-
Notifications
You must be signed in to change notification settings - Fork 2
Open
Description
The openmpi package in Spack is quite complicated to configure. One aspect is the scheduler option. It seems the basic case is that with it enabled it installs a mpirun wrapper script around srun. This helps with applications such as orca which call mpirun inside a prebuilt executable. One example where this is beneficial is setting network options in slurm for single_node_vni. Without the mpirun wrapper script, openmpi on a single node will break with error such as
Open MPI failed an OFI Libfabric library call (fi_domain). This is highly
unusual; your job may behave unpredictably (and/or abort) after this.
Local host: xxxx
Location: mtl_ofi_component.c:1007
Error: Function not implemented (70368744177702)
This is due to libfabric expecting Slingshot 11 capabilities to be setup but Slurm doesn't do this by default. It requires the srun option --network=single_node_vni.
Lets set OpenMPI to use the scheduler option to create the wrapper script.
Reactions are currently unavailable
Metadata
Metadata
Assignees
Labels
No labels