[CIG-SHORT] OpenMPI and Infiniband: workaround for segmentation faults when PyLith starts
Brad Aagaard
baagaard at usgs.gov
Tue Nov 13 18:03:48 PST 2012
Hi all,
On a few clusters with Infiniband and OpenMPI, I have encountered
segmentation faults when PyLith is starting up. The symptom is that at
some point during the initialization process you will see segmentation
faults when running on more than one compute node. This appears to be a
result of how PyLith spawns subprocesses from Python using fork().
Infiniband will sometimes attempt to do something special with fork().
The workaround is to turn off Infiniband support for fork in OpenMPI.
This is most easily done by setting the following environment variables
(bash style):
export OMPI_MCA_mpi_warn_on_fork=0
export OMPI_MCA_btl_openib_want_fork_support=0
I have added the error message and the workaround to the PyLith FAQ:
http://www.geodynamics.org/cig/community/workinggroups/short/workarea/pylith-wiki
Regards,
Brad
More information about the CIG-SHORT
mailing list