Rob, first of all, thank you for your support and help. Here is what we found out: > How many servers do you have running in this test? > 24 servers
> I have two thoughts: > > - Your backtrace shows you linked with /lib64, and you're running > OpenSuse. I presume then that you're running in a bi-arch > environment. Could you have possibly built pvfs2-server as a 32 bit > executable but ended up linking it with 64 bit libraries? I have to > confess that this theory is a bit of a longshot... > > - When you built OPENMPI you might have compiled against some oddball > pvfs2.h header file or linked with an incompatible libpvfs2. Do you > have any other pvfs installations on your system? Are you sure? > Check the configure output: was configure able to find pvfs2-config? > Check your mpicc wraper script: is it including links to the > expected libpvfs2? > let me start with the second bullet. You were in fact right, that our open MPI configuration did not pick up the right directory. However, since we had CFLAGS, LDFLAGS and LIBS set up accordingly before running configure, the pvfs2 adio did get compiled. Regarding the first bullet, we are a little confused right now. We managed to generate a more stable configuration which didn't crash by adding in CFLAGS a '-m64' before recompiling PVFS2. Alternatively, if we do not set the '-m64' but force PVFS2 to create a non-optimized version e.g. by setting CFLAGS to '-g -O0', we do achieve the same, i.e. a stable pvfs2 configuration. Do you think that it is possible that we are hitting here a gcc 4.3 bug? Thanks once again for your support Mohamad > I've run your test code on my (32 bit) laptop (4 procs, one server) > and on a 64 bit Ubuntu system (4 procs, 4 servers) and did not see a > segfault. Thanks for sending along a testcase, but I'm afraid I'm > not going to be able to help very much if I can't reproduce the crash > on my end. > > Sometimes I get weird behavior when the PVFS + MPI + application > software stack gets out of sync: the one other suggestion I can make > is to 'make clean' and rebuild everything, in case symbols from an > earlier iteration are somehow floating around (they shouldn't be, but > sometimes it happens) > > ==rob > > -- Mohamad Chaarawi Research Assistant http://www.cs.uh.edu/~mschaara Department of Computer Science University of Houston 4800 Calhoun, PGH Room 526 Houston, TX 77204, USA _______________________________________________ Pvfs2-users mailing list [email protected] http://www.beowulf-underground.org/mailman/listinfo/pvfs2-users
