No effect of MPI_Wait if -trace flag is set

[JonasFriedrich]

Hello,

for the analysis of my Fortran MPI code, I added the -trace flag. The environment was set by source /opt/intel/oneApi/setvars.sh. Not often throughout the entire code, but right at the beginning, the MPI_IBcast is called. Without the -trace flag, it runs as expected and after the MPI_Wait call every rank has the same values.

But when the -trace flag is set, the MPI_Wait afterwards does not seem to work accordingly. Because of the missing values, a following error occurs and the program crashes.

My system consists of a i9-9980XE in openSuse Leap 15.4 and I built the  problem with mpiifort of oneApi mpi version 2021.12.

I tried I_MPI_ASYNC_PROGRESS=1, but that did not solve the problem.

I attached a snippet of the problem, so that

>/opt/intel/oneapi/mpi/2021.12/bin/mpiifort test_MPI_Wait_problem.F90 -o test

>mpirun -n 4 -l ./test

gives in the output of every proc "buffer after non-blocking bcast 1"

But if the trace flag is added

>mpirun -trace -n 4 -l ./test

gives in the output of the procs receiving from the broadcasting root "buffer after non-blocking bcast 0"

which is not correct as it should be 1.

Am I missing something or is this a bug?

Thanks in advance,

Jonas

[Vipin_Singh1]

Hi Jonas,  we would like to inform you that we are routing your query to the dedicated team for further assistance.

[Rafael_L_Intel]

Hello Jonas!

Thank you for reporting this. Indeed, libVT seems to be misbehaving for the non-blocking Bcast. I will report this to the ITAC developers as a bug.

Cheers,

Rafael

[Rafael_L_Intel]

Hey Jonas,

I hve reported the issue to the engineering team. However, I would like to highlight that ITAC is going to be deprecated, as you can see in the link below:

https://www.intel.com/content/www/us/en/developer/articles/technical/trace-analyzer-deprecation.html

So a fix will not be made public at this time.

Cheers!

Rafael