[Pkg-scicomp-devel] Bug#498213: libmpich1.0gf: SIGSEV in MPI_Bcast when the size of the array is>40000

Ralf Wildenhues wildenhues at ins.uni-bonn.de
Tue Feb 2 15:56:35 UTC 2010


* Ralf Wildenhues wrote on Mon, Feb 01, 2010 at 04:33:53PM CET:
> This is a packaging bug.  It does not occur if you self-build
> mpich-1.2.7p1, so upstream seems innocent.

It's even better.  I've now rebuilt Ubuntu package mpich_1.2.7-8 
(which is probably exactly the same as the Debian package) on this
Ubuntu Hardy, with
  apt-get source mpich
  cd ...
  debian/rules build

and guess what, the resulting mpich ch_p4 works!  I expect the same
to be the case on Debian.

> This bug has about the same impact as a silent wrong-code in GCC
> for those wanting to do serious work with MPICH.  Please don't
> ignore this bug.

To clarify this statement a bit, there are incarnations of this bug
(when just happening to use the right message size) that don't outright
segfault, but cause silent data corruption on the transferred message.


This bug report is
- easily reproducible,
- has a test case (two even, if you look at the corresponding Ubuntu page)
- and is likely very easily fixable by just doing a simple rebuild of
  the package, and whether that worked is easily verifiable too, and it is
- very painful bug for those that have to experience it.

Is that enough of a prerequisite that we can get somebody to look at it?

Thanks,
Ralf





More information about the Pkg-scicomp-devel mailing list