[Pkg-scicomp-devel] Bug#498213: libmpich1.0gf 1.2.7-8 SIGSEGV bug confirmed, affects 3rd party apps

[Andrew Reid]

  Hi all --

  Just wanted to add my voice to this report.

  A number of users in my lab here have run into this, the 
initial symptom was that locally-commpiled versions of the 
molecular-dynamics code LAMMPS were failing with SEGV when the 
model systems got beyond a certain size.  I conjecture that
LAMMPS makes use of large Bcast operations.

  I have also confirmed the work-around -- I downloaded the
source package, and built it unmodified on a 64-bit system
via dpkg-buildpackage, and installed three of the resulting
.deb files (mpich-bin, libmpich1.0gf, and libmpich1.0-dev)
on several of our systems.

  Both the problem LAMMPS jobs and the test program filed
with the related Ubuntu bug 231062
(https://bugs.launchpad.net/debian/+source/mpich/+bug/231062)
now operate correctly.

  I know that mpich 1.x is deprecated, and in the fullness of
time, we will migrate to Debian-packaged openmpi and/or mpich2,
but 64-bit "lenny" is our workhorse for now, and a fix would
definitely be appreciated here.

  Many scientific codes are distributed as source, and are 
MPI-capable, and in these cases, often the user building the code 
is not a developer, and can't easily break up Bcast operations 
into smaller chunks.

				-- A.
-- 
Dr. Andrew C. E. Reid
Computer Operations Administrator
Center for Theoretical and Computational Materials Science
National Institute of Standards and Technology, Mail Stop 8910
Gaithersburg MD 20899 USA
andrew.reid at nist.gov