DUMPI with a Fortran code

[sudheerchunduri]

Hi,
I am using the dumpi library with instrumentation using libdumpi_enable_profiling and call libdumpi_disable_profiling()routines in a FORTRAN77 code.

I build the code with "-I/projects/Performance/chunduri/sst-dumpi/install/include -L/projects/Performance/chunduri/sst-dumpi/install/lib -ldumpif77 -ldumpi".

The build process completed without any issues, however, when running the code, it generates the following error:

Rank 1378 [Thu Jul 20 02:48:45 2017] [c1-1c2s10n2] Fatal error in MPI_Attr_get: Invalid argument, error stack:
MPI_Attr_get(141): MPI_Attr_get(MPI_COMM_WORLD, keyval=1681915906, attr_value=0x7fffffff63e4, flag=0x7fffffff6290) failed
MPI_Attr_get(99).: The attribute value is not the address of a pointer or pointer-sized integer. A common error is to pass the address of an integer to any of the MPI_Xxx_get_attr routines on systems where the size of a pointer is larger than the size of an integer.
.....
....
Similar error reported from each MPI rank. I suspect this to be a FORTRAN - C interoperability issue.

Can you suggest anyways to resolve this.

Thanks
Sudheer

[jjwilke]

Sorry about that. The GitHub notifications apparently aren't working for that repo so I missed the issue.
I am your point of contact. I assume this issue occurs regardless of whether libdumpi_enable/disable calls are happening?

Your best bet for now is to simply not add that function to the trace. It doesn't really provide any useful info for analysis/replay anyway.
When configuring, add the following to CPPFLAGS:

-DDUMPI_SKIP_MPI_ATTR_GET

You can keep turning off functions that don't matter and are breaking the code. Hopefully the interop bug isn't present in any MPI_Send,collective etc calls.
It's obviously not ideal, but your best bet for now until we can fix the interoperability bug.