[SciPy-User] scipy import problem in apache-mod_wsgi environment
Ralf Gommers
ralf.gommers at googlemail.com
Sun Nov 6 16:49:27 EST 2011
On Sat, Nov 5, 2011 at 8:52 PM, Myunghwa Hwang <mhwang4 at gmail.com> wrote:
> Hello, list!
>
> First of all, this will be a long message, due to the complexity of our
> environment.
> So, please be patient with my question.
>
> I am trying to run a simple django application in a cluster environment.
> And, my application hangs while it imports scipy.linalg, and both scipy
> and apache do not write out error messages.
> When I run my application in my local python shell, it imports
> scipy.linalg. But, somehow it does not when it is run by apache.
> So, after reading this message, please share any ideas about how to debug
> this problem or new solutions to address this issue or deploy my
> application.
>
> Now, let me explain our current setup.
> 1. OS
> -- The server is a compute cluster where each node runs centos 6 that was
> installed from a clean version of centos6 minimal.
> 2. Apache
> -- Apache 2.2 is also manually installed from one of default linux
> repository.
> To be specific, it is installed from source code together with httpd-dev.
> 3. Python
> -- Python 2.7.2 is also installed from source code across all nodes in the
> cluster. Its source code is downloaded from python.org's ftp.
> 4. Python packages: nose, numpy, scipy
> -- Nose 1.1.2 was downloaded from pypi.python.org and installed from its
> source code.
> -- numpy 1.6.1 was downloaded and installed from a linux repository. When
> building numpy, gnu95 fortran complier was used.
> -- To install scipy, we installed atlas-3.8.4, lapack-3.3.1, and blas from
> their source code.
> ----- atlas was from sourceforge's 3.8.4 stable version. To compile altas,
> gcc was used.
> ----- lapack and blas was obtained from netlib.org's repository. To
> compile the package of lapack and blas, gforan was used.
> ----- Finally, after exporting paths to blas, lapack, and atlas,
> scipy-0.9.0 was installed from its source code.
> scipy was obtained from sourceforge.net's repository.
>
> All of the above were installed in the same way across all nodes in our
> cluster.
> Since I am the only user of the cluster who needs to run python web
> applications,
> I installed python virtualenv package in my local directory.
> Within my virtual environment, django-1.3 and pysal-1.2 (our own package)
> were installed.
> To deploy my web applications, we used mod_wsgi.
> mod-wsgi was compiled with python-2.7.2 and loaded into apache-2.2.
> My application is attached. Basically, it is 'hello world' application
> that tests if numpy, scipy, and pysal can be imported.
> In the attached file, lines 4-9 are just adding paths to django and pysal
> so that apache knows where to find these packages.
> Also, to let apache know where to find atlas-related packages, the path to
> those packages was added to the LD_LIBRARY_PATH environment variable in the
> /etc/sysconfig/httpd file.
>
> When I first ran my application, it just hung and wrote no message.
> So, across scipy.linalg modules, I added print out statements to figure
> out at which point the import was broken.
> Here is the messages I got when I imported scipy.linalg in my local python
> shell.
>
> 1. ########################
> 2. starting linalg.__init__
> 3. pre __init__.__doc__
> 4. pre __init__.__version__
> 5. pre __init__.misc
> 6. pre __init__.basic
> 7. #######################
> 8. Starting basic
> 9. pre basic.flinalg
> 10. pre basic.lapack
> 11. pre basic.misc
> 12. pre basic.scipy.linalg
> 13. pre basic.decomp_svd
> 14. pre __init__.decomp
> 15. ################
> 16. starting decomp
> 17. pre decomp.array et al.
> 18. pre decomp.calc_lwork
> 19. pre decomp.LinAlgError
> 20. pre decomp.get_lapack_funcs
> 21. pre decomp.get_blas_funcs
> 22. ####################
> 23. Starting blas
> 24. pre blas.scipy.linalg.fblas
> 25. pre blas.scipy.linalg.cblas
> 26. pre __init__.decomp_lu
> 27. pre __init__.decomp_cholesky
> 28. pre __init__.decomp_qr
> 29. #################
> 30. Starting special_matrices
> 31. pre special_matrices.math
> 32. pre special_matrices.np
> 33. pre __init__.decomp_svd
> 34. pre __init__.decomp_schur
> 35. ##################
> 36. starting schur...
> 37. pre decomp_schur.misc
> 38. pre decomp_schur.LinAlgError
> 39. pre decomp_schur.get_lapack_funcs
> 40. pre decomp_schur.eigvals:1320454147.23Fri Nov 4 17:49:07 2011
> 41. schur testing
> 42. pre __init__.matfuncs
> 43. #####################
> 44. Starting matfuncs
> 45. pre matfuncs. asarray et al
> 46. pre matfuncs.matrix
> 47. pre matfuncs.np
> 48. pre matfuncs.misc
> 49. pre matfuncs.basic
> 50. pre matfuncs.special_matrices
> 51. pre matfuncs.decomp
> 52. pre matfuncs.decomp_svd
> 53. pre matfuncs.decomp_schur
> 54. pre __init__.blas
> 55. pre __init__.special_matrices
>
> When scipy.linalg is successfully imported, I should get these messages.
> But, when my web application tried to import scipy.linalg, the output
> messages stop at line 41.
> At line 41, decomp_schur.py tries to import decomp.py. Since decomp.py was
> already imported at line 16, scipy ignores it and continues to import other
> modules in my local shell.
> But, somehow, in apache-mod_wsgi environment, scipy failed to ignore or
> reload decomp.py and seems to kill my web application.
> This is really odd, because python does not give any message about this
> error and neither does apache. apache just hangs without sending out any
> response.
> Since lapack and blas functions were imported successfully, the problem
> seems not related to path setup.
>
> If anyone in the list has any insights into or experience into this kind
> of symptom,
> please share your insights and experience. In particular, debugging
> techniques or less-known installation/compilation problems would be helpful.
> I feel like I am at a dead end. So, please help me.
>
> Thanks for reading this post.
> I will look forward to yo
>
Looking at linalg/__init__.py the register_func calls are an obvious
candidate for causing this strange issue. of the few functions left in
decomp_schur.py, norm() and dot() have both gone through register_func,
which messes with the call stack. Could you comment out all calls to that
function and see if that helps? Replace functions that then become
unavailable with numpy ones of the same name where needed.
Cheers,
Ralf
>
>
> --
> Myung-Hwa Hwang
> GeoDa Center
> School of Geographical Sciences and Urban Planning
> Arizona State University
> mhwang4 at gmail.com or Myunghwa.Hwang at asu.edu
>
> _______________________________________________
> SciPy-User mailing list
> SciPy-User at scipy.org
> http://mail.scipy.org/mailman/listinfo/scipy-user
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.scipy.org/pipermail/scipy-user/attachments/20111106/d260284b/attachment.html>
More information about the SciPy-User
mailing list