[Pkg-openmpi-maintainers] Bug#814451: openmpi: stderr output is an error message
Gilles Gouaillardet
gilles.gouaillardet at gmail.com
Thu Feb 25 05:11:42 UTC 2016
Folks,
i was able to reproduce this behaviour.
here is what the openib (e.g. infiniband) btl does :
- if running on Linux, and there is no /sys/class/infiniband, then the
btl disqualifies itself (see ompi_common_verbs_check_basics()(
- if the btl did not previously disqualified itself, and there is no
infiniband port available, then a warning message is issued (we
consider there is likely something wrong)
i am able to reproduce this scenario on a VM with no infiniband
hardware by loading the mlx4_core kernel module.
can you confirm you are running into this very same scenario ?
is the mlx4_core.ko or similar module loaded ?
if yes, is there any reason for it to be loaded ?
is the __linux__ macro defined on debian ?
does the /sys/class/infiniband directory exists ?
is this directory empty ?
from an OpenMPI point of view, it could be possible to disqualify the
openib btl with no warning message if /sys/class/infiniband is empty,
would that be helpful ?
Cheers,
Gilles
More information about the Pkg-openmpi-maintainers
mailing list