[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: Bug#944631: python3-mpi4py: import sends messages to stderr



On 2019-11-15 18:06, Alastair McKinstry wrote:
Hi all,

Do you think it is worth setting :

export OMPI_MCA_btl_base_warn_component_unused=0

in the defaults for OpenMPI ?



Hi Alistair, I think it could be reasonable to do that. Debian is "the universal operating system", so our standard build is intended for a wide variety of systems. Obviously many of those systems will not have OpenFabric/Infiniband transport and we would not expect them to have it. Especially for computational workstations, which are not HPC but for which MPI is still very useful. So the warning is not particularly helpful for the general case. For the specific systems where the hardware does support OpenFabric, the warning won't appear anyway.

The only case where the warning is useful is as a diagnostic where OpenFabric transport is available, but cannot be used by OpenMPI for some reason (e.g. if ib_core kernel module is not available). But the default Debian Linux image does provide ib_core (builds with CONFIG_MLX5_INFINIBAND=m) so we don't expect it to not work. If a HPC centre supporting OpenFabric hardware is building its own kernels then we can expect them to know what they're doing.

Perhaps if you set OMPI_MCA_btl_base_warn_component_unused=0 by default, then it would be worth adding a comment to README.Debian suggesting to set OMPI_MCA_btl_base_warn_component_unused=1 if a user or administrator wants to verify that no components like OpenFabric have gone unused. Are there any other components other than OpenFabrics that OMPI_MCA_btl_base_warn_component_unused=0 might hide?

Another question, in common cloud computing (e.g. Amazon), or full HPC systems, how likely is it that the installation will have OpenFabric hardware anyway? How often would these systems not have it?

Drew


Reply to: