openmpi/debian/README.Debian

27 lines
1.3 KiB
Plaintext

Please be aware that as of release 4.1.0-5, Debian now sets the following:
OMPI_MCA_btl_base_warn_component_unused=0
btl = ^uct,openib
pml = ^ucx
osc = ^ucx
This is set in /etc/openmpi/openmpi-mca-params.conf and turns off warnings that OpenFabric is configured
but not used on the system. Many or most users of Debian will not have OpenFabric present on their systems
and turning this off prevents spurious warnings that were breaking higher-level libraries and
applications.
However the Libfabric EFA provider currently does not support applications calling fork()
when memory registrations are handled by the EFA provider. This is to prevent applications
from encountering potential data corruption or crashes when operating in this mode.
Most applications, including the majority of MPI and NCCL applications, will still function
with this constraint. Some applications do require fork() support, however,
these applications do not require memory registrations to be handled by the EFA provider.
If you wish to use OpenIB and know the application to be safe, comment out the exclude line.
Similarly, UCX is disabled by default as it gives spurious errors on missing RDMA fabric.
Please enable this if you have the appropriate hardware.
-- Alastair McKinstry <mckinstry@debian.org>, 2021-01-14