$I_MPI_WAIT_MODE is no longer supported by recent Intel MPI versions

[boegel]

worker defines $I_MPI_WAITMODE in the job script that it generates:

export I_MPI_WAIT_MODE=enable

This causes trouble with recent Intel MPI versions (tested with Intel MPI 2021 update 4 + 2019 update 9):

Abort(566799): Fatal error in PMPI_Init: Other MPI error, error stack:
MPIR_Init_thread(82)...:
MPIR_T_cvar_init(19376): Unable to parse environment variable (I_MPI_WAIT_MODE)

Apparently support for $I_MPI_WAIT_MODE was removed in Intel MPI recently, see also https://community.intel.com/t5/Intel-oneAPI-HPC-Toolkit/I-MPI-WAIT-MODE-replacement-in-Intel-MPI/m-p/1173203.

Is it strictly necessary to set $I_MPI_WAIT_MODE, or would it be OK to just strip this out?
Will worker still work correctly if it's not set (perhaps suboptimally)?

[gjbex]

We noticed similar issues. It is okay to simply leave out I_MPI_WAIT_MODE, the consequence is that slaves which are no longer active will spin, but worker will work correctly.

The risk of load imbalance may increase as well when the -master flag is used, but that would mostly impact scenarios where worker is used inappropirately in the first place.

[gjbex]

The bugifix/conformance branch (1.6.14) removes I_MPI_WAIT_MODE and has some code clean up.

[stevenvdb]

I don't want to reopen the discussion on whether or not worker needs I_MPI_WAIT_MODE. I just want to mention that this variable is still supported (for sure in Intel(R) MPI Library, Version 2021.6, click), but now only supports 0 or 1 as values (no longer enable or disable).

[gjbex]

Indeed, I_MPI_WAIT_MODE was resurrected in 19 update 5, probably because its intended alternative didn't live up to expectations.