest matplotlib build):

numpy-1.10.4 |py27_blas_openblas_201 6.8 MB conda-forge

that installs:

libgcc-5.2.0 | 0 1.1 MB defaults

breaking everything that depends on numpy but somewhow fails to get the conda-forge version later on.

I already reported this in conda-forge/numpy-feedstock#15, but the discussion got derailed. I hope we can fix this soon.

jakirkham · 2016-06-23T15:45:36Z

Please see PR ( conda-forge/staged-recipes#855 ), which splits out a libgfortran package for Mac and Linux. The next step would be to make openblas only depend on libgfortran at install time. Thus there would be no libstdc++ pulled in by openblas. If that is the problem, that would solve it cleanly.

pelson · 2016-06-23T16:58:13Z

Doubtful. The libraries are pointing to the system directory.

Actually the ldd tells us that we are taking libstdc++.so.6 dynamically from the linker PATH, so having a libstdc++ in the prefix would be picked up before that in /usr/lib64, no?

This is interesting. So, why are there 2 libstdc++, @pelson?

I'm on a coorporate machine and it is likely to be stock RHEL6 installation, but could be that it has been modified by our sys-admins (though I doubt it)

breaking everything that depends on numpy but somewhow fails to get the conda-forge version later on.

Nice digging @ocefpaf. Interesting problem - @jakirkham sounds like you might be onto something - let's take a look at doing that and see if we can manoeuvre out of the problem that way. We have a hangout tomorrow and we can asses the likelihood of us fixing the issue in the short term vs. needing to remove the build.

ocefpaf · 2016-06-23T17:29:47Z

Nice digging @ocefpaf.

I am receive a torrent of e-mails from IOOS users complaining about our numpy breaking their envs for almost a month now. Sadly I am suffering a big resistance to even discuss the matter here.

pelson · 2016-06-23T17:33:14Z

our numpy breaking their envs for almost a month now

OK, we will have to take stock and address this. Tomorrow's dev meeting will need to have a decisive solution - the worst solution IMO is to remove numpy, but it most certainly is and option and is definitely on the cards.

ocefpaf · 2016-06-23T17:41:34Z

OK, we will have to take stock and address this. Tomorrow's dev meeting will need to have a decisive solution - the worst solution IMO is to remove numpy, but it most certainly is and option and is definitely on the cards.

Thanks @pelson it is good te be heard. I never actively merged anything to "fix my problem." All I asked was for this: to have a voice in the community. I feel I am losing mine here in conda-forge.

jakirkham · 2016-06-24T15:59:04Z

@pelson could you please share with @msarahan and myself some more info about this target machine? In particular, it would be nice to know the results of the following commands.

gcc --version
strings /usr/lib64/libstdc++.so.6 | grep LIB

jakirkham · 2016-06-24T15:59:51Z

This would also be good info from you too, @ocefpaf, if you have access to the affected machines or from your users if possible.

pelson · 2016-06-24T16:03:14Z

$> strings /usr/lib64/libstdc++.so.6 | grep LIB
GLIBCXX_3.4
GLIBCXX_3.4.1
GLIBCXX_3.4.2
GLIBCXX_3.4.3
GLIBCXX_3.4.4
GLIBCXX_3.4.5
GLIBCXX_3.4.6
GLIBCXX_3.4.7
GLIBCXX_3.4.8
GLIBCXX_3.4.9
GLIBCXX_3.4.10
GLIBCXX_3.4.11
GLIBCXX_3.4.12
GLIBCXX_3.4.13
GLIBC_2.2.5
GLIBC_2.3
GLIBC_2.4
GLIBC_2.3.2
GLIBCXX_FORCE_NEW
GLIBCXX_DEBUG_MESSAGE_LENGTH

$ > gcc --version
gcc (GCC) 4.4.7 20120313 (Red Hat 4.4.7-16)

pelson · 2016-06-24T16:03:54Z

$ uname -a

Linux eld118 2.6.32-573.22.1.el6.x86_64 #1 SMP Thu Mar 17 03:23:39 EDT 2016 x86_64 x86_64 x86_64 GNU/Linux

ocefpaf · 2016-06-24T16:06:51Z

I don't have access but I do know that most of the users have gcc 4.4.7 or gcc 4.4.8. (That is why I want to pin to 4.4*)

jakirkham · 2016-06-24T16:36:27Z

So, this PR ( conda-forge/openblas-feedstock#6 ) fixes the import issue that was mentioned with matplotlib.contour above ( with a corresponding version bump build of matplotlib ). Unfortunately, that is the only issue I can reproduce. Could either of you please try this?

Basically, use the binaries from my channel under the dev label to test this. Please let me know what you find.

conda install -c jakirkham/channel/dev openblas matplotlib

Alternatively you can try rebuilding them from source. Just note building OpenBLAS takes ~30mins. Though matplotlib is much faster.

ocefpaf · 2016-06-27T14:01:24Z

Fixed in #34

pelson · 2016-06-27T14:58:13Z

Thank you very much to @jakirkham and @ocefpaf for helping to fix this.

$ /downloads/miniconda/bin/python
>>> import matplotlib.contour
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/downloads/miniconda/lib/python2.7/site-packages/matplotlib/contour.py", line 16, in <module>
    import matplotlib._contour as _contour
ImportError: /usr/lib64/libstdc++.so.6: version `GLIBCXX_3.4.15' not found (required by /downloads/miniconda/lib/python2.7/site-packages/matplotlib/_contour.so)
>>> 

$> /downloads/miniconda/bin/conda update --override-channels -c conda-forge -c defaults matplotlib

The following packages will be downloaded:

    package                    |            build
    ---------------------------|-----------------
    matplotlib-1.5.1           |      np111py27_4         6.7 MB  conda-forge

The following packages will be UPDATED:

    matplotlib: 1.5.1-np111py27_3 conda-forge --> 1.5.1-np111py27_4 conda-forge

$> /downloads/miniconda/bin/python -c "import matplotlib.contour"

ocefpaf · 2016-06-27T15:23:13Z

Thank you very much to @jakirkham and @ocefpaf for helping to fix this.

Thanks @pelson for getting this issue under people's radar. I clearly failed at that. Could you please consider conda-forge/numpy-feedstock#16 as that is the last step for our envs to be re-built.

ocefpaf mentioned this issue Jun 23, 2016

Test pin gcc conda-forge/openblas-feedstock#7

Closed

This was referenced Jun 27, 2016

Bump build for newer openblas. #34

Merged

rebuild conda-forge/numpy-feedstock#16

Merged

ocefpaf closed this as completed Jun 27, 2016

jakirkham mentioned this issue Oct 4, 2016

Should fix #61 #71

Closed

kmuehlbauer mentioned this issue Dec 18, 2017

V0.11.2 conda-forge/wradlib-feedstock#29

Merged

jakirkham mentioned this issue Mar 5, 2018

Missing dependencies: compilers and C++ runtime conda-forge/pandas-feedstock#40

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Incompatible with older OSes (RHEL6) #32

Incompatible with older OSes (RHEL6) #32

pelson commented Jun 23, 2016

ocefpaf commented Jun 23, 2016 •

edited

Loading

jakirkham commented Jun 23, 2016

ocefpaf commented Jun 23, 2016 •

edited

Loading

ocefpaf commented Jun 23, 2016 •

edited

Loading

jakirkham commented Jun 23, 2016

pelson commented Jun 23, 2016

ocefpaf commented Jun 23, 2016

pelson commented Jun 23, 2016

ocefpaf commented Jun 23, 2016 •

edited

Loading

jakirkham commented Jun 24, 2016

jakirkham commented Jun 24, 2016

pelson commented Jun 24, 2016

pelson commented Jun 24, 2016

ocefpaf commented Jun 24, 2016

jakirkham commented Jun 24, 2016 •

edited

Loading

ocefpaf commented Jun 27, 2016

pelson commented Jun 27, 2016

ocefpaf commented Jun 27, 2016

Incompatible with older OSes (RHEL6) #32

Incompatible with older OSes (RHEL6) #32

Comments

pelson commented Jun 23, 2016

ocefpaf commented Jun 23, 2016 • edited Loading

jakirkham commented Jun 23, 2016

ocefpaf commented Jun 23, 2016 • edited Loading

ocefpaf commented Jun 23, 2016 • edited Loading

jakirkham commented Jun 23, 2016

pelson commented Jun 23, 2016

ocefpaf commented Jun 23, 2016

pelson commented Jun 23, 2016

ocefpaf commented Jun 23, 2016 • edited Loading

jakirkham commented Jun 24, 2016

jakirkham commented Jun 24, 2016

pelson commented Jun 24, 2016

pelson commented Jun 24, 2016

ocefpaf commented Jun 24, 2016

jakirkham commented Jun 24, 2016 • edited Loading

ocefpaf commented Jun 27, 2016

pelson commented Jun 27, 2016

ocefpaf commented Jun 27, 2016

ocefpaf commented Jun 23, 2016 •

edited

Loading

ocefpaf commented Jun 23, 2016 •

edited

Loading

ocefpaf commented Jun 23, 2016 •

edited

Loading

ocefpaf commented Jun 23, 2016 •

edited

Loading

jakirkham commented Jun 24, 2016 •

edited

Loading