Intel MPI slow startup on Windows for local single-machine execution

[Bruno22]

We're experiencing ~10 second startup overhead with `mpiexec -localonly -np 8` on Windows for local (single-machine) execution.

With `mpiexec -localonly -np 1` it takes 1 second

Current config: set I_MPI_OFI=0 && mpiexec -localonly -np 8 MODL.exe
We tested various optimizations (ASYNC_LAUNCH, PIN_DOMAIN=off, etc.) but they all made performance worse.

Questions:

•  Are there Windows-specific optimizations for local-only execution?
• Is ~10 second MPI overhead normal on Windows?

Details: Migrated from MS-MPI to Intel MPI. No cluster, no network, no CUDA needed.

[Sergey_K_Intel3]

What version of Intel MPI are you using?

Can you try setting I_MPI_FABRICS=shm and possibly passing "-bootstrap fork" option to mpiexec?

[Bruno22 · Accepted solution]

Thanks for your feedback. We investigated this further and found the root cause on our side: our company antivirus was scanning each MPI process at startup.

This explains the behavior very well:
- With 1 process, startup is around 1 second.
- With 8 processes, startup increases to around 10 seconds.
- The overhead scales with the number of launched processes.

So in our case, the delay is not caused by Intel MPI itself, but by security software overhead on Windows during process creation.

We also found that signing the binaries with a trusted certificate completely resolved the issue.