[Wien] Segmentation faults

L. D. Marks L-marks at northwestern.edu
Thu Apr 28 14:27:33 CEST 2005


Like others, we've seen some segmentation faults. I've managed to cure 95%
of them, but there is one I've not been able to and maybe someone can make
a suggestion.

Quick intro: in general a segmentation fault occurs when a program tries
to access memory it should not. In most cases this is due to an error in
the code or the something similar. For instance, not every "allocate" call
in Wien2k checks the IOSTAT flag to see if there is enough memory
available. They can also occur with "broken" systems; perhaps 90% of linux
systems (and others) and not correctly setup.

One issue we've traced and cured concerns multiple versions of ifc and the
mkl libraries. One way to switch versions is to use the LD_LIBRARY_PATH
variable, but this can be dangerous (see
http://www.visi.com/~barr/ldpath.html for some interesting comments), and
we've found that it does not always work the way we expected. Much more
stable (needs su) is to edit /etc/ld.so.conf on every node and run
ldconfig, although this might not be the safest method. The ldd tool is
also useful to check that you really have the libraries set the way you
think you do. It's also advisable to use "which ifc" to see if you are
really using the version that you think you are. Compiling with one
version then running with another will give problems almost impossible to
trace.

The one problem we have not solved is that our nodes seem sometime to be
hogging some memory in the cache at times, and not releasing it. In one
case this was 1.6G out of 2G (using vmstat). When several codes (lapw1 for
a large calculation) started they were not able to get enough memory and
crashed with a segmentation fault. The "cure" was to reboot the node. I'd
appreciate any suggestions; I don't think this is an intrinsic Wien
problem, although ...

-----------------------------------------------
Laurence Marks
Department of Materials Science and Engineering
MSE Rm 2036 Cook Hall
2220 N Campus Drive
Northwestern University
Evanston, IL 60201, USA
Tel: (847) 491-3996 Fax: (847) 491-7820
email: L - marks @ northwestern . edu
http://www.numis.northwestern.edu
-----------------------------------------------




More information about the Wien mailing list